arXane › Computer Science › Computer Science
Wednesday, 4 November 2026: 148 new papers, 104 cross-lists and 193 replacements in Machine Learning (cs.LG), in arXiv's order, with full abstracts. The same papers as arXiv's own listing, set as a newspaper; every title links to arXiv.
Lifelong (continual) anomaly detection systems must learn a sequence of tasks without forgetting how to recognize anomaly types seen earlier -- a property known as backward transfer. A common way to try to improve backward transfer is to reshape a model's internal representation around learnable concept centroids, typically coupled with a new way of scoring anomalies based on distance to those centroids. Because these two design choices are usually introduced together, it is unclear which one is actually responsible for any performance gain. In this paper, we unbundle the two mechanisms within a VAE-based lifelong anomaly detector and study them separately. We find that used purely as a training-time regularizer -- while leaving anomaly scoring unchanged -- latent cluster separation can improve detection quality and backward transfer, though whether it outperforms the alternative scoring approach depends on the dataset's underlying representation. Switching to distance-based scoring is instead where most of the instability commonly associated with this class of method actually originates; we identify and explain the underlying mechanism. This benefit depends on the dataset's underlying representation, and we report this as practical guidance for choosing between the two design choices, rather than as a list of caveats.
Bubeck and Sellke's universal law of robustness says that fitting $n$ noisy samples below the noise level with a class of $p$ real parameters forces a Lipschitz constant of order $\sqrt{nd/p}$, when the covariate distribution on $\mathbb{R}^d$ is isoperimetric: smooth interpolation costs a factor $d$ of overparametrization. Its hypotheses are Gaussian-type concentration, a homogeneous covariate law, scalar outputs, and independent samples; we relax each one. Exponential concentration suffices, at the unchanged rate and absolute constant and at the price of a capacity burn-in, some form of which we show is necessary; this gives the law for normalized log-concave covariates, with a $\log d$ loss from the current KLS bound. For heterogeneous mixtures the worst component constant is replaced by the weighted average of the component constants, under a typicality condition on the mixture composition. For vector-valued predictors the bound at the Frobenius Lipschitz scale is free of the output dimension $K$, and the Euclidean bound loses only $K^{-1/2}$, a loss that covariate-independent noise removes. Stationary Markov covariates retain the i.i.d. rate, the pseudo-spectral gap entering only the sample-size conditions. One master theorem carries the shared argument, and constants are explicit throughout.
Cardiac auscultation traditionally detects the presence of a heart murmur, but clinicians further characterize it by timing, shape, pitch, quality, and grade. Most automatic phonocardiogram analysis systems stop at binary detection and do not estimate this richer description. In this work, we formulate automatic murmur characterization as five patient-level classification tasks and compare three multi-spot strategies on recordings from the aortic, pulmonary, mitral, and tricuspid auscultation locations. First, classical and deep-learning classifiers are trained independently at each location, and predictions are combined via a maximum-severity late-fusion rule. Second, a synchronous multi-input CNN jointly processes one representative segment per location for a single task. Third, a unified multi-input, multi-output CNN predicts all five attributes from one shared multi-spot representation. Using 735 patients from the CirCor DigiScope dataset with complete four-spot recordings, the best late-fusion configurations reach an unweighted mean sensitivity between 85.56% and 93.81% across the five tasks, with class-specific recalls between 80.00% and 100.00% for every relevant category. The multi-input CNN trails late fusion by 7-16 points depending on the task, while the multi-output model reaches values between 72.59% and 79.40%, estimating all five attributes from one shared representation. We further show that naively balancing training data by adding all minority-class segments degrades patient-level performance under the maximum-severity rule, since it introduces non-representative, noisy segments not filtered by the clustering-based segment selection used for evaluation. These results indicate multi-spot murmur characterization is feasible with both fusion-based and jointly learned architectures, outlining a path toward automatic, structured murmur reports complementing binary screening.
Language models often spread a concept across many internal dimensions, making their features difficult to interpret. Sparse autoencoders (SAEs) recover clearer features after training. We ask whether models can instead learn interpretable output coordinates directly. We introduce Mixture of Wide Experts (MoWE), a sparse feed-forward layer in which selected experts read the full representation but write to fixed, disjoint output coordinates. On a controlled task with known hidden causes, this design preserves predictive performance while localizing causes more clearly than dense and standard mixture-of-experts (MoE) models. At approximately matched compute, we train two 19-layer runs of each architecture for 10 billion tokens. MoWE achieves evaluation cross-entropy 2.417 versus 2.452 for dense, while standard MoE reaches 2.22. We measure native-coordinate readability by asking a language-model judge to describe each coordinate from passages where it is strongly active and test that description on new passages. Correcting for descriptions that also fit other coordinates gives a specificity score of 0.523 for MoWE, 0.375 for dense, and 0.268 for standard MoE. TopK SAEs trained on the same FFN updates score 0.489 for dense and 0.556 for MoWE, making native MoWE comparable to the dense SAE. In the first case study, we identify a four-coordinate causal path supporting the model's prediction of degrees after the prompt A right angle is 90. We then copy seven angle-related coordinate values into the model as it processes A rectangle has four, changing its next-token prediction from sides to degrees. In the second, one fixed coordinate set detects hazardous requests and another steers the next token toward refusal. These examples show that MoWE combines sparse conditional computation with coordinates that can be identified and reused directly.
Why do some of the world's most powerful Artificial Intelligence (AI) systems still struggle to reason about cause and effect? Causal discovery, the ability to discern the underlying drivers of a system, is fundamental to scientific inquiry and has traditionally relied on structured data and statistical methods. Recently, Large Language Models (LLMs) have been explored as an alternative paradigm for causal inference, leveraging their broad knowledge and reasoning capabilities. Yet, despite their linguistic fluency, LLMs often rely on surface-level pattern recognition rather than genuine causal logic. While it is known that LLMs stumble in this domain, we lack a precise map of where, why, and how significantly these failures occur. In this paper, we propose a systematic stress-test framework across six critical dimensions: model scale, causal graph complexity, prompting strategy, metadata quality, data quality, and sensitivity to uncertainty. Our findings reveal a significant "optimism bias" in LLMs; models consistently overestimate causal links, leading to low precision across the board. Furthermore, we demonstrate that traditional evaluation metrics can be deceptive: while reasoning-heavy models lead in performance, the highest scores often occur when uncertainty in inference is overlooked. Once uncertainty is integrated into the evaluation, performance drops, providing a more honest reflection of causal reliability. We also find that, while scaling model size improves outcomes, these gains are fragile; performance degrades sharply as the quality of metadata decreases. This work serves as a practical guide to the blind spots of LLMs, offering a clear-eyed assessment of when these tools can be integrated into a causal discovery pipeline and when they are likely to lead researchers astray.
Deep neural networks extract powerful latent representations from images, enabling fast and efficient downstream analysis. However, these latent spaces lack explicit geometric structure, leading to possibly poor interpolation fidelity and misleading extrapolation. Latent interpolation techniques assume symmetric transition costs which are often not the case for real world data. Their symmetric nature cannot capture directional processes, such as disease progression, where certain transitions are impossible (e.g. tumor growth dynamics or cognitive improvement in Alzheimer's disease). Although Riemannian metrics extend Euclidean geometry by incorporating curvature, they remain symmetric. We show that understanding asymmetry and not curvature alone directly addresses these problems. To do so we propose a new Finsler-Randers metric learning framework. It incorporates direction-dependent distances, thus enabling the computation of trajectories respecting monotonic sequential constraints. We demonstrate its potential for interpolation and extrapolation on controlled cross-sectional datasets.
In receiver-directed federated learning, each site draws a trust boundary. For every source and model component it judges, on its own data, whether the transfer helps (admit), harms (reject) or cannot yet be told (abstain). We ask whether this boundary holds over hundreds of rounds. Our benchmark has a federation of thirteen semi-synthetic sites, an exact oracle for every evaluated transfer, paired seeds, an attack protocol, fourteen methods over 600 rounds and Qwen3-1.7B LoRA sites over 500. The answer depends on which boundary is measured. Across 20 paired CPU seeds, a gate judging against a fixed reference better identified transfers useful relative to the site's starting model. Yet it admitted 29% of transfers harmful to the model the site now has. A gate judging against the current model admitted 2% of harmful transfers on that target in the CPU tier, but missed many transfers useful to the starting model. Two short propositions explain these distinct failures. We also evaluate a held out audit that leaves gate verdicts intact. After training, it decides whether each final model is deployed or the site falls back to local training, and reports admissions that have gone stale. With independent audit data its guarantees hold after any training history. At 600 rounds the audit deployed 65 to 72% of models, none with negative benefit on its target, and kept 98% of the gain. No stale report had the wrong sign in our runs. With 20,000 audit observations, a variance-adaptive bound found 99% of stale transfers that cost at least 0.01 in log loss. A common-target Qwen3 comparison on 60 selected transfers shows the same tension over 100 rounds. With 38% adversarial sites, FedFomo and CS-Gate were the only evaluated methods with positive mean gain in every honest group.
Benchmarks built from model hubs combine interaction data, knowledge graphs (KGs) and resource documentation, but valid files do not ensure that these artifacts describe the same objects. Treating the release as a data product, we audit HuggingBench (SIGIR 2025), a benchmark for recommending Hugging Face resources, through data validation, a replication of its eleven recommendation methods and controlled experiments with three training seeds. In three of its four external KGs, the files pass key and range checks but violate the item-entity identity constraint on which KG recommenders rely. Correcting this join improves every KG-based method in every seed, by up to 73% in mean Recall@10, and consistently reverses ten of 27 comparisons between KG configurations. The protocol matters as well: the 100-epoch budget ends before early stopping in seven to eight of 29 validation-selected settings, and 21 of 23 such setting-seed pairs improve with a longer budget. The released item text, which no reported method uses, is the strongest signal: closed-form linear models with text and the item graph exceed the best reported Recall@10 by 18% while fitting in minutes on a CPU. Checks of other recommendation releases show that individually valid releases of one dataset can use incompatible identifiers. We release corrected files, mappings, validators and logs, and derive practices for preserving identity and provenance across the artifacts of benchmarks and model lakes.
Training a classifier means letting a validation metric decide which model is kept and when training stops. A natural requirement on that metric is superiority: every correctly classified forecast should score better than every incorrectly classified one, however close the underlying probabilities are. Neither the Brier score nor the logarithmic score satisfies this on its own. We show that superiority is governed by the separation between two endpoints -- the supremum of the score on correct predictions and the infimum on incorrect predictions -- and derive both in closed form for each score. Their difference is the infimal penalty, the greatest lower bound on the penalty needed to restore superiority. Prior work restores superiority with a penalty calibrated from the first endpoint alone; for the Brier base, our infimal penalty is smaller than that published penalty by a factor ranging from 4.00 at three classes to 2.11 at twenty. We further embed the published and infimal penalties in a one-parameter family and show that, under a single-valued tie convention, every member of the family is strictly proper. In a pre-specified two-dataset case study, mean selected-model macro-F1 rises with the penalty in the reported ten-class regimes, though not distinguishably from noise once seed/selection-mode nesting is accounted for; a fifth regime at forty-seven classes does not reproduce even this descriptive pattern. The intended no-superiority baseline is algebraically a rescaling of the family under test, so the experiment cannot attribute the observed pattern specifically to superiority. When superiority itself is the objective, any penalty strictly above the infimal value suffices; choosing a larger penalty is a separate calibration trade-off rather than a requirement for the guarantee.
Extracting semantically meaningful staypoints from raw trajectory streams is a foundational prerequisite for mobility analytics. However, real-world trajectory data is frequently degraded by severe spatial noise and protracted temporal dropouts. Classic threshold-based extraction methods, such as the widely adopted sliding-window algorithm, are highly sensitive to these anomalies, routinely fracturing continuous stationary episodes into disjointed fragments. To overcome this limitation, we propose the Sequential Staypoint Extraction (SSPE) algorithm. SSPE employs a resilient, state-driven architecture that replaces static spatial anchors with a dynamic centroid and introduces an $n$-step lookahead heuristic to structurally tolerate transient coordinate drift. Concurrently, it utilizes temporal bridging mechanisms to prevent artificial fragmentation during prolonged signal loss. Since optimal extraction parameters vary significantly across datasets, we further introduce a fully unsupervised, consensus-based parameter estimation framework. This framework derives baseline statistical scales directly from the raw trajectory data and identifies the optimal parameter configuration by maximizing structural agreement, thereby eliminating the reliance on annotated ground truth. We evaluate SSPE on the ExtraSensory dataset, a real-world mobility corpus with manually verified ground-truth staypoints, against the sliding-window implementation of Trackintel, one of the most widely adopted extraction baselines in the field. Empirical results demonstrate that SSPE achieves superior accuracy, yielding higher F1-scores while simultaneously improving computational efficiency over this standard.
Accurate forecasting and temporal downscaling of solar irradiance is critical for reliable renewable energy integration and grid management. Traditional methods often struggle to capture rapid irradiance variability. In this work, we develop a Transformer-based architecture for simultaneous forecasting and temporal downscaling. The model takes input data at arbitrary resolutions and generates outputs at any resolutions represented within the training data, leveraging both global and local contextual information. To improve training efficiency, a genetic algorithm-based optimization framework is implemented to identify suitable model hyperparameters, while reducing the computational cost of exhaustive grid search. The effectiveness of the proposed model is shown by analyzing real data in Dallas, Texas. The proposed model achieved root mean square error (RMSE) values of 28.785, 29.544, 30.157, 32.132~W/m\(^2\) for 15-, 30-, 60-minute, and irregularly spaced global input and producing forecast at 5-minutes resolution, respectively. Furthermore, this model produced reliable uncertainty estimates, exhibiting wider prediction intervals for high-resolution inputs during volatile solar conditions, whereas coarser-resolution inputs result in narrower intervals and significantly lower training time and computational cost. Our approach provides a scalable and efficient framework for multi-resolution solar irradiance forecasting and temporal downscaling, suitable for practical energy forecasting and grid management applications.
Knowledge tracing (KT) estimates learners' evolving knowledge states from interaction histories to support adaptive learning. Retrieval-augmented KT improves these estimates using cross-learner behavioral evidence, but existing approaches typically separate evidence retrieval from downstream prediction. This creates a retrieval-inference alignment gap because evidence relevance varies with the learner's evolving state and prediction context. We propose CaRe-KT, a query-adaptive framework that maintains relevance continuity between evidence acquisition and inference without equating retrieval relevance with predictive utility. CaRe-KT introduces an Evidence-Graph Cache (EGC) that combines persistent question-level relational compatibility with interaction-level temporal validity to construct query-specific evidence subgraphs. Historical learning trajectories retrieved through these subgraphs are encoded alongside the target learner's history in a shared representation space using outcome-supervised contrastive learning. An attention mechanism combines acquisition relevance with learned query-evidence compatibility, allowing relevance to inform, rather than determine, each trajectory's predictive influence. Experiments across five benchmarks show that CaRe-KT outperforms strong KT baselines and remains robust under sparse-history and cold-start conditions. These findings support relevance continuity as a design principle for retrieval-augmented knowledge tracing.
Financial workflows need fast, calibrated, auditable decisions, not just generated text. We train System-1 decision models, small language models fine-tuned to answer typed questions (choice/score/noul) with a probability in a single forward pass, across two model families (K2-Horizon and Qwen3.5/3.8) and nine rungs from 0.8B to 32B parameters, using one frozen LoRA recipe on a public decision benchmark and two new finance corpora: a programmatic-gold policy corpus (rule-reading with exact labels) and a reality-labeled outcome corpus (19,640 realized forward-return questions in walk-forward splits, plus a sealed 2026 test of 1,360 questions). Narrow fine-tuning costs general decision competence 9-45 percentage points, a data-mixture effect that shrinks with model scale and is associated with narrow-gradient volume. On reality-labeled questions, every trained rung from 0.8B to 32B lands at the horizon-conditioned prior baseline (accuracy 0.37-0.38, no detectable conditional edge under our controls); the models respond with flat, well-calibrated probabilities that track per-horizon base rates: naturally honest uncertainty, with a mid-scale calibration sweet spot (ECE 0.026 at 2B, degrading to 0.043-0.044 at 9-32B under prior drift). A mixed-corpus remediation carries all competencies in one adapter at modest cost: it matches or slightly exceeds the generalist specialist on the public benchmark while holding the policy ceiling and near-specialist outcome calibration, yielding a deployable artifact with per-domain temperature regimes and a band-dispatch escalation analysis: sharp, dispatchable bands where rules determine answers; flat, review-routing bands where reality does not.
Language models are typically trained with a single choice of architecture on diverse domains and data sources. Modern transformers have been extensively refined for such general-purpose learning, but are they equally effective across different subsets of the pretraining data? We investigate this question by training small language models from scratch on representative subsets of pretraining data such as web text, code, and mathematics. We introduce variants of transformers in which standard activation functions are replaced with learned splines tuned on specific datasets, providing a simple way to modify their inductive biases. Using a mix-and-match of architectures and datasets, we evaluate cross-domain transfer, i.e. how well does an architecture tuned for one domain perform on data from another? We find that the tuned transformer variants improve training efficiency or generalization on all seven of our tested datasets, but the gains are often domain-specific. For example, architectures tuned on a natural language dataset such as FineWeb perform well on TinyStories, whereas those tuned on mathematics perform poorly on natural language. We identify several mechanisms underlying these improvements, including wider loss basins and more abundant "lottery tickets". Our results show that different subsets of pretraining data can benefit from different learning mechanisms, highlighting a limitation of the prevailing trend towards a small set of general-purpose architectures and optimizers. This points to an opportunity to improve data efficiency by developing domain-specific learning strategies that retain the ability to train general-purpose models.
Diffusion models trained on distinct datasets can produce highly similar samples when prompted with the same noise input, suggesting that they learn a generalized data distribution. We investigate the role of noise injected during the sampling process, and show that it plays an important role in making or breaking this strong generalization effect. Based on this finding we investigate the robustness of sample similarity to changes in the injected noise, and find that its impact is largest in the sampling regime where broad structure already exists but the precise identity of the sample is not yet determined.
As general-purpose Large Language Models (LLMs), such as Gemini and DeepSeek, exhibit increasingly advanced reasoning capabilities, a critical question arises: how do their property prediction abilities compare to those of domain-specific pretrained materials models? In this study, we evaluate predictions for Ti, steels, high-entropy alloy (HEA) and Al, demonstrating that general-purpose LLMs surprisingly outperform SteelBERT, MatSciBERT, and SciBERT. To further enhance performance, we introduce a hybrid approach that utilizes numerical priors from a prior model, such as CatBoost or SteelBERT, to guide the general LLMs (Deepseek V3.1, Gemini-2.5-Flash). By inputting these initial estimates from a prior model as prompts, the LLM acts as a reasoning corrector, achieving accuracy superior to the prior model and "unguided mode". Furthermore, we investigate the impact of data strategy using a similarity-based metric, revealing that data quality dominates quantity; using merely 5% of high-similarity reference data of training set as input yields performance similar to the results of using 100% of training set. Additionally, the LLMs demonstrate excellent uncertainty calibration, where self-reported "high" confidence strongly correlates with minimal deviation from ground truth. Our findings suggest that synergizing general-purpose LLMs with prior model numerical baselines and high-quality data curation creates a cost-effective and accurate paradigm for materials discovery, reducing the need for expensive domain-specific pretraining.
Sparse attention cuts the cost of long-context inference by making the attention matrix sparse: each query attends only a subset of the sequence, chosen by a fixed pattern or, in recent production LLMs, learned per query. In DeepSeek Sparse Attention (DSA), the attention of DeepSeek-V3.2 and the GLM-5 family, every layer is preceded by a lightning indexer that scores the whole cache and keeps the k=2048 highest-scoring tokens for the layer to attend. At long context this indexer scan, not the sparse attention it feeds, is where the decode time goes. But adjacent layers select nearly the same tokens, so a layer can attend the selection of the previous indexer layer instead of its own. This removes the dependency: each indexer then runs concurrently with the attention of its layer. The resulting Multi-layer Indexer (MLI) needs no retraining and no new weights. Its default realization changes no kernel, running inside SGLang's production decode graph; an optional fused decode kernel accelerates it further on DeepSeek-V3.2. MLI composes with IndexCache (the cross-layer reuse GLM-5.2 ships as IndexShare): IndexCache drops the indexer from some layers, and MLI hides the scan on the rest. We serve DeepSeek-V3.2 and GLM-5 under IndexCache's protocol at retentions 1/2 to 1/32 (the fraction of layers that keep an indexer) and contexts 10K to 195K. At equal retention the composition decodes faster than IndexCache: up to 14.5% faster on DeepSeek-V3.2 and 9.8% on GLM-5 at 1/2, and 7.9% and 5.1% at 1/4, largest where IndexCache keeps the most indexers and growing with context. At retention 1/r it is faster still than IndexCache at 1/(2r) while keeping twice the fresh selections: at 128K it reaches 1.32x the unmodified model's decode throughput at 1/4, above IndexCache's 1.28x at 1/8. Its LongBench, RULER and LongBench v2 accuracy stays within a few percentage points of IndexCache's.
Therapeutic antibody development is often framed as comprising two discrete stages: initial discovery of novel binders followed by iterative optimization to refine these hits into therapeutic-grade molecules. Models and methods often reflect these stages, with generative structure models applied to discovery and probabilistic black-box optimization methods applied to refinement. Here we propose to unify these stages by smoothly transitioning a generative discovery model into iterative optimization, allowing the handoff between the structure-based prior and supervision from wet lab data to be guided by the posterior uncertainty of a probabilistic surrogate. We show that this yields improvements in optimization efficiency in the low-N regime, as well as improved diversity amongst the optimized molecules.
Large foundation models often answer even when the available visual or textual evidence is insufficient to support a reliable conclusion. This failure to abstain makes them hard to trust in high-stakes or information-seeking settings, where a confident but unsupported answer can be worse than no answer at all. It also raises a natural question: can the internal mechanisms behind this behaviour be identified and controlled at inference time? We present Dr. No, a feature-steering approach that mines sparse autoencoder (SAE) features associated with abstention and commitment, then modulates a small set of them during the prefill pass, without retraining the base model. Dr. No is evaluated on LLaVA-1.6 Mistral 7B across multimodal abstention benchmarks, including UNK-VQA and MM-UPD, and additionally benchmarked on language reasoning tasks using Qwen3.5-2B on AbstentionBench subsets. Against multiple uncertainty-based abstention baselines, feature steering achieves a more favourable balance between correct answering and appropriate abstention than post-hoc rejection methods. The gains are strongest on UNK-VQA and on MM-UPD cases involving image--question or question--answer incompatibility, and transfer most clearly to mathematical reasoning subsets of AbstentionBench. However, performance is more limited on absent-answer detection and retrieval-heavy multi-hop QA, such as MuSiQue. These results suggest that sparse internal features provide lightweight control handles for shaping uncertainty-aware behaviour in foundation models, while also showing that abstention control is source-dependent rather than governed by a single universal signal.
Hyperparameter optimization (HPO) returns a single optimal configuration, systematically discarding the explored performance landscape that encodes how results depend on configuration choices. Fitness landscape analysis comprises a robust collection of methods from evolutionary computation offering ruggedness indices, autocorrelation functions, and fitness-distance correlation that have never been applied to machine learning hyperparameter spaces at scale, given that fine-grained ground-truth enumeration data has not existed. We present ExaTune, an open-source HPC system to produce a new artifact class: the hyperparameter landscape map. ExaTune exhaustively evaluates every point on a discretized hyperparameter grid via $K$-fold cross-validation, aggregates results into reproducible parquet benchmarks, and applies a suite of five formalized landscape metrics: spatial autocorrelation (Moran's $I$), ruggedness, local optima density, fitness-distance correlation (FDC), and variance-based parameter importance. We release a benchmark collection of 8 complete landscape maps (totalling 285,080 exhaustively evaluated configurations) spanning five model families (SVMs, random forests, XGBoost, ridge regression, and $k$-nearest neighbors) across six classification and two regression benchmarks, hosted on HuggingFace. Empirical analysis demonstrates that Bayesian optimization, Optuna, Hyperband, and Random Search at standard trial budgets systematically miss instability regions, secondary optima, and parameter interaction effects that are visible in complete landscapes; information that is structurally unavailable to any optimizer by design.
Large Language Models (LLMs) often generate fluent yet incorrect responses, making un- certainty estimation important for reliable deployment. Semantic Entropy (SE) estimates uncertainty by clustering responses based on semantic equivalence, but it relies on hard entailment-based clustering. We propose Embedding Semantic Entropy, which replaces entailment-based clustering with K-Medoids in semantic embedding space and uses dynamic cluster selection to account for intra-cluster compactness and inter-cluster separation. We evaluate the method on TriviaQA, SQuAD, NQ-Open, SVAMP, and BioASQ using Falcon- 7B-Instruct, LLaMA-2-7B, LLaMA-2-7B-Chat, and Gemma-7B. Results show competitive AUROC and AUARC performance, suggesting that NLI-based entailment clustering can be replaced with embedding-space clustering with limited degradation while improving the inspectability of semantic grouping. This matters because semantic uncertainty scores depend on underlying semantic groups; by making these groups explicit, compact, and separable in embedding space, Emb-SE turns uncertainty estimation from a black-box en- tailment decision into an inspectable clustering problem.
Neural controlled differential equations (Neural CDEs) evolve hidden states along continuous paths constructed by interpolating discrete observations. Sparse measurements leave the input between observation times undetermined, and admissible continuous paths consistent with the same data can produce different terminal states. We characterize this ambiguity for planar drivers with bounded speed. The set of compatible terminal states admits a leading-order zonotopic approximation determined by feasible signed areas and transported Lie brackets. Its projection onto a scalar prediction yields a computable, target-specific ambiguity certificate. We also study the number of measurements needed to reduce the remaining uncertainty. For nondegenerate planar CDEs, the optimal worst-case radius over deterministic adaptive sampling policies scales as $N^{-1}$ when speed is bounded and $N^{-2}$ when both speed and acceleration are bounded, where $N$ is the number of observation intervals. Synthetic experiments assess the certificate using admissible continuous completions and examine the role of path regularity.
Koopman representations describe nonlinear dynamics through linear evolution in a transformed space, offering an attractive framework for their linear simplicity and generality. Yet these advantages do not necessarily translate into improved performance in deep latent data assimilation (LDA). We systematically compare seven representative methods and structural baselines across five systems spanning approximately periodic to strongly chaotic dynamics and a broad range of state dimensions. We evaluate forecast and assimilation accuracy, physical state recovery, latent-trajectory behavior, and dynamical consistency, and assess Gaussian error assumptions, objective convexity, and uniqueness of analysis solutions. Our results reveal a trade-off between representational flexibility and optimization tractability. Linear evolution of the latent state can stabilize short-term forecasts but restrict expressiveness in strongly chaotic and high-dimensional systems, without consistently improving assimilation accuracy. Parameterized linear operators offer greater flexibility, yet their joint optimization with the encoder--decoder remains nonconvex. Methods with closed-form analysis updates generally perform better and exhibit greater robustness in our experiments, although matrix inversion remains a computational bottleneck. These findings show that the practical benefits of Koopman structure depend on both sufficient observational information and the choice of assimilation procedure.
Graph forecasting systems often rely on prediction intervals to quantify uncertainty across connected locations, but the true outcomes needed for evaluation may be only partially observed. We study how an auditor can selectively retrieve missing outcomes to determine whether there exists a connected group whose outcomes violate the same side of their respective forecast intervals. The audit stops when the observed labels determine the event or the query budget is exhausted. On three public graph time series, MOMS reduces mean queries by 19.6% to 38.5% relative to screened random acquisition when 30% of labels are hidden. Under localized masking on METR-LA, shortest witness finds more positive cases within ten queries, although the difference depends on node ordering. Direct component traversal reproduces all 9,696 screened random traces without enumerating connected subsets. The experiments show that reducing queries to a decision does not necessarily improve early event discovery or reduce local computation.
Fast Dominance Estimation Clustering (FastDEC) is an efficient variant of Density Peak Clustering (DPC) that achieves high clustering quality through dominance-based density estimation, but remains constrained by the computational cost of exact k-nearest-neighbor (k-NN) construction and dominance component (DC) processing. We present SymWarp-FastDEC, a GPU-accelerated implementation of FastDEC. At its core is SymTrikNN, an exact self-k-NN kernel that exploits search symmetry using upper-triangular traversal and warp-cooperative top-k selection. We also introduce WP-DC, a warp-parallel DC reduction algorithm designed for modern GPUs. Combined with memory-adaptive execution planning and stream scheduling, WP-DC eliminates DC reduction as a runtime bottleneck. Experiments on synthetic and real-world datasets show that SymWarp-FastDEC produces clustering outputs identical to the original FastDEC on all evaluated equivalence benchmarks. Compared with a FastDEC variant using FAISS GpuIndexFlatL2 for exact k-NN search, our approach achieves approximately 1.7x to 4.9x higher end-to-end performance, with the advantage generally increasing with dimensionality. Compared with the original single-threaded Python FastDEC implementation, speedups reach approximately 2.5 x 10^4 in a very high-dimensional case where its k-d-tree-based search degrades; this extreme result is not representative of typical acceleration.
Multi-teacher on-policy distillation consolidates expert capabilities by supervising student-generated trajectories, whereas task arithmetic offers a training-free alternative through linear parameter combinations. We view the linear combination of student and expert parameters as a single differentiable model and propose P$^2$-MOPD, which jointly optimizes the mixing coefficients and the student weights under the same on-policy distillation objective. Unlike task arithmetic, which stays within the span of the frozen student and expert parameters for any choice of merging coefficients, P$^2$-MOPD keeps the student trainable, so the experts act as parameter priors without confining it to this span. Across four domains at 1.7B and 4B scales, P$^2$-MOPD achieves faster and more balanced recovery than MOPD from the same initialization, surpassing its final average recovery in half the training steps. At the end of training, our 1.7B model reaches 94.7% average recovery, versus 80.9% for MOPD.
Large language models (LLMs) can memorize and reproduce fragments of their training data, raising privacy and legal concerns that motivate machine unlearning. Yet poisoned training examples can remain statistically detectable after unlearning, which raises the question of whether current unlearning objectives remove what the model has memorized at all. In this work, we introduce PINT, a novel algorithm for machine unlearning that uses the Hilbert-Schmidt Independence Criterion (HSIC) to reduce the statistical dependence between tokens to be forgotten and the internal representations used to predict them. We formally relate our HSIC-based objective to Exact Memorization through a computable upper bound, connecting the training objective directly to the model's ability to reproduce the forgotten tokens. In extensive experiments, we compare forgetting-utility trade-offs across eight settings with models up to 7B parameters and demonstrate scalability to 70B parameter models. Our evaluations cover unlearning Gaussian poisons, personal information, hazardous knowledge in biology and cybersecurity, and copyrighted books and news articles. PINT achieves the strongest forgetting-utility trade-off on average and is the only method to make Gaussian poisons undetectable while maintaining near-original model utility. Code is available at: https://github.com/shkabrii/PINT
Many causal inference problems require sampling from interventional outcome distributions rather than estimating low-dimensional summaries. Existing approaches often rely on fitting generative models of the target distribution together with additional nuisance components, which can be computationally demanding for high-dimensional outcomes. We propose Interventional Maximum Mean Discrepancy Flow (IMMD-Flow), a particle-based method that instead represents the target interventional distribution through its kernel mean embedding. This kernel embedding admits a doubly robust, Neyman-orthogonal estimator with finite-sample guarantees and can be efficiently reused to sample from the interventional distribution via MMD gradient flow. IMMD-Flow accommodates both hard and soft interventions, admits theoretical guarantees for estimation and particle approximation, and extends to high-dimensional image outcomes through frozen encoders. Across image-based benchmarks, IMMD-Flow matches or improves upon state-of-the-art generative approaches while showing significant speed advantages.
Evolution equations with time-dependent coefficients force a neural solver to spend capacity on a scale that the equation already determines. We show that a single classical object, the Ermakov-Pinney scale function $\rho$, can be written into the input layer of a physics-informed neural network as a parameter-free embedding. One construction covers three problem families through the two signs of the Pinney constant $c$: linear diffusion ($c=-1$, $\omega\equiv0$), diffusion in a time-dependent confining trap ($c=-1$, $\omega\neq0$), and the quantum oscillator with time-dependent frequency ($c=+1$). An exact moment identity gives $\rho$ in closed form from the initial data alone, so no auxiliary network or knowledge of the solution is needed. Our central finding is comparative: each family requires a different component of the construction and has a different honest control. In family I, the diffusive-time reparametrisation alone reproduces most of the gain, so that family is a poor test of the idea and shows why an ablation is mandatory. In family II no time reparametrisation exists, the control degrades accuracy, and the amplitude prefactor becomes essential (40.8x). In family III the adiabatic approximation also fails, and the quadratic phase factor is what matters (5.7x). The method does not help when the reduced solution does not converge to a fixed point: its extrapolation advantage then disappears. The reduction itself is the classical scaling-variable change of the self-similar-asymptotics literature; what is new is the moment identity that makes it architectural for time-dependent coefficients, and the comparative anatomy above.
Large Language Bayes (LLB) transforms a natural language problem description into a collection of candidate probabilistic programs and performs inference within each model, combining their predictions using Bayesian model averaging. While this provides a flexible framework for probabilistic inference without manual model specification, the resulting candidate model set is often misspecified in practice, and likelihood-based aggregation can concentrate on a single model, under-representing uncertainty across the sampled models. We propose a stacking-based aggregation method for LLB that selects model weights by optimizing predictive performance rather than marginal likelihood, while preserving the original language-model-driven model generation and within-model inference pipeline. Because leave-one-out predictive quantities are intractable in latent-variable models, we derive a variational lower bound that yields a tractable ELBO-based stacking objective. Across real-world probabilistic inference tasks, our method achieves comparable predictive performance to Bayesian model averaging while yielding more dispersed model weights, reducing concentration on a single candidate model under misspecification.
Probing frozen representations layer by layer is now standard practice, but reporting when a signal first becomes decodable requires a rule for reading a depth curve, and the obvious rule is wrong. On synthetic curves with known ground truth, requiring a probe score to exceed a null-calibrated threshold for three consecutive stages reports an onset nineteen stages early under transformer-like depth correlation, with every resample returning an onset. The failure follows from correlation between neighboring layers, a hazard known in neuroimaging but not, to our knowledge, checked for depth curves. We give an estimator whose error rate is calibrated rather than assumed. Null trajectories are built by permuting labels within each evaluation unit and holding each permutation fixed across depth, preserving the correlation structure the rule must survive; thresholds are per-layer order statistics; and the persistence requirement is calibrated rather than chosen. Applied to pocket information in a frozen co-folding trunk, stratifying 1,820 post-cutoff targets by protein-family and pocket familiarity, the five-stage rule fires on 4 to 7 per cent of null trajectories against a nominal 5 per cent and the three-stage rule on 12 to 17: the synthetic failure reproduces on real data. Pocket membership is decodable above threshold at all 48 blocks in all four regimes, so the onset cannot discriminate them; the pre-declared accumulation depth does, and is latest for familiar families with structurally novel pockets in every feature set. That regime is also the weakest, by a margin surviving controls for length, pocket measure and alignment depth. Separately, a widely used public co-folding checkpoint loads against its own code with 128 trunk parameters missing; those would be randomly initialized, and every representation extracted would be partly noise while appearing normal.
This paper develops a mean-field formation transport method with low online computational cost for large-scale multi-agent systems. Classical mean-field planners often require repeated solutions of coupled forward and backward partial differential equations, making online replanning computationally demanding. Density and control are therefore approximated by a common parameterized nonlinear map with separate scalar density and vector control outputs. Parameters are optimized using Fokker--Planck and initial-condition residuals, formation-matching costs on numerically simulated closed-loop particle trajectories, and control effort. Numerical computations are restricted to a bounded region of the state space, with an additional penalty on probability flux across its boundary. For the periodic-domain model, we establish well-posedness of the induced density evolution and bound the density-approximation error in terms of the dynamics and initial-condition residuals. Parameter optimization precedes online operation, during which controller coefficients remain fixed. Experiments quantify the accuracy--runtime trade-off relative to a mean-field planner, show lower transport error than classical feedback baselines, and demonstrate the benefits of the common parameterization.
Prestress governs the stability, stiffness, and dynamic behavior of tensegrity structures, but it is not directly observable from sparse structural measurements, and it may decrease during service life. Under strong excitation, the cables may slacken and the dynamics become nonsmooth. Existing identification methods are limited to the taut regime, and physics-informed networks for tensegrities solve only the forward problem of predicting the response from known member parameters. This study formulates the inverse problem. Neural networks are trained on the equation-of-motion residual in constrained Cartesian coordinates, where the mass matrix is constant and the force densities enter linearly. The force densities, damping, and axial stiffness are trainable parameters, and the tension-only cable law is kept. A Fisher-information analysis shows, before inversion, which member parameters a record can determine. The method is tested on simulated planar tensegrity frames with up to five modules and a three-dimensional prism, and on measured records of a biomimetic leg and a full-scale tensegrity simplex. On the one-module frame, the identifiable force densities are recovered within 1.5% from noisy accelerations at half the free coordinates, on records with and without taut-slack transitions. Records with transitions determine members that taut records leave undetermined. On the five-module frame, a 16-cable two-span bridge, the median force-density error rises moderately, to 4.2%. On measured records, an identified spring stiffness is within 0.1% of its independently measured value on average, and the identified simplex prestress at 13 levels agrees with its cable force transducers to 2.9% root-mean-square.
To enable effective generalization in offline meta-RL, it is essential to accurately infer the underlying task through latent task representations, which provide a compact and transferable summary of task-specific characteristics for guiding policy generation. In this paper, we propose Contrastive Task-Aware Decision Transformer (CoTA-DT), a DT-based offline meta-RL framework that learns task representations by jointly integrating four complementary components: (i) contrastive learning from offline trajectories, (ii) alignment with natural-language task descriptions, (iii) one-step supervision via a task-conditioned world model, and (iv) long-term supervision via a task-conditioned Q function. The learned task representations are used as conditioning signals for a DT-based policy, enabling task-aware meta-policy learning. To further support generalization in both zero-shot and few-shot settings, we design a language-grounded, self-adaptive strategy that enables effective and reliable adaptation to unseen tasks. Notably, in CoTA-DT, the pretrained LLM serves as a shared backbone for language encoding, world modeling, Q-function estimation, and policy learning. This unified design facilitates parameter sharing, improving both efficiency and generalization. Extensive experiments across diverse offline meta-RL benchmarks demonstrate that CoTA-DT consistently outperforms prior methods, particularly in challenging zero-shot generalization scenarios.
We study score learning from denoising pairs when the data contain a hidden non-Gaussian component. If covariance reveals the component, its subspace can be recovered directly. If covariance is isotropic, the first distinguishing moment has order $k^\ast$, and the learner must extract that moment from clean points and their noisy views. Correlations in the denoising loss are Stein discrepancies against a baseline; for polynomial fields they determine exactly the corresponding moment differences. Averaging fresh noisy views conditional on a clean point gives an exact moment query on the clean distribution. In the covariance-matched regime, orientation-independent dictionaries satisfying an $L^4/L^2$ condition need order $d^{k^\ast-1}$ features, and the stated dot-product kernel ridge models need the same power in samples. Small-step online SGD on one unit has this exponent up to logarithms. A two-stage learner uses the curvature of a smoothed denoising loss to recover the hidden direction, then sharpens it and fits a scalar readout. For every even $k^\ast$, it obtains a fixed score gain over the Gaussian baseline from $O(d^{k^\ast/2}\log d)$ clean points. In the covariance-visible regime, the frozen-feature threshold occurs at $d^J/J!$, while projection onto principal components makes the fitted-model, score-risk and generated-law distributions independent of ambient dimension. In the covariance-matched experiments, the measured partial-trace overlaps are consistent with the proposed spectral transition. A width-8 network trained with clean-point reuse shows a $d^2$-like training-set threshold over $d=64$--$192$, with fixed validation and calibration overhead.
Prior-data fitted networks are typically trained using synthetic data generated by a single, manually constructed generator. In cases where such a generator combines multiple components, e.g., in TabPFN, Mitra, or TabICLv2, this is achieved through a mixture using manually specified weights, the impact of which on the network only becomes apparent after pre-training. Instead of such a hand-built generator, this work proposes a structured prior: a small set of named component priors, linked by operators whose algebraic properties are formally proven and whose influence on the trained network has been validated. Within a measure-theoretic framework, the operators of mixture, convolution, amplitude scaling, local transform, and projection are defined. Their governing laws are proven. These operators are implemented in the macl2l software package. Experiments demonstrate that the implemented generators reproduce the identities within the resolution of a permutation test. Furthermore, it is shown that trained networks obey the distributive law; that a mixed prior composed of small components performs as well as or better than any individual component in seven out of eight open regression benchmarks; and that on two of these tables a structured prior reaches, with a ninth of the data budget, an error that a generic random-function prior does not reach even with the full budget. A theorem on model averaging demonstrates that networks trained on individual components constitute the Bayes predictor for the mixture, provided that the component networks are themselves the Bayes predictors for their respective priors. However, the pre-registered test of this averaged predictor with trained component networks failed, because the networks identify the generating component too rarely. The paper proposes a research program focused on the algebra of structured priors for prior-data fitted networks.
Learn-then-Test (LTT) provides a general framework for certifying deployment policies against a prescribed risk target, but its standard formulation targets a fixed population risk estimated from calibration losses. In sequential deployment, losses may instead be dependent and nonstationary, so the risk of a fixed policy can evolve with the observed history. We introduce StreamLTT, which replaces this fixed population risk with average predictable risk, defined by conditioning each decision's loss on the information available before its outcome is revealed. Conditional centering turns empirical deviations from this target into a martingale sum, enabling both a test that adapts to predictable variance and a finite-sample Hoeffding alternative. Combined with family-wise error rate control, StreamLTT yields a simultaneously certified policy region that supports subsequent deployment selection. Across seven forecasting datasets and seven deployment risks, StreamLTT places the certified frontier substantially closer to the prescribed risk target than IID, robust-variance, block-bootstrap, and concentration-based alternatives, while empirical family-wise error closely follows the prescribed level. A finite-sample ablation further shows that the predictable-risk formulation drives the tighter risk frontier, while adaptation to predictable variance makes more effective use of the available family-wise error budget. StreamLTT therefore extends risk-certified policy selection to dependent deployment streams by addressing both what risk should be certified and how it should be tested.
Transolver has recently gained widespread adoption for solving diverse physical problems on unstructured geometric meshes by leveraging its Physics-Attention mechanism, which seamlessly extends standard softmax attention to handle physical domains at scale. For time-dependent problems, existing literature relies on time-conditioned or autoregressive rollouts to preserve temporal causality while retaining Physics-Attention as the spatial backbone. However, a fundamental limitation of this formulation is its sequential execution model, wherein computational costs scale linearly with the temporal horizon, rendering standard Transolver inefficient for transient problems despite its exceptional spatial efficacy. This bottleneck has historically hindered its broader application in time-dependent physical modeling. To address this limitation, we introduce TB-Transolver (also referred to as Temporally Bundled Transolver), an efficient framework built on temporal bundling that scales Transolver for time-dependent PDEs. Furthermore, we propose \emph{spatiotemporal coherence} (${\rm C(K)}$) as a general metric to systematically quantify snapshot correlation within a temporal bundle, enabling the \emph{a priori} selection of optimal bundle widths using training data alone, thereby avoiding expensive grid searches. We demonstrate the computational efficacy of TB-Transolver across four physically distinct domains, namely, plasticity, fluid dynamics, reaction-diffusion, and global weather forecasting; achieving training speedups of up to $29\times$ over conventional autoregressive and time-conditioned baselines. Finally, these computational gains are further amplified through distributed multi-node and multi-GPU parallelism, yielding aggregated speedups of up to $45.6\times$ at optimal bundle widths for systems exhibiting high spatiotemporal coherence compared to single-GPU implementations.
Conformal prediction (CP) provides marginal coverage guarantees, but marginal coverage does not determine where prediction errors occur on a graph. Queries from different graph neighborhoods can exhibit substantially different levels of uncertainty even when aggregate coverage remains near its target. We introduce GLocalCP, an online CP framework for inductive node classification and edge prediction that combines marginal and graph-local calibration with uncertainty-adaptive prediction sets. Specifically, GLocalCP puts forth a novel uncertainty-adaptive nonconformality score that relies on the mutual information obtained from Monte-Carlo (MC) dropout based prediction. Then, it augments a global online threshold with a kernel-based local correction that propagates coverage feedback among similar graph queries. On the theoretical horizon, we establish pathwise $O(N^{-1/2})$ marginal calibration error without exchangeability or stationarity and, for the exact-kernel formulation, an explicit $O(N^{-1/2})$ graph-localized calibration guarantee. Experiments on seven node-classification and seven edge-prediction benchmarks show that similar marginal coverage can conceal substantially different local reliability. Our proposed GLocalCP maintains coverage near the target while consistently reducing graph-local calibration error. Uncertainty scaling in GLocalCP further reduces prediction-set size across all evaluated node and edge datasets while leaving marginal coverage essentially unchanged.
We study the problem of average reward Multi-Agent Reinforcement Learning (MARL) where agents interact within a network. Each agent in the network conducts local updates with information within its $k$-hop neighborhood and collaboratively maximizes the overall reward of the entire network. We first provide impossibility results in achieving convergence to the global optimal joint policies in our setting. These impossibility results highlight the challenges of decentralized learning with local information in multi-agent systems, namely the issues of multiagency, where each agent independently optimizes its policy, and partial observability, where agents have limited visibility into the global state. Given these challenges which indicate that decentralized policy optimization can generally be certified only up to stationarity, we provide finite-sample convergence guarantees for a decentralized actor-critic algorithm with linear function approximation that converges to an approximate stationary point with small gradient norm. To overcome such impossibility results, we further show that adding fixed-state entropy regularization reshapes the objective landscape which leads to even stronger global convergence guarantees for our proposed actor-critic algorithm.
Recent work has improved computer-use agents (CUAs) by combining retrieved task-relevant tutorials and look-ahead planning using world models with MLLM judges. However, the contribution of each component -- tutorials, the judge, and the world model -- to overall improvement has not been evaluated in isolation. To fill this gap, we first construct OSWorld-Tutorial and ScienceBoard-Tutorial: two datasets for CUAs with curated multimodal tutorials, covering everyday and scientific domains. We then propose RAG-PolicyJudge (RAG-PJ), a policy-judge framework for CUAs conditioned on task-specific tutorials: the policy proposes candidate actions, and the judge selects among them. To isolate the effect of world models given tutorials, we compare RAG-PJ with and without common world models: using either zero-shot text-based descriptions of next states or an oracle that returns ground-truth visual observations. We find that RAG-PJ gains little benefit from these rollouts when tutorials are present, even with perfect visual futures. This suggests prior world-model gains in tutorial-conditioned methods come mainly from tutorial understanding and the judge's ability to select the proposed actions, rather than from predicted rollouts. Error analysis reveals that the bottleneck is not rollout fidelity, but the judge's ability to reason over future states and recognize task completion. This highlights the need to improve not just next-state prediction in world models, but also reasoning over their predicted rollouts. Datasets and code are available at https://github.com/realjoenguyen/RAGPJ-official.
Knowledge distillation is a technique used to transfer the expertise of large teacher models into small, efficient student models. Recent advances have integrated the distillation of internal activations, more directly training students to learn similar representations as the teacher. One of the core challenges of internal distillation is the dimensional incompatibility between the student and teacher internal representations. Existing methods add learnable matrices to project the larger teacher activations into the smaller representation space of the student. This is a notable point of inefficiency, as learning additional parameters for the projectors increases training costs, only for them to be discarded immediately after training. We introduce RAPID (RAndomized Projection for Internal Distillation) a novel distillation methodology which eliminates the wasteful projection learning by leveraging randomized methods. RAPID utilizes an efficient transform construction based on the Johnson-Lindenstrauss lemma to compress the teacher representations into the student embedding space with strong theoretical guarantees of structural preservation. RAPID also employs a combination of exponential moving average (EMA) and homoscedastic uncertainty weighting in a novel multi-objective function designed for internal distillation. Our experiments in post-training for reasoning demonstrate that RAPID increases training efficiency by as much as 21.1% while maintaining performance comparable to existing methods.
Selecting small, effective instruction-tuning datasets requires identifying which aspects of the original data to retain. We introduce Follow the Features (FtF), a data-selection method that uses sparse autoencoder (SAE) features to characterize training data and guide subset construction. FtF ranks features using dataset-level activation statistics, scores samples by their expression of highly ranked features, and incorporates feature resampling to encourage diversity. The LLM-SAE pair used for selection is independent of the model being fine-tuned, enabling reuse across target models. We evaluate FtF on two model-dataset settings at a training budget of 100 samples with a fixed sequence length of 256 tokens. FtF achieves performance competitive with GRAPE and S2L, with relative rankings varying across settings. On Qwen 3 8B trained on Alpaca, FtF exceeds full-dataset training by 1.59 percentage points in the harmonic mean of an AlpacaEval-based win rate and Open LLM Leaderboard performance, although full-dataset training retains a higher Leaderboard score. These results support SAE feature statistics as a useful signal for data-efficient instruction tuning.
Individual-level trajectories are essential for applications that rely on user-level mobility behavior. Existing mobility-generation approaches face a fundamental tension between individual-level and population-level statistical fidelity: top-down models reproduce aggregate mobility statistics but struggle to guarantee realistic individual trajectories, while bottom-up generators synthesize individual trajectories yet often fail to preserve key population-level mobility regularities after aggregation. In particular, many bottom-up generated populations fail to reproduce the displacement distribution, whose heavy tail is well summarized by a truncated power law. We formalize this phenomenon as the \textbf{Population-Level Displacement Gap}. To investigate its prevalence, we conduct a systematic evaluation of representative model-driven and data-driven trajectory generators. Our results provide evidence that the Population-Level Displacement Gap is prevalent across diverse generation paradigms. To diagnose this gap, we introduce the \textbf{Scale Fidelity Profile (SFP)}, a scale-decomposed metric that quantifies discrepancies in trip allocations across spatial scales. SFP indicates that population-level displacement mismatches co-occur with systematic errors in how trips are allocated across spatial scales. Motivated by this observation, we develop \textbf{ScaleTraj}, an autoregressive Transformer that decomposes location prediction into a Scale Selection Head that decides the movement scale and a Scale-Aware Cell Selection Head that predicts the destination conditioned on the selected scale. Experiments on two datasets show that ScaleTraj reduces the Population-Level Displacement Gap versus model-driven and data-driven baselines, while preserving fidelity. Our code is available at https://github.com/SCU-TrustworthyAI/ScaleTraj.
Routing benchmarks often judge a router against a hindsight oracle that counts a query as solved whenever any model's recorded answer is correct, but under stochastic decoding that oracle mixes two different objects: the clairvoyant ceiling of committing to one model per query and the premium of calling several models and keeping any correct answer. A companion study measured this premium at 3.5 percentage points (pp) for 11 small open models on MATH-500; how large it is once current reasoning models enter the pool was unknown. We measure it with a fixed-protocol, pre-committed generation on a single NVIDIA B300: all 500 MATH-500 problems, 30 seeded samples per problem, Llama-3.3-70B-Instruct and NVIDIA's Nemotron Nano 9B v2 and Nemotron Super 49B v1.5 (a NAS derivative of the same 70B) under vendor sampling, a 65,536-token cap, a pinned Math-Verify scorer, and an append-only attempt ledger, for 45,000 scored responses. The two reasoning models both reach 96.1% accuracy against 76.2% for the 70B baseline. The empirical product-coupling premium is 0.215-0.588 pp across all two- and three-model pools; the empirical plug-in selection headroom over always calling Nano is 1.27 pp; and resampling Nano three times gives practically the same empirical any-correct rate as calling all three models once. Only 31 responses reached the cap, 25 of them with exactly periodic tails, yet the capped responses account for up to 60.2% of a model's output-length-squared proxy, and long generations made a fourth planned model, Llama-3.1-8B, too slow to finish. The premium is far below the companion's 3.5 pp, but models, pool size and protocol changed together; an oracle gap should therefore be reported with the pool, the protocol, the samples per cell and the decoding budget that produced it.
Optimizing accelerator code across hardware architectures is difficult. ML compilers search predefined spaces of schedules and micro-kernels, while LLM-guided optimization can explore alternative algorithms, operator decompositions, and hardware-specific techniques. However, existing LLM agents typically target a single programming interface, toolchain, or accelerator family, focus on kernel-level optimization without model-level transforms, and offer limited reuse of experience across hardware targets. We present ModelBlaster, a Memory-Augmented In-Context Reinforcement Learning (MAIC-RL) framework for optimizing data-parallel kernels and whole models across heterogeneous SIMD and SIMT accelerators. Its mutable policy is a persistent knowledge base updated at inference time by Actor-Critic Textual Gradient Descent (AC-TGD). The knowledge base maps performance states to actions, outcomes, and constraints, enabling experience reuse across kernels, models, and hardware targets. A model-level optimizer uses kernel-level knowledge to inform whole-model decisions. The same optimization loop and knowledge schema support CUDA on NVIDIA GPUs, OpenCL on Qualcomm GPUs, and RVV intrinsics on RISC-V vector cores, with only lowering and measurement layers specific to each target. Compared with PyTorch, ModelBlaster achieves geometric-mean speedups of 1.85x, 5.58x, and 5.18x on KernelBench Levels 1, 2, and 3, and 1.69x on SOL-ExecBench Levels 1 and 2 on a B200 GPU. On Qualcomm Adreno GPUs, it achieves 5.8x over the vendor SDK and 1.7x over Apache TVM; on RISC-V vector cores, it achieves 2.7x over ExecuTorch with XNNPACK across 70 KernelBench problems.
What makes a pretraining corpus transferable? We describe a corpus by three properties: its structure, its dimensionality, and the distribution shape of each unit's values. We propose the Transfer-Invariance Principle: the same property can be irrelevant to transfer under one input transform and decisive under another, so the model's input transform, not the data modality, decides which properties can matter. To test it we build TransferForge, a generator that places a synthetic population at any prescribed structure, dimensionality, and distribution shape. KernelSynth, the synthetic prior of Chronos, cannot, because it draws independent series. When the model standardizes each input window (PatchTST on UCR), the distribution shape is inert and the structural setting governs. Synthetic pretraining then matches a real corpus of the same size and exceeds it once scaled. When the model encodes calcium fluorescence signals without normalization (POYO Plus), the shape becomes a governing constraint, and a shape-matched synthetic corpus replaces real pretraining with zero real neural activity. Flipping only the input normalization, with the model and data fixed, changes the effect of a shape manipulation by a factor of 5.3, so the pipeline, not the modality, sets the switch. The principle yields a recipe that chooses what to synthesize before any training. Pretraining data then becomes something to manufacture for the model, most valuably where real data is scarce, expensive, or private.
Model merging combines multiple fine-tuned models without additional training, and arithmetic methods have emerged as a simple yet effective way to merge models without data. However, existing arithmetic methods treat task vectors in parameter space and apply a single element-wise rule to directions that can differ in mergeability. In a shared directional basis, we find a consistent two-regime structure: a low-dispersion common bulk shared across tasks and a high-dispersion task-dominated tail that carries disproportionate energy. These findings motivate Common--Unique Subspace Decomposition (CUSD), a data-free pre-merging method that separates task vectors into common and unique directional components, merges the common components with downstream arithmetic operator, and retains unique residuals separately. Experiments across $11$ base models, four model families, and four arithmetic-based merging operators show that CUSD improves 42 of 44 model--operator combinations and achieves up to 22.5% improvement over the corresponding baselines. Our code is publicly available at https://github.com/stwang666/CUSD
Under temporal nonstationarity, model performance rankings reverse across training horizons due to heterogeneous sensitivities to memory. While existing literature addresses adaptive lookbacks, effective model complexity, and empirical forecast evaluation in isolation, no unified framework explains the structural mechanics driving these rank reversals. We bridge this gap by deriving a model-agnostic decomposition of out-of-sample $R^2$ into target alignment, miscalibrated aggressiveness, and predictive instability. Within this framework, we mathematically characterize complexity phase boundaries, establish closed-form optimal-memory equations for ridge regression under local drift, and extend these guarantees to complex model architectures via surrogate bounds. We validate these theoretical dynamics by benchmarking 14 linear, tree-based, and neural architectures across large-scale financial panels. In doing so, our framework demonstrates that empirical model rank reversals are driven by predictable surges in miscalibrated aggressiveness, providing a rigorous mechanistic diagnosis of generalization under temporal shift.
Looped transformers apply a shared block several times per token to improve accuracy without adding parameters. For models such as Ouro that keep a separate key-value (KV) cache per loop, cache memory limits the batch size that amortizes repeated weight reads. Uniform-bit quantization and fixed rules per loop do not allocate precision according to each layer's sensitivity in each loop. Measurements on Ouro show that quantization sensitivity varies across (loop, layer) cells, and the sum of individual cell damages approximates the overall output damage. Sharing damage also varies by cell, with later loops generally tolerating sharing better. Skipping low-attention values to save reads produces errors that compound across loops. LoopKV uses a one-time offline profile to choose 4-bit storage, 2-bit storage or reuse of the final loop's KV for each cell. The planner minimizes reads under budgets for predicted damage and storage. During decoding, each loop reads the high 2 bits of each 4-bit value and fetches the low bits only where its own attention weights require 4-bit precision. A single fused kernel in vLLM executes the plan. Serving Ouro-1.4B on one RTX PRO 6000, the memory-optimized LoopKV-M plan stores 35% less KV than uniform 4-bit quantization (KIVI-4) on the same fused kernel and CUDA graph. It sustains 1.7-2.2x KIVI-4's goodput under a latency service-level objective with a 16K-token maximum length. None of its averages on six public benchmarks falls significantly below BF16's.
Diffusion models represent a powerful class of generative models known for their solid theoretical foundations and remarkable performance across diverse tasks and domains. Existing node level diffusion methods synthesize nodes and edges within a graph, but are not designed to learn a robust generator from adversarially corrupted graphs. In this study, we introduce the Robust Graph Diffusion Model (RGDM), designed to generate labeled synthetic graph structures consisting of nodes and edges that integrate seamlessly into a given graph. The RGDM consists of a Robust Graph Autoencoder (RGAE) and a Latent Diffusion Model (LDM). Leveraging an edge selection mechanism and an innovative low-rank regularization on the latent feature, the RGDM produces clean and high-quality synthetic graph structures, even when trained on graphs subject to adversarial attacks. Comprehensive experimental evaluations reveal that Graph Neural Networks (GNNs) trained on the augmented graph, which is formed by merging the original attacked graph with the synthetic graph structures, exhibit significantly improved robustness against various graph adversarial attacks in the context of semi-supervised node classification. The code of the RGDM is available at https://github.com/Statistical-Deep-Learning/RGDM.
Temporal-difference (TD) learning trains neural representations using bootstrapped targets generated by evolving value estimates, yet how predictive features emerge and propagate under incomplete neural updates remains poorly understood. We investigate this problem using finite-width two-layer ReLU critics on a constructed family of smooth stochastic horizon-$H$ MDPs with unknown quadratic structure and $d$-dimensional observations. We first show that an informative TD update acquires the quadratic Bellman signal and yields a representation admitting a bounded predictive readout. The simultaneous hidden-weight and head update achieves a strict finite-sample prediction-risk improvement over matched head-only SGD, with a $2(d+1)$-fold larger population quadratic response. We then establish that an imperfect learned continuation can already provide informative supervision to an earlier zero-reward stage. By tracking coupled neural updates with one SGD step per target copy, without requiring intermediate Bellman regressions to be solved, we obtain a uniform prediction-error transient $\exp[-\Omega(K/(Hd))]$, up to action-independent shifts and explicit sampling and finite-width errors. On the constructed signed-action MDP family whose total reward is bounded by 1, we show that an exploration-then-commit algorithm can achieve $\widetilde{O}(d^3H^{1/3}T^{2/3})$ regret over $T$ episodes, using $K=\widetilde{O}(Hd)$ global updates. Our analysis connects bootstrapped feature acquisition, concurrent neural refinement, and finite-budget policy performance.
Reasoning vision-language-action (VLA) models pair physical control with natural-language rationales that supervisors can read to monitor embodied safety. Prior attacks on VLA reasoning corrupt the rationale itself and leave visible evidence in the text. We show that visual perturbations can instead alter the physical action while keeping the generated rationale word-for-word identical. At the shared text/action interface of Alpamayo for driving and InstructVLA for manipulation, our text-neutral action-subspace (TNAS) diagnostic shows that the locally most action-sensitive direction is nearly orthogonal to the active text-loss gradients, with 1.87% and 5.28% norm overlap. At equal hidden-state perturbation norms across both models, interventions along this text-neutral residual shift actions 4.64x to 47.8x more than text-gradient perturbations, whereas text-gradient perturbations induce 21.2x to 277x more text drift. Steering visual perturbations along this direction reaches the same decoupling from the image. Across 154 obstacle-annotated driving scenes in Alpamayo, TNAS-guided attacks induce a new collision under a word-for-word identical rationale in 52.6% of scenes under full-image perturbations and 35.1% under a localized forward-camera patch, compared with 4.55% for unguided action attacks. Relative to a text-preserving baseline, TNAS guidance raises exact rationale preservation from 83.0% to 89.5% and cuts teacher-forced text drift by 68.4% at a comparable collision rate. In InstructVLA, TNAS guidance cuts text drift by 40.5% and preserves the exact response on 85.0% of examples alongside a 1.73-unit action-chunk shift. An unchanged rationale therefore cannot certify a safe action, so monitors for reasoning VLAs must verify the predicted action channel directly.
Wearable sports sensors can recognize actions on the device itself, where lower-precision numbers reduce model storage and memory traffic. Two 4-bit block-scaled formats, MXFP4 from the Open Compute Project, introduced in 2023, and NVFP4 from NVIDIA, introduced in 2025, are drawing interest. Published evaluations cover general deep-learning benchmarks and large language models; to our knowledge, nobody has yet measured what they cost a small wearable-sensor model. We train a 159,492-parameter 1D CNN to recognize four basketball actions from wrist-accelerometer data in Hang-Time HAR and test it on players it has never seen, using five-fold cross-validation grouped by player. Every comparison between number formats is made fold by fold. In full precision the model reaches 69.26\% macro-F1. Measured TFLite FP16 and INT8 models lose at most 0.16 percentage points. In simulation, 4-bit weights cost 1.45 points with MXFP4 and 0.72 points with NVFP4. When activations are quantized to 4 bits as well, MXFP4 loses 5.14 points and NVFP4 only 0.75, and NVFP4 comes out ahead in all five held-out-player folds. The ranking survives two alternative simulator settings, with NVFP4 leading by 3.49 to 4.40 points. A $2 \times 3$ ablation with 4-bit weights and activations (two block sizes and three scale rules) points to scale selection as the main source of MXFP4's loss. The reference MXFP4 conversion we simulate rounds the scale down, which can pick scales that are too small, so many blocks clip their largest values. Rounding the scale up recovers 3.18 points with 32-value blocks and 3.63 points with 16-value blocks, and switching to an 8-bit floating-point scale, as NVFP4 does, adds another 0.84 and 1.20 points. Block size matters less. Under this round-down rule, 16-value blocks score 0.44 points below 32-value blocks in all five folds; under the other two rules there is no consistent effect.
Deployed language models are updated continually, and regulators increasingly accept changes that were declared in advance. Showing that an update changed behavior only where it was declared to is a different statistical problem from bounding how much it changed: a useful update is meant to change some decisions, so global disagreement and regression bounds cannot certify it. We formalize the problem with declared change envelopes. Before training, the developer commits to a region of inputs, and every decision that changes outside it counts as leakage. Leakage compares two models, so it can be certified from unlabeled traffic with anytime-valid tests whose sample cost we bound from above and below, and per-update certificates compose into a bound on lifetime drift outside all declared scopes. We measure leakage for LoRA and full fine-tuning, preference optimization, reinforcement learning, and knowledge editing on models from 1.5B to 72B parameters. Narrow updates leak widely: domain and policy LoRA updates keep only about 7-36% of the decisions they change inside the declared scope. Across 256 trained variants, leakage tracks the update's KL divergence to the incumbent on out-of-scope inputs (Spearman 0.94). None of the customary remedies we tested makes a rule-like update certifiable without giving up much of its intended change. Distilling the incumbent's decisions on unlabeled prompts in the deployment format does: anchored LoRA, GRPO, and (at a small learning rate) full fine-tuning updates certify 5% leakage from a few hundred to a few thousand unlabeled samples, and the certificates replicate on held-out traffic and on written answers. The certificates cover decision-like outputs only; on open-ended chat, re-batching alone changes 11-21% of responses and no update certifies. Code: https://github.com/Usama1002/declared-change-envelopes
How much label noise has a network memorized, and when? We study a multilayer perceptron and what statistical complexity adds to a description of its predictions. Standard identities decompose predictive information into signal, memorization, and heterogeneity. The memorization term vanishes in the population under the stated noise-independence condition. A Fano-type bound relates it to training accuracies, but the memorized fraction M provides no upper bound. Across 89 MNIST runs with ten pre-specified hypotheses (five supported, four unsupported, one not testable), normalized memorization information rose with M on training samples, remained near zero on held-out samples with fresh noisy labels, and was close to its Fano bound. Unlike raw M, it has a zero population null under the independence condition and chance correction largely removed this difference. When every label was replaced by an incorrect class, test accuracy fell to 0.5%, yet predictions retained information about the true class. Mean statistical complexity peaked before memorization at corruption rates up to 40%. At higher rates, its peak fell within the memorization phase, near the epoch when mean training confidence reached 0.6 and the two could not be separated. As a stopping rule, it matched established rules only at low noise and detected corrupted samples less well than simple label-free scores. An exploratory extension of 280 runs covering Fashion-MNIST, pair-flip noise, and four training variants yielded similar results under uniform noise, but not pair-flip noise, where complexity peaked within two epochs. Conclusions are limited to this architecture, two small datasets, and synthetic noise.
Predicting stock returns requires modeling persistent company relations and time-varying market signals. Graph Neural Network (GNN) approaches combine static priors with dynamic relational graphs, but treat every extracted edge as a trustworthy binary fact, ignoring heterogeneous reliability across relational sources. We propose LENS (LLM-Extracted priors and News-driven Semantics), a static--dynamic graph framework for stock return prediction and ranking that models relational reliability as a first-class signal across the pipeline. On the static side, we extract a multi-relational company graph from a Large Language Model and aggregate it over multiple runs with confidence weighting to downweight low-consistency edges. On the dynamic side, we build a news-driven semantic graph with exponential-decay temporal aggregation to capture changing market context under uneven coverage. A context-dependent gated fusion module rebalances the two graphs at the per-stock, per-day level. On S&P 500, LENS improves AR@5 by 16.4% and SR@5 by 22.6% over the strongest baseline, and on NASDAQ 100 it achieves comparable improvements (15.5% and 20.9% respectively). Ablations isolate each component's contribution, and a human-validation study of sampled relations supports that the aggregation is grounded, providing evidence that reliability-aware LLM relations and news-based semantic graphs carry complementary information for graph-based financial prediction.
Low-rank adaptation (LoRA) has become a widely used parameter-efficient fine-tuning method for large pretrained models. However, standard LoRA typically assigns a fixed rank profile despite layer- and stage-dependent adaptation needs, while recent adaptive-rank variants often rely on heuristic or first-order signals that provide only an indirect link between rank reallocation and local loss behavior. This leaves the rationale behind each rank reallocation insufficiently interpretable. Motivated by this gap, we propose LIRA (Loss-Informed Rank Allocation), a dynamic-rank LoRA method designed to make online rank reallocation more interpretable. LIRA scores each layer by the estimated magnitude of the local loss perturbation associated with its recent LoRA-factor movements under a second-order approximation, making the basis for each rank change interpretable term by term. By periodically applying DELTA (Displacement-Evaluated Loss-Trajectory Allocation), a displacement-based algorithm for LoRA rank reallocation, LIRA produces discrete effective ranks under matched budgets while remaining compatible with standard (A,B) adapters. On RoBERTa-base GLUE under the La-LoRA/LoRA protocol, LIRA outperforms the uniform-LoRA and adaptive-rank baselines reported under this protocol; Trajectory EMA attains performance competitive with Fixed-point Fisher at lower estimation cost.
Standard spatiotemporal graph convolutional networks need a fixed-size adjacency matrix, coupling model weights to a specific sensor count N; they cannot produce any output at a sensor absent from training. We present OmniST, a topology-invariant forecaster whose computation has zero structural dependency on the number of nodes, and which, composed with inductive split-conformal calibration, gives distribution-free, coverage-guaranteed prediction intervals at arbitrary (including never-seen) sensor locations, preserving that guarantee as sensors are added or removed. Topology invariance follows from coordinate-based tokenization, k-nearest-neighbor attention from geographic coordinates at each forward pass, and per-node temporal processing via a bidirectional GRU. On real Delhi air-quality data (37 stations, 2019-2023), OmniST attains a 30-day PM2.5 MAE of 31.0 ug/m3 and degrades at most 1.1 percent when 40 percent of stations are removed. Its inductive forecasts at unseen stations beat nearest-neighbor interpolation (about 12 percent) and a learned kriging GNN (IGNNK, about 21 percent), and, unlike either, carry a finite-sample coverage guarantee. The guarantee holds at 90 percent or above under 30-50 percent node dropout across four real datasets over three domains (Delhi and Beijing air quality, NOAA GHCN weather, Caltrans PeMSD7-M traffic), and still holds at 325 real sensors (full PeMS-BAY). We also fix a temporal-drift failure: static conformal under-covers over time (about 0.50 late), but a rolling adaptive-conformal estimator restores target coverage (0.89 overall, 0.96 late) at the cost of wider intervals.
As deep learning (DL) models become increasingly complex and hardware architectures grow more heterogeneous, DL compilers play a critical role in optimizing computational efficiency and enabling scalable deployment. This paper presents a systematic analysis of DL compilers, covering their architectural evolution, compilation workflows, optimization techniques, and emerging research directions. We organize DL compiler architectures into frontend, middle-end, and backend stages, emphasizing the role of the middle-end in connecting high-level model representations with hardware-specific optimizations. We further examine representative DL compilers and recent optimization techniques across different compilation stages. Beyond architectural and technical analysis, we conduct an empirical evaluation of compiler behaviors to investigate the effectiveness of graph optimization, compilation overhead, and dynamic-shape handling. Experiments using PyTorch/TorchInductor on an NVIDIA RTX A6000 demonstrate that compiler benefits depend strongly on workload characteristics, with speedups of up to 8.21\times for fusion-friendly computation graphs, while other workloads exhibit limited improvements. Our results further reveal that compilation overhead can require thousands of executions to amortize, and dynamic-shape compilation reduces cumulative shape-transition costs by approximately 70\% at the expense of modest steady-state performance degradation. By combining systematic architectural analysis with quantitative experimental evidence, this work provides insights into the design, optimization, and practical deployment of modern DL compilers.
Magnetic graph networks encode direction through phase, but a learned graph can erase the intended directional mechanism before prediction. We characterize this failure for support-based magnetic Laplacians. Exact identities relate imaginary energy and distance from zero phase to amplitude-weighted reciprocity. Top-$k$ selection forces complete collapse when the observed-node count satisfies $m\leq k+1$, and dense support imposes a lower bound on reciprocal-arc fraction. We then construct paired propagation examples with identical node features and symmetric amplitudes but opposite labels, giving a provable chance-level limit for direction-erasing representations. In 72 fixed neural fits, support phase achieved test AUROC 1.0000 on a unilateral ring but 0.5000 on reciprocal ring and complete supports; continuous score phase and a real directed comparator achieved 1.0000 throughout. All models solved the direction-free control. These are oracle-direction mechanism tests, not evidence of learned clinical direction. In the preserved public DeepDRiD development study, continuous phase was active but its robustness difference from zero phase was only $+0.0036$ (descriptive 95\% interval $-0.0014$ to $+0.0095$) on validation patients used for checkpoint selection. The contribution is a structural and experimental framework for distinguishing erased direction from unestablished task utility. It provides a precise circumstance in which magnetic phase preserves useful information, without asserting universal or clinical superiority.
Aggregate scores for continual learning algorithms rest on a distribution over task orders that is rarely stated. We formulate algorithm comparison as an expectation over a declared order distribution, which separates training variation from order variation, and apply it to motion forecasting with four road-layout tasks from the INTERACTION dataset, all 24 task orders, four algorithms, and three training replicates (288 runs). Task order accounts for 65\% to 91\% of the variation in current-task error and backward transfer, and 90\% to 99\% of this order effect comes from the identity of the starting task: the smallest task cannot be learned from a cold start. Beyond the first stage, the order effect is comparable to training variability. In most orders, the leading algorithm in current-task error and backward transfer is within one replicate standard deviation of the runner-up. Backward transfer depends on the order mainly through its reference term, which weights the first task most heavily, so relative improvement and final error identify different leading algorithms for displacement error. A transition-direction descriptor computed from task statistics tracks the position of the smallest task, and weighting orders by it leaves the algorithm ranking unchanged. These results call for benchmarks that declare their order distribution and report order-resolved, starting-task-resolved, warm-start, and final errors alongside aggregate scores. The code and data are available \href{https://anonymous.4open.science/r/task-order-cl-4f7b/}{here}.
We introduce a training-free sparse attention method that builds a prompt-specific key index to accelerate both prefill and decoding. The method clusters the keys into large semantic blocks that group tokens from noncontiguous positions. Each query position selects blocks by scoring their centroids and attends to the original keys and values within them. Queries that select the same block are processed together, so prefill stays fast even though each query chooses its own blocks. Large blocks make centroid scoring and this grouped processing cheap, so each query can attend to more tokens in the same time. Larger blocks with a larger token budget can thus be both faster and more accurate on long-context tasks than smaller blocks with a smaller budget. We evaluate two hybrid LLMs, Qwen3.5-4B and 35B-A3B, on one NVIDIA H200. Including the cost of building and using the index, whole-model wall-clock speedups over dense attention are 2.3-2.4x for prefill and 1.7-2.0x for decoding at 512K tokens, and 2.9-3.0x for prefill at 1M. The 35B-A3B model stays close to dense attention on five RULER tasks through 1M, on LongBench-v2 overall, and on MRCR through 256K. The 4B model loses accuracy mainly on multi-key retrieval, yet outperforms dense Qwen3.5-2B, whose latency is similar at 512K, on all three benchmarks.
Parameter-efficient linearization turns a pretrained autoregressive Transformer into an efficient model by hybridizing sliding-window attention (SWA) with a constant-size recurrent memory. But does every layer's recurrent state contribute to long-context retrieval? Investigating an SWA and Test-Time Training (TTT) hybrid, we find that retrieval is sparse across depth, concentrated in a few anchor layers, and that the remaining states hold similar information expressed through layer-specific transformations. We exploit both observations in Anchor: TTT memory placed only at the anchor layers, and a cross-layer recurrent state written once and read by many layers, each through its own learned read map that realizes a layer-specific stretch and rotation, along with demonstrated generalization to other hybrid attentions. On Llama-3.1-8B and Mistral-7B, Anchor delivers 2.3-3.1x higher recall and 6.7-8.9 points higher MMLU accuracy than LoLCATs with a 10x smaller recurrent state, retains 91-95% of the teacher's MMLU accuracy, and cuts end-to-end latency by up to 4.3x relative to softmax attention.
We give a proper constant-factor agnostic learner for affine halfspaces under full-dimensional log-concave marginals. For arbitrary binary labels, the algorithm uses independent labeled samples and returns a single affine halfspace with error at most $C\mathrm{OPT}+\varepsilon$. Its sample and real-arithmetic complexity are polynomial in dimension, inverse accuracy, and the logarithm of the reciprocal failure probability. The mean and covariance of the marginal need not be known, and the sample complexity can be made linear in this logarithmic confidence factor. The main geometric result compares global disagreement with disagreement on a balanced restriction: for a minority halfspace of mass $p$, global disagreement is within universal constant factors of $p$ times conditional disagreement, provided the latter is sufficiently small. To construct the restriction from noisy labels, we estimate the minority cap's mean and covariance robustly, approximately minimize the classical zonoid-depth dual objective, and expand a superlevel set to a marginal quantile. The disagreement comparison cancels the noise amplification caused by conditioning and allows the learned halfspace to be used on the entire space. The analysis combines degree-four sum-of-squares moment bounds for log-concave laws with a translation-based reduction from balanced affine learning to homogeneous halfspace learning.
In this work, we study end-to-end FP4 reinforcement learning (RL) post-training for Mixture-of-Experts (MoE) large language models. Recent work on FP4 RL has introduced effective recovery recipes that restore accuracy under aggressive low-precision quantization for dense models, but we find that the same approach yields no improvement over the naive FP4 baseline in the MoE setting. We conduct an in-depth study to understand why. Our analysis identifies the root cause in the routing mechanism unique to MoE: because the router makes discrete expert-selection decisions, training-inference policy alignment becomes acutely sensitive to whether the same recovery technique is applied symmetrically on the training and inference sides. When an asymmetric recipe is used-including the dense-model approach, which corrects only the rollout-training-inference alignment deteriorates sharply, as measured by the Pearson correlation coefficient between training and inference probabilities, and router selection matching rates. Building on this finding, we propose MoeResQ, a quantization-recovery recipe for MoE RL post-training that enforces symmetry between the training- and inference-side corrections. MoeResQ restores training-inference alignment and substantially closes the accuracy gap to full precision, establishing symmetric recovery as the key principle in FP4 RL for MoE models.
Language models are accurate on most long-arithmetic problems, but their accuracy is estimated on samples and carries no guarantee for an unseen input; tool use is exact but moves the computation outside the network. We show that exactness at all input lengths becomes a finite, checkable property of a network when each of its steps reads only a bounded view of an external working memory. In Sketchpad, Qwen3.5-0.8B with a low-rank adapter executes arithmetic procedures on a rendered sheet. Each forward pass reads a three-cell view and emits one step: a digit, a carry or borrow, and the next operation. A fixed runtime moves the view and writes no digit of a result. We prove in Lean that the model is then exact on all inputs of any length if and only if it is exact on the finite set of views the runtime presents, that no finite test suffices when steps also read earlier steps, and that the guarantee tolerates bounded numerical perturbation and extends to compositions of procedures. All 2,446 views pass on the deployment device, a laptop GPU running the model in a web browser. The certified model performs signed addition and subtraction, long multiplication and division, square roots and greatest common divisors at arbitrary length. It solved 200 of 200 pre-registered problems in the browser, and 220 of 220 sums of up to 1,000 digits and products of up to 20 digits, of which GPT-6.1 Sol, the strongest of ten other models, solved 217. A symbolic runtime built on these procedures solves, expands and factors polynomials with every number computed by the model, and a multiplicative gate on the adapter recovers the base model exactly when the sheet is not in use.
Among 291 surveyed open-weight language model repositories (280 distinct embedding tables), 52.6% of untied gated and hybrid checkpoints (20/38) contain reachable input embeddings with norms below $10^{-6}$, compared with 0% of weight-tied models. Such near-zero rows can arise during pretraining when unseen rare or template tokens receive zero input gradients while AdamW weight decay shrinks untied embeddings. In tested Transformer configurations, non-initial positions or a BOS token prevent this failure mode. In gated sequence models (Mamba, Gated DeltaNet, xLSTM), a near-zero token can amplify input-embedding gradients by more than 30 orders of magnitude (up to 35). At zero input, current-token readout factors (SiLU gates or linear queries) vanish, suppressing historical-state readout. Residual RMS values can then fall below the square root of the RMSNorm floor across consecutive layers, causing Pre-RMSNorm to amplify backward gradients by up to $10^3$ per layer. When a gradient entry exceeds $1.84 \times 10^{19}$, FP32 global-norm clipping in PyTorch, Hugging Face Trainer, and FSDP2 can overflow to infinity and zero the current batch's gradients. In Telugu continual pretraining of Nemotron-3-Nano-4B, this discards updates on approximately 20% of steps and leaves triggered rows frozen. We propose two remedies. For released checkpoints, GroupClip decouples input-embedding clipping using a 64-bit or rescaled norm; its rescaled implementation eliminates update loss and heals triggered rows in one step with negligible measured end-to-end overhead. For pretraining, GateFloor ($\mathrm{SiLU}(z) + \delta$, $\delta \in [0.01, 0.1]$) keeps the gate positive at zero input, reducing near-zero gradients by 28 orders of magnitude in experiments up to 370M parameters, with perplexity comparable to the baseline.
Low-rank adaptation (LoRA) has attracted significant attention in the parameter-efficient fine-tuning (PEFT) landscape. However, its design suffers from two fundamental limitations. First, its invariance manifold induces flat directions in the loss landscape, creating an optimization bottleneck for adaptive optimizers. Second, its effective rank scales only linearly with the parameter budget. In this paper, we propose Scalable Morpho Adaptation (SaMA), a PEFT method grounded in a principled generalization of the Kronecker product. By relaxing the shared-block constraint in the perfect-shuffle factorization, SaMA spans from low to full rank and achieves a practical rank that scales quadratically with the parameter budget, up to full-rank. Furthermore, we show that its invariance manifold collapses to a diagonal group, making it structurally more optimizer-friendly. Empirically, SaMA consistently outperforms strong PEFT baselines on commonsense and arithmetic reasoning benchmarks while being more parameter-efficient.
Muon has recently shown strong empirical performance in training Transformer-based large language models (LLMs), yet its optimization dynamics remain poorly understood from a theoretical perspective. In this paper, we analyze the training dynamics of Muon for next-token prediction (NTP) in Transformers. The training procedure consists of two phases: the preprocessing phase learns the output-value component, while the main phase learns the key-query component. For the preprocessing phase, we characterize the optimization trajectory exactly and show that the normalized spectral margin is optimal after each update. For the main phase, we establish nonasymptotic convergence rates for the training loss and the normalized margin gap under softmax attention. These results suggest that Muon's update rule can help models efficiently learn which tokens matter for prediction, offering a possible explanation for its strong performance in large language model training. The learned weight structure also yields a generalization guarantee for unseen contexts. Furthermore, we validate our theoretical insights with experiments on next-token prediction tasks.
Compiling a natural-language agent skill into model parameters removes the skill text from the prompt and leaves behind a small artifact that can be distributed and reused. Existing methods differ in where they inject the compiled quantity, and all of them assume that the artifact carries the content of the skill. We test that assumption. Through three controls on a retrieval agent, we find that a compiled skill shift carries a domain-general behavioural prior, not the content of the skill. A wrong skill document performs on par with the correct one, a single averaged bundle matches per-skill routing under a two-sided equivalence test, and a same-norm random direction collapses to the no-injection baseline, which shows that the trained direction stays load-bearing. We then introduce SKILL2STEER, a one-pass hypernetwork that reads a skill document once and maps it to low-rank activation shifts added at the output of every decoder block. Each skill compiles to 2.36M parameters, 9.25 times fewer than a weight-space adapter, while the base weights stay frozen and the generator runs only at compile time. The shift can be scheduled across retrieval steps and deployed as one routing-free bundle. On SearchQA, a seven-dataset benchmark, SKILL2STEER reaches 37.9 exact match, 9.8 points above the frozen model and 2.3 points above a weight-space method that ships 9.25 times more parameters per skill.
Machine unlearning aims to erase the influence of a designated forget set from a trained model while preserving strong utility on the retain set. Approximate unlearning methods efficiently achieve this goal without full retraining, but their effectiveness is primarily assessed using dataset-level aggregate metrics. These aggregate metrics can mask deviations from retraining at the level of individual samples. In this work, we focus on per-sample unlearning effectiveness. We introduce SALA (Sample-grained Likelihoods ALignment), a plug-in optimization framework that enables approximate unlearning methods to operate at the level of individual samples by steering each example toward its correct target distribution. To quantitatively assess this behavior, we further propose the sample-grained distribution gap, a metric that captures per-sample discrepancies between approximate unlearning and exact retraining output distributions. We provide theoretical guarantees showing that SALA contracts this gap and yields a bounded per-sample discrepancy to retraining. To make this framework practical, we develop a low-cost estimation method for the member and non-member target distributions. Extensive experiments on classification and generation tasks demonstrate that SALA reduces the sample-grained distribution gap by up to $86.67\%$ when paired with existing approximate unlearning baselines, while also improving robustness to unlearning hyperparameters.
Variational Bayesian learning for deep neural networks offers many potential benefits. The difficulty in implementation arises from constraints on the parameters of the variational distributions, as well as the need for low-variance gradient estimators. Our recent work proposes Lie group methods for multiplicative or affine groups that naturally overcome such difficulties. Lie groups are essentially parameter spaces with a binary operator such as addition or multiplication, which allows for natural updates within the parameter space. In this article, we streamline and extend our Lie group framework to any general Lie group, and give an explicit relation to standard natural-gradient variational learning methods such as the Bayesian learning rule (BLR). In particular, we prove that linear approximations of the Lie group algorithm give generalizations of the BLR. The key ingredient is the harmonic exponential family, a general class of Lie group exponential families associated to representations of Lie groups. Our work provides a systematic method to construct from such representations efficiently implementable algorithms for variational learning.
We study the sample complexity of list replicable algorithms for estimating the biases of $d$ coins. An algorithm is $(k,\varepsilon,\delta)$-list replicable if for every bias vector $\mathbf{p}\in[0,1]^d$, there is a set of outputs $L_{\mathbf{p}}\subseteq \mathbb{R}^d$ of at most $k$ elements such that every $\mathbf{y}\in L_{\mathbf{p}}$ satisfies $\lVert\mathbf{y}-\mathbf{p}\rVert_\infty\leq\varepsilon$, and the algorithm output is in $L_{\mathbf{p}}$ with probability at least $1-\delta$. Prior work gave a $(d+1,\varepsilon,\delta)$-list replicable algorithm with sample complexity $\widetilde{O}\left(\frac{d^3}{\varepsilon^2}\right)$, but no nontrivial sample complexity lower bound was known. We establish the first non-trivial sample complexity lower bound for this problem and give an improved algorithm that matches it up to polylogarithmic factors. Together, these results determine the optimal sample complexity for list size $d+1$ to be $\widetilde{\Theta}\left(\frac{d^2}{\varepsilon^2}\right)$, showing that the quadratic dependence on dimension is inherent up to polylogarithmic factors. More generally, for every $k\in[d+1,2^d]$, we establish matching upper and lower bounds of $\widetilde{\Theta}\left(\frac{d^2}{\varepsilon^2\log k}\right)$ on the total sample complexity.
Pretrained generative policies provide strong behavior priors for robot control, but adapting them efficiently with online reinforcement learning remains difficult. A central challenge is how to improve a frozen generative policy without destroying the structure learned from offline demonstrations. Noise-space residual methods address this by modifying the policy's input noise, but existing noise-space methods typically regulate adaptation through fixed latent feasible regions, which may exclude valuable noise edits. In contrast, action-space residual methods are more sample-efficient, but their efficiency can still be further improved. We propose Conservative Noise Directional Optimization (CNDO), a noise-space residual RL method that improves a frozen generative policy by learning state- and noise-conditioned directions around Gaussian latent samples. CNDO removes predefined latent feasible regions and instead regulates noise-space adaptation through the soft conservative constraint, which consists of conservative value estimation and gated execution. We further introduce Conservative Action-Noise Directional Optimization (CANDO), which adds an action-space residual, combining structured latent-space improvement with local action correction. Experiments show that, once CNDO narrows the gap between noise-space and action-space residual learning, CANDO further improves over single-space optimization, achieving 76.1 NAUC and a 91.4% final success rate, surpassing the best single-space results by 6.6 NAUC and 4.1% success rate. More demos and info are available on https://cando-rl.github.io/.
During learning, humans often benefit from curricula that teach simple rules before exceptions. However, we lack a computational explanation for this phenomenon. Here, using a canonical behavioural task and modelling with neural networks, we develop and test a theory for why simple-to-complex curricula succeed. We show that during learning, both humans and neural networks show simplicity biases, a tendency to learn simple rules that are underfit to the data. In neural networks this bias is accompanied by compression of neural representations. Simple-to-complex curricula help both types of learners overcome this compression, developing representations (and policy) with appropriate dimensionality for the task. We test and confirm a further prediction from this theory: learners that have been pre-exposed to simple tasks and with more compressed initial representations benefit more from curricula. These findings suggest that curricula help with ``unlearning'' of simple rules.
This paper presents a CFD dataset of a highly parametrized 3D wing operating in ground effect, designed for benchmarking neural surrogate models for aerodynamic prediction, using a wing shape and flow conditions typically encountered in motorsport applications. The dataset comprises 770 RANS simulations of 194 unique wing configurations, sampled across 27 airfoil profile variants and 4 distances from the ground, spanning variations in wing sweep, chord, taper, angle of attack, and tip twist. Six neural network architectures representing distinct paradigms --- convolutional (DoMINO), graph-based (X-MeshGraphNet), transformer-based (AB-UPT, GeoTransolver, GAOT), and JEPA-based (AeroJEPA) --- are evaluated on their ability to predict surface pressure coefficient and wall shear stress distributions, from which integral drag and lift coefficients are recovered. The dataset is intended to be used as a challenging benchmark for surrogate CFD models, due to the uniqueness of the ground effect being absent from other wing and airfoil datasets.
Training neural operators on million-cell meshes can exceed GPU memory limits. Models are therefore trained on subsamples and deployed on whole meshes or selected regions. However, subsampling can distort boundary information, degrading accuracy. We ask when boundary conditioning under sampling stays consistent with whole-mesh training, and how to aggregate and inject boundary data. Representing the solution as an integral of boundary data against type-dependent measures, we show that subsampling distorts this conditioning in three ways: incomplete boundary coverage or a sample-dependent global state, face weighting by count rather than area, and compression of boundary inputs into a low-dimensional vector. This analysis yields four measurable consistency conditions checkable before training. We satisfy them with two backbone-agnostic modules. Boundary Quadrature Attention (BQA) aggregates all boundary faces, typed and area-weighted, into slice tokens that every query point reads at every layer. Anchored Global State (AGS) estimates the global state from anchors independent of the sampled points. Adding both to a scalable message-passing and slice-attention backbone yields SCANO. We evaluate the approach on controlled and public benchmarks and a data-centre thermal simulation dataset (DCTS) containing million-cell meshes. On both benchmarks, BQA improves every backbone, reducing whole-mesh error up to 15-fold. On all three datasets, AGS aligns regional with whole-mesh predictions, cutting regional error at least tenfold for most backbones on the public benchmark. Together, the modules nearly halve the best original backbone's whole-mesh error on both benchmarks. SCANO predicts a room-scale field in under one second, over 100 times faster than whole-room anchor self-attention.
Per-step selective back-propagation forwards every example of a mini-batch but back-propagates only some of them. Most rules rank examples by loss or gradient magnitude, but such scalar scores ignore cancellation between per-example gradients: the gradient removed by a set can be much smaller than its individual magnitudes suggest. We introduce CAPSkip (Cancellation-Aware Prefix Skipping). Among the examples the model already classifies correctly, CAPSkip skips the longest prefix, in increasing order of last-layer gradient norm, whose vector sum stays within a relative budget $b$ of the batch's last-layer gradient. In exact arithmetic this bounds the change of the raw final-layer gradient, using rank-one contributions already available from the forward pass; the bound does not extend to deeper layers, optimizer updates, or the training trajectory. Because the cumulative distortion is not monotone in the prefix length, CAPSkip scans all prefixes; a factorised Gram matrix makes the scan cheap and keeps it on the device. With activation reuse and selected-row backward, CAPSkip fine-tunes DeiT-S on CIFAR-100 for 103 epochs in 0.576x the time of 103-epoch dense training, and is +0.25 pp [+0.08, +0.42] above an approximately time-matched 59-epoch dense run (n=3).
Diffusion models have shown remarkable progress in AI-generated content, including images and videos. In this paper, we explore whether diffusion models can generate the parameters of policy networks in reinforcement learning (RL), and whether the generated policies exhibit generalization capabilities. To achieve this, we propose Diffusion-based Conditional Policy Network Parameter Generation (DPPG), which consists of a Transformer-based autoencoder and a task-conditioned diffusion model. The autoencoder achieves a high compression rate by patchifying network parameters, while the task-conditioned diffusion model generates corresponding parameters conditioned on different environments. Our experiments demonstrate that, with a single unified DPPG model, we can synthesize high-quality policies across a diverse set of tasks from different environment domains, including Classic Control, Atari games, and MuJoCo. Additionally, our method can generate networks with a larger number of parameters compared to previous approaches. Furthermore, we find that in previously unseen environments, DPPG can generate parameters that achieve faster convergence and reach competitive performance, showcasing the potential future applications of parameter generation via diffusion.
Supervised fine-tuning (SFT) is a key component of language model post-training, enabling effective task adaptation and serving as an important foundation for subsequent reinforcement learning. Recent methods seek to improve SFT by introducing token-level weights into the training objective, where these weights are typically derived from a model's output distribution at each token position. To better understand how such weighting schemes affect SFT, we analyze their optimization objectives and theoretically characterize the target distributions they induce. Our analysis shows that dynamic weighting methods, such as Dynamic Fine-Tuning (DFT), induce a heavily sharpened distribution, whereas static weighting methods fit a target distribution proportional to the product of the data distribution and a frozen prior. These findings suggest that static weighting provides a more stable training target, while its effectiveness critically depends on the quality of the prior distribution. Building on this observation, we propose Self-Privileged Fine-Tuning (SPFT), which uses a frozen copy of the initial model conditioned on privileged information (such as ground truth for mathematical questions) as a prior to provide guidance for token-level weighting during fine-tuning. Through a Bayesian decomposition, we further show that SPFT reweights the standard static prior according to how informative each target token is for predicting the ground-truth outcome, thereby assigning larger weights to more predictive tokens. Experiments on mathematical reasoning and code generation show that SPFT improves over standard SFT and other baselines across the main settings.
We show that replacing LinUCB's usual confidence width $\beta$ by any fixed $\alpha>0$ preserves sublinear regret when the action set is $M$-smooth and $m$-strongly convex. Our high-probability regret bound for this generalisation, $\alpha$-LinUCB, establishes that valid confidence sets are not necessary for the success of LinUCB. We identify the mechanism behind this result: action-set curvature drives $\alpha$-LinUCB to explore every direction, even when optimism is not guaranteed. To analyse this mechanism, we introduce a logarithmically capped spectral potential that, unlike the standard log-determinant potential, caps contributions from well-explored directions. This yields lower bounds on the smallest eigenvalue of the design matrix. At the standard confidence width $\alpha=\beta\asymp\sqrt{d\log T}$, our $T$-round regret bound in $d$ dimensions is $O\bigl(\frac{M}{m}d\sqrt{T\log T}\bigr)$, up to iterated-logarithmic factors, improving the horizon dependence of the classical LinUCB guarantee by a factor of $\sqrt{\log T}$.
Attention is known to be sparse in vision transformers. In particular, we analyze the unequal relevance of attention-matrix columns to the final computation and show that this phenomenon becomes increasingly pronounced with network depth. This can be interpreted as saying that tokens have an absolute importance for the attention process, independently of queries. Inspired by this, we propose a reformulation of the attention formula that makes this phenomenon explicit by adding an input-dependent bias to the query-key dot product. We show that this new term gives each token a scalar value indicating its anticipated influence on other tokens during the attention process, and can thus serve as a proxy for pruning attention maps. Across multiple transformer model sizes, this allows us to substantially reduce the computational cost of attention, in terms of both FLOPs and inference latency, while retaining most of the original classification accuracy. We publicly released the code and models at https://github.com/EIDOSLAB/AttentionWithBias.
Graph neural networks are increasingly proposed for network intrusion detection, yet many reported results come from protocols in which test flows shape the graph used for training, duplicated flows sit on both sides of the split, or a single split is reported. This study measures what a residual graph convolutional network gains over simpler models once these sources of optimism are removed. CICIDS2017 and NF-UNSW-NB15-v2 were cleaned with label-aware removal of exact duplicates, removing 69% of NF-UNSW-NB15-v2 flows. Flows were linked in a nearest-neighbour graph built from training flows only, and validation and test flows were attached to their nearest training flows at prediction time. A two-layer residual network was compared with eight models under the same settings, including a nearest-neighbour vote over the same graph, gradient-boosted trees and a random forest, over three seeds that each draw a new split, with ablations, leave-one-family-out and feature-noise tests. On CICIDS2017 the residual network reached a binary F1 of 98.11% (SD 0.18) and on NF-UNSW-NB15-v2 90.57% (0.38), indistinguishable from a multilayer perceptron without the graph and from other graph networks. Tree ensembles were 1.48 to 1.57 F1 points higher on CICIDS2017 and led on the area under the precision-recall curve on both datasets; their CICIDS2017 multiclass macro-F1 reached 93.08% against 64.51%, a gap mainly from benign flows the residual network assigned to small attack classes. No ablation, including transductive graph construction, changed F1 beyond seed variation. Under Gaussian feature noise (SD 0.5), the residual network kept an F1 of 96.16%, while XGBoost fell to 0.67%. Under a leakage-controlled protocol, message passing over flow graphs adds no measurable accuracy over a perceptron, and tree ensembles remain stronger on both; the advantage of graph models lies in robustness to feature noise.
Looped Transformers extend inference-time computation through repeated latent-state updates, but final-depth training does not directly encourage useful intermediate alternatives. We introduce Best-of-Depth (BoD), which trains a recurrent trajectory as a candidate set by minimizing prediction loss across depths. Different valid targets can be matched at different depths without architectural changes or depth-specific labels. An exact-match analysis connects this objective to solution coverage. On N-Queens completion and graph coloring, BoD improves coverage over final-depth and intermediate-supervision objectives. N-Queens coverage reaches 91.03%, versus 45.75% for final-depth training. Further experiments show discovery beyond the training depth and recovery of solutions withheld from teacher pools on familiar inputs. Fine-tuning a pretrained looped language model with BoD yields higher pooled code-generation success than final-depth training when programs are generated separately at evaluated fixed depths. These results support training recurrent computation for solution discovery along its trajectory.
We compare feature-based machine learning and end-to-end deep learning for variable star classification using the Zwicky Transient Facility (ZTF) StarEmbed dataset, which contains about 40,000 objects across seven imbalanced classes. Classical models trained on physically motivated features consistently outperform deep networks trained directly on raw light curves. We find that limited labeled data contributes to the performance gap, while the multivariate, irregularly sampled, and morphologically diverse nature of light curves poses additional challenges for deep learning models. While deep learning holds promise for learning representations directly from observations, our results indicate that current models fail to generalize with the available training set size. These findings motivate further investigation of the benefits of scaling labeled datasets and self-supervised pretraining on large unlabeled surveys. This work establishes a quantitative baseline for future research and highlights the need for further innovations to unlock the potential of deep learning in astronomical time series analysis.
Federated Learning (FL) enables collaborative training of machine learning models across distributed clients while preserving data privacy. In practical deployments, FL relies on selecting a subset of clients in each communication round to reduce communication cost, mitigate stragglers, and improve training efficiency under heterogeneous data and system conditions. In heterogeneous edge computing environments, both statistical heterogeneity and system heterogeneity significantly affect convergence efficiency and model performance. However, existing client selection methods often address these factors independently, overlooking their coupled and dynamic impact across communication rounds. To address this limitation, we propose FedCSGA, a Client Selection method for Federated learning based on a Genetic Algorithm. FedCSGA encodes client subsets as chromosomes and evaluates these subsets through a non-linear fitness function that jointly integrates resource-aware metrics and history-aware participation weighting. This design captures subset-level interactions, mitigates persistent selection bias, and balances statistical utility with system efficiency. Extensive experiments on FMNIST, CIFAR-10, CIFAR- 100, and Tiny-ImageNet demonstrate that FedCSGA consistently outperforms state-of-the-art baselines across diverse heterogeneity settings, achieving accuracy improvements of up to 7.0% and faster convergence with fewer communication rounds. These results validate the effectiveness of jointly modeling statistical and system factors to achieve robust and efficient learning in heterogeneous federated environments.
Standard Graph Neural Networks (GNNs) implicitly assume that neighboring nodes share the same feature space, which contributes to oversmoothing and poor performance on heterophilic graphs. Sheaf Neural Networks (SNNs) remove this assumption by giving each node and edge its own vector space and learning restriction maps that transform features as they move between neighbors. These maps define a sheaf Laplacian that drives diffusion over the graph, but existing implementations often assemble it explicitly as a global sparse block matrix, which adds memory and computational overhead and can hinder scalability. In this paper, we introduce Sheaf-MPNN, a PyTorch Geometric library that reformulates sheaf diffusion as local incidence-wise transport and aggregation, eliminating the need to construct the global Laplacian. Sheaf-MPNN implements fifteen families of SNNs under a unified message-passing framework, each supporting both homogeneous and heterogeneous graphs, and directed and undirected settings, with diagonal, general, orthogonal, and low-rank restriction maps and attention-based predictors. We further provide a unified benchmarking suite covering node- and graph-level tasks, heterogeneous graph learning, and computational efficiency, enabling systematic and reproducible comparisons across architectures. We establish theoretical equivalence between these message-passing formulations and their original counterparts and identify the architectural mechanisms needed to preserve their propagation dynamics. This unified perspective provides a principled foundation for implementing, comparing, and extending sheaf-based architectures within standard graph learning frameworks. Source code is available at https://github.com/alessioborgi/sheaf-mpnn.
Open-set test-time adaptation (TTA) aims to adapt source-trained models to unlabeled test streams where covariate-shifted in-distribution (csID) samples and covariate-shifted out-of-distribution (csOOD) samples may coexist. Existing TTA methods are typically designed under a closed-set assumption and therefore lack explicit mechanisms for reliably distinguishing csID from csOOD samples. Consequently, csOOD samples may be mistakenly included in confidence-enhancing adaptation, producing harmful gradients that degrade known-class performance. To address this, we propose Uncertainty-Aware Prototype Learning for Open-Set Test-Time Adaptation (UAPA), a plug-and-play method for reliable open-set adaptation. Specifically, we first introduce Uncertainty-Driven ID/OOD Separation, which combines a classifier-prototype score with a mini-batch-normalized evidence-derived uncertainty score to distinguish csID from csOOD samples. We then propose a Dynamic Prototype Refinement strategy that periodically refines class-specific target prototypes using buffered csID features to progressively follow the estimated class directions. Lastly, we introduce Contrastive Prototype Entropy Adaptation, which applies entropy minimization to csID samples and entropy maximization to csOOD samples, while prototype-level contrastive learning aligns csID features with their corresponding target prototypes and reduces the maximum similarity between csOOD features and known-class prototypes. Empirically, UAPA outperforms the strongest baselines on the CIFAR benchmarks and Tiny-ImageNet-C, improving average OSCR by 2.08 and 1.28 percentage points, respectively.
Phase-field models resolve microstructural evolution during rapid alloy solidification, but their computational cost limits long trajectories, stochastic ensembles, and studies across process conditions. We present a variance-preserving denoising diffusion bridge model (VP-DDBM) for state-conditioned prediction of two-dimensional Al-Cu solidification. Given the current phase, concentration, and prescribed temperature fields, the model predicts the subsequent state and is applied autoregressively. Across two held-out thermal gradients, VP-DDBM produces lower long-horizon errors than deterministic U-Net and SI-ConvNeXt baselines in domain-mean phase, domain-mean Cu concentration, and leading-tip position. It also retains fine dendritic structure over long rollouts, whereas the deterministic U-Net progressively smooths the morphology. VP-DDBM predicts the phase and concentration evolution approximately $73\times$ faster than the phase-field solver, showing potential for faster process-condition screening and ensemble-based microstructure studies.
Transaction event sequences support diverse prediction tasks in online commerce and digital financial services. The same transaction records can yield prediction targets with different semantics, including values, changes, and relations to past events. These targets impose different constraints on history representations, but their predictability alone does not establish the downstream utility of a target set. Moreover, target-wise prediction losses do not explicitly model conditional co-occurrence among targets. To address these challenges, we propose TAPES, a task-aligned joint predictive pretraining framework for transaction sequences. TAPES constructs a finite collection of candidate target sets from admissible field-operator combinations. It selects a target set according to the downstream utility of representations learned through short joint pretraining trials under matched budgets. It then reinitializes the encoder and pretrains it on the selected targets through a joint likelihood parameterized by a shared discrete latent variable, before downstream supervised adaptation. Experiments in a real-world transaction-risk setting and on multiple public financial benchmarks spanning diverse downstream tasks demonstrate that TAPES outperforms representative baselines. Ablation studies further validate the contributions of predictive target selection and joint modeling.
Graph neural network explanations are often evaluated via deletion fidelity, i.e., by deleting their top-ranked edges and measuring the resulting change in model output. Yet, in a message-passing GNN, "deleting an edge" does not specify a unique intervention. Structural deletion removes the edge from the graph and recomputes quantities such as degree normalization and attention, whereas message suppression leaves this structural context intact while blocking the message carried by the edge. The meaning of deletion fidelity therefore depends on what the intervention targets in the model computation. We formalize this distinction and derive sufficient conditions for exact equivalence. The two interventions coincide for additive message passing with edge-set-independent coefficients, but can diverge when aggregation depends on the presence of zero messages or when removing an edge changes the coefficients applied to surviving messages. We then construct intervention-specific edge attributions and evaluate each resulting ranking under both interventions, testing whether their relative fidelity changes with the evaluation intervention. Across nine datasets and a panel of GIN, GCN, and GATv2 models at depths 2--5, all nine GIN controls satisfy the equivalence predicted by the theory. On the target logit, the relative fidelity of the two rankings depends on the evaluation intervention in 10 of 11 reportable GCN and GATv2 cells, and in nine of them each ranking has the larger deletion area under the intervention it was built with. Deletion fidelity is thus not a property of an edge ranking alone. Evaluation protocols should explicitly report the intervention, output functional, target, and perturbation unit, and test alternative intervention semantics when the scientific question does not uniquely determine one.
Hybrid models let neural networks supply the unresolved parameters of a process-based simulator, and are fitted to observations that constrain aggregate fluxes and storages but not the internal partitioning that produces them. Such a fit is equifinal: configurations indistinguishable in the observed aggregates can disagree substantially about the unobserved internals, and therefore about the mechanism they attribute to the same data. We study evapotranspiration, whose total flux is observed while its split into transpiration and evaporation is not. A model ensemble built from random seeds alone spans only a narrow band of the partitionings the data admit. We produce the spread deliberately, training members under auxiliary repulsion on transpiration, subject to the fit to observations not degrading. Non-interference is usually arranged by making the auxiliary gradient orthogonal to the main one, but under AdamW the applied step is a preconditioned momentum direction, so orthogonality between raw gradients no longer implies it. We introduce Non-interfering Asymmetric Network Optimization (NANO), an update whose first-order effect on the main loss is exactly that of the main objective alone, whose auxiliary step is bounded, and which concentrates repulsion in the coordinates the main gradient leaves idle.
Probabilistic time-series forecasting requires models that simultaneously capture structured temporal dynamics, expressive uncertainty, and efficient inference, yet existing approaches fall short of this goal: latent dynamical models impose structure but rely on restrictive generation mechanisms, while modern generative methods such as diffusion and flow matching achieve flexibility at the cost of iterative and computationally expensive sampling. We introduce the Koopman Generative Operator (KGO), a probabilistic forecasting method that conceptualizes prediction as the evolution of structured uncertainty. KGO integrates three core components: (1) Koopman Patch Embedding (KoPE) for temporally consistent latent trajectory extrapolation; (2) Koopman Flow Matching (KoFM), which enables fast, single-step generation through a closed-form matrix exponential in a Koopman latent space, bypassing iterative sampling; and (3) an Adaptive Uncertainty Gate (AUG), which provides probabilistic forecasts by adapting uncertainty per-variable and per-horizon, corresponding to a learned decomposition of aleatoric uncertainty without manual tuning. By unifying these components, KGO achieves state-of-the-art accuracy on 13/17 ProbTS benchmarks in CRPS and 16/17 in NMAE. Further, by eliminating iterative sampling, KGO delivers at least a 25$\times$ reduction in inference time compared to iterative generative models. These results position KGO as a practical and principled framework that advances the accuracy-efficiency trade-off in probabilistic time-series forecasting, combining leading accuracy on the evaluated benchmarks with fast, single-step generation.
We study cooperative multi-agent reinforcement learning with decoupled dynamics and a monotone submodular reward function over joint trajectories. Due to the non-additive reward, an optimal policy may depend on the history of all agents. This dependence makes exact optimization exponential in the horizon and the number of agents. To overcome this challenge while retaining theoretical guarantees, we introduce pre-window curvature $c_m$, which quantifies how states visited more than $m$ steps earlier reduce current marginal gains. Using this curvature to control the error from restricting the policy of each agent to depend only on its last $m+1$ states, we propose a greedy algorithm that exploits the decoupled dynamics to optimize local policies sequentially. For fixed $m$, the algorithm achieves a $\frac{1-c_m}{2-c_m}$ approximation guarantee up to estimation error, and its computation time is polynomial in the number of agents and the horizon. We recover the known $1/2$-factor approximation for time-additive rewards and obtain an approximation factor no smaller than the known total-curvature bound for a single agent. We formulate multi-agent traffic monitoring as an example of this framework and evaluate our algorithm on it.
Reinforcement learning objectives that maximize expected return can concentrate probability mass on a small subset of high-reward outcomes, limiting diversity when multiple distinct solutions are desirable. We introduce Density-Aware Reward Scaling (DARS), a simple reward-processing mechanism that promotes reward-proportional sampling in continuous outcome spaces while retaining standard policy-gradient training. DARS estimates the local density of outcomes within each on-policy rollout group and inversely scales rewards by this density. The method requires no learned density model, auxiliary objective, critic, or modification to the underlying policy-gradient objective. In a controlled multimodal navigation task, DARS-GRPO avoids the single-mode collapse exhibited by GRPO and closely recovers the target mode allocation. Applied to text-to-image prompt adaptation, DARS-GRPO consistently improves image and prompt diversity across datasets, Policy LMs, and reward settings while maintaining good semantic alignment. These results show that density-aware reward scaling provides a lightweight way to improve coverage of diverse high-reward outcomes within existing reinforcement-learning pipelines.
Large language models (LLMs) on personal devices keep user data local, but limited memory restricts their general capability. Since most user-specific tasks do not require the breadth of frontier models, a practical solution is to adapt a shared on-device model into lightweight specialists. However, backpropagation is prohibitive in memory, and users rarely provide labeled data. We present MobileSpec, a system that trains specialists with forward passes alone and learns from automatically checked task outcomes, as in reinforcement learning with verifiable rewards (RLVR). Because forward-only updates are built from random perturbations, training is efficient only when these perturbations resemble the weight changes of RLVR, which largely preserve the singular values of pretrained weights and rotate their singular vectors. The core algorithm of MobileSpec, FuRA-S, produces such perturbations while training only the block-spectral coordinates of pretrained weights, namely the singular values of their column blocks. As we show theoretically and empirically, rescaling these blocks independently shifts their relative contributions to each weight matrix and thereby rotates its singular vectors. Since this update can be rewritten as a transformation of the layer input, FuRA-S also adapts INT4 models without modifying their weights. With specialists of only 1-5 MB that update 0.015-0.05% of the parameters, MobileSpec raises the success rate of a phone-use agent from 6-23% up to 89-100% and trains a personalized 8B model on an iPhone 16 Pro overnight. On RLVR benchmarks, FuRA-S is competitive with full-parameter evolution strategies (ES) and outperforms ES over LoRA adapters and quantized ES on INT4 models.
Finite-sum nonconvex-strongly-concave minimax problems arise in adversarial training and distributionally robust learning. Their known lower bounds cover only linear-span algorithms, zero-respecting algorithms, or a restricted proximal protocol. They leave open the power of exact component proximal steps, whose responses are global saddle points of regularized component problems and can expose directions that gradients do not reveal. We show that neither unrestricted adaptivity nor exact proximal access improves the polynomial complexity. We consider $n$ components with joint mean-squared smoothness $L$ and condition number $\kappa$, started from a point with primal gap at most $\Delta$ and joint gradient norm at most $\sqrt{2L\Delta}$. For $\varepsilon^2/(L\Delta)$ below a universal constant, every randomized adaptive algorithm with exact component values, gradients, and proximal responses needs $\Omega(n+\min\{\sqrt n\,\kappa,n^{3/4}\sqrt\kappa\}L\Delta/\varepsilon^2)$ component queries in the worst case to find an $\varepsilon$-stationary point of the value function. A method using only component gradients attains this rate up to a factor $\log^2(2+\kappa)$, including initialization. Zhou (arXiv:2610.04944) concurrently proves the same rate for zero-respecting first-order methods. In our hard instances, every component admits proximal steps with the same step sizes, and a response for a component reveals hidden directions only up to the first one assigned to another component.
This work investigates the application of Quantum Recurrent Neural Networks to the binary classification of sleep-disordered breathing events in time-series data, aiming to distinguish between apneic and normal breathing episodes. A variational classical--quantum architecture is proposed and evaluated against classical Recurrent Neural Network models. Experiments were conducted on multiple real-world datasets with varying levels of complexity to assess the strengths and limitations of the quantum approach. In the comparative analysis, both classical and quantum models showed comparable performance. The quantum model achieved superior performance on the smallest dataset, while yielding slightly weaker results in the most complex setting. However, the quantum approach exhibited notable strengths, such as competitive performance despite a drastically reduced parameter count and low-qubit regime, as well as good generalization capabilities. These results suggest that, despite current hardware and algorithmic limitations, variational quantum recurrent models represent a promising alternative for time-series classification as Quantum Computing technologies continue to mature.
Anomaly detection at the edge requires the deployment of models that achieve competitive performance, while minimizing computational cost and memory footprint. In this article, we investigate the suitability of the Mahalanobis Distance of Reservoir States (MDRS), an unsupervised time-series anomaly detection method, for resource-constrained devices and propose two optimizations to reduce its computational requirements. Specifically, we replace random reservoir connectivity (MDRS-Rand) with a circular shift (MDRS-Circ) and show that the anomaly score for an entire time series can be computed from the final reservoir state alone. Furthermore, we show that exploiting the reservoir's temporal memory enables more efficient input preprocessing, further reducing computational cost. We evaluate the MDRS variants on the MLPerf Tiny anomaly detection benchmark against the reference autoencoder and other anomaly detection methods, and deploy their inference kernels on a NUCLEO-F401RE microcontroller to assess latency, estimate energy consumption, and measure memory footprint. Among the evaluated methods, only MDRS and the autoencoder meet the benchmark's performance target, with MDRS requiring fewer operations and smaller models. With optimized preprocessing, MDRS achieves an ROC AUC of 0.869, a model size of 24 KB, an estimated energy consumption of 24 uJ per inference, and a latency of 0.32 ms, corresponding to a 75.9x speedup over the autoencoder with standard preprocessing.
Can a search over innate drives find intrinsic rewards that help a reinforcement-learning agent learn, transfer to a task it never saw, and resist an exploit that was absent during the search? Each candidate reward combines four generic events (a new view, a pickup, a drop and a toggle), each with a weight, a relief and a refill rate. We ran three pre-registered stages; the search protocol was locked in a public, timestamped commit before any search result existed. CMA-ES and an equal-budget random search used 1,560 PPO training runs, and the final evaluation and tamper test used 340. On the held-out task, KeyCorridorS3R2, the CMA-ES-selected reward (E) beat task-only PPO, episodic novelty and the random-search-selected reward (R) in all 3 registered comparisons, all of them meaningful by the registered size rule. Among these generic rewards E transferred best; the earlier paper's satiation reward, which on this task is given hand-wired key and locked-door events, did better (20 of 20 seeds solved against 4 of 20; descriptive). On UnlockPickup, where that reward uses generic events, E's mean area exceeded it by 0.276 (descriptive). On the training tasks there was no registered difference between E and R: both E-over-R tests missed. A free reset action, absent during the search, collapsed E's learning on DoorKey-8x8 (mean area 0.829 without it, 0.136 with it; registered) to below task-only PPO (descriptive). The searched drives helped on every task tested, but a reward given task-specific events did better on the held-out task, and selection gave E no protection against an exploit absent from the search.
Online continual learning aims to continuously train neural networks on a non-stationary data stream in a single pass. Rehearsal-based methods alleviate catastrophic forgetting by replaying stored samples, but the imbalance between current and previous data often leads to predictions biased toward current classes. Many approaches have been proposed to overcome the forgetting problem by correcting the bias; however, they still need to be improved in an online fashion. By decomposing the dot-product logits into an angle factor and a norm factor, our analysis reveals that the angle factor efficiently captures novel knowledge but is particularly susceptible to prediction bias, whereas the norm factor helps preserve historical knowledge. Based on these observations, we propose Dual-Logit Experience Replay (DLER), which exploits dual-logits for both model optimization and memory management. Specifically, DLER learns current samples using cosine logits and incorporates additional dot-product logits when replaying previous samples to mitigate prediction bias. To strengthen dual-logit learning under limited replay memory, an auxiliary supervised contrastive objective promotes angular compactness within each class, extending corrections learned from limited replay samples to other samples of the same class. Meanwhile, we introduce a memory update strategy that preferentially retains samples for which dot-product prediction assigns a higher probability to the correct class than cosine prediction, aligning sample selection with dual-logit training. Extensive experiments on three datasets demonstrate the effectiveness of DLER, particularly under limited memory budgets.
Real-world sensing systems are inherently imperfect: sensors may fail, intermittently disconnect, operate with asynchronous sampling rates, and be subject to noise, leading to measurement streams that are irregular, incomplete, and heterogeneous. Nevertheless, sensing strategies for real-time state reconstruction typically assume synchronous and continuously active sensors, a constraint that substantially limits their applicability in operational environments. In this work, we introduce fault-tolerant sparse sensing strategies in the context of shallow recurrent decoders (SHRED) to enable high-dimensional state estimation from sparse, asynchronous, and missing sensor measurements. Specifically, we consider naive imputation strategies, fault-tolerant recurrent temporal encoders, and we introduce four different transformer-based masking mechanisms to make SHRED explicitly aware of sensor availability through the binary data-availability mask. Moreover, we further introduce a multi-step decoding training objective for SHRED, resulting in smoother latent trajectories and more robust predictions. Throughout a synthetic benchmark and three real-world problems in hydrodynamic and ocean modeling, we show that (i) transformer-based masking mechanisms outperform standard imputation techniques and fault-tolerant recurrent architectures, (ii) the multi-step decoding training objective enhances reconstruction fidelity and robustness in the presence of heterogeneous missing-data patterns, while (iii) the computational efficiency of SHRED is preserved, with no additional inference-time overhead.
Recurrent memory requires learning useful updates, but attention added to assist learning can also compromise length generalization. We study bounded delta attention: causal attention over uncorrected update proposals whose bounded correction enters the recurrent update. Its implementation in TraceRelay combines additive phase recurrence with attention computed before a parallel scan. On prefix-supervised modulo-5, trained at lengths 1--40, three findings characterize its behavior at length 80. First, upper delta attention attains $70.81\pm8.26$% accuracy across seven seeds, versus $69.21\pm22.42$% for the local-attention recurrent baseline (mean $\pm$ sample SD): a 63.2% reduction in SD with a similar mean. Upper state attention also reduces variation, reaching $74.75\pm6.24$%, so this benefit is not unique to update-level attention. Disabling the upper delta correction retains $70.62\pm8.60$%. Second, blocking gradients at the delta-attention inputs changes accuracy to $65.25\pm12.40$% in the matched seven-seed comparison, increasing SD by 50.2%. It also delays a 99% validation threshold on average among the same six seeds that reach it; the 95% threshold is reached earlier on average with input-gradient blocking across all seven seeds. Third, lower delta attention exhibits substantially less extrapolation degradation than lower state attention: 62.60% versus 19.40%, against a 71.06% baseline. Removing its correction reduces accuracy to 22.56%; its advantage over lower state attention also persists under input-gradient blocking. These observations motivate bounded update-level attention as a recurrent learning scaffold, combining reduced seed variation with less severe degradation in a placement where state attention fails.
Nanodevices that absorb energy from their thermal environment can act as quantum refrigerators. Finding suitable configurations, however, is a needle-in-a-haystack problem: the many-body state space scales exponentially with the size of the device, where each candidate configuration requires an expensive quantum-chemical computation. We introduce a search framework that couples a Graph Variational Autoencoder (GVAE) with an uncertainty-driven Active Learning (AL) loop to efficiently identify energy-absorbing configurations in electronic junctions with a 2D "working material" to be optimized. The GVAE maps physical parameters - site energies, coupling matrices, temperature - into a smooth latent space that distinguishes cooling from dissipative parameter regimes, achieving a classification discriminability of $0.920 \pm 0.004$ on held-out configurations and $0.895 \pm 0.010$ on unseen extrapolation topologies. An AL pipeline based on Monte Carlo dropout then targets the decision boundary, iteratively selecting the candidates where the model is least certain for simulation. Over 32 AL cycles, the uncertainty-driven strategy discovers energy absorbing systems at rates consistently exceeding the random baseline across all tested topologies (where only 1.9-8.4% of configurations exhibit the desired property), providing a data-efficient pathway for navigating a combinatorically large design space. The learned latent space organizes configurations along physically interpretable axes, that correlate with underlying microscopic model parameters. This hints at the existence of "cooling coordinates" - low-dimensional, physically interpretable quantities separating cooling from non-cooling regimes. Our framework, thus, paves a new way towards effective cooling of nanoscale electronic information processing devices, to reduce the environmental footprint of high performance computing.
Continual learning (CL) aims to adapt models to new tasks while preserving previously learned knowledge. Recently, Low-Rank Adaptation (LoRA), a representative Parameter-Efficient Fine-Tuning (PEFT) method, has gained increasing attention in CL due to its efficiency and scalability. Several LoRA-based merging methods maintain a single inference model by scaling the newly learned LoRA update with one coefficient and merging it into the previous weights after each task. However, they usually control the whole update with a shared scalar factor, although different update elements may affect previous and current tasks in different ways. To address this issue, we propose Merging Strength Modulation via Joint Loss Estimation (M2LE). It consists of Hessian Information-guided Adaptive Merging (HIM) and Task Simulation Perturbation for Merging (TSP). Instead of searching for a global merge coefficient, HIM directly solves for the target merged update under a joint objective over previous and current tasks. Solving this objective gives an analytic solution that naturally takes the form of element-wise modulation over the current task LoRA update. To improve the stability of the learned update under subsequent merging, we introduce TSP which simulates different tasks interference. Experiments on multiple benchmarks show that M2LE consistently outperforms existing PEFT-based continual learning methods. Code is available at https://github.com/bbdeeer/M2LE_NeurIPS2026.
In this paper, we establish a universal approximation theorem for solution operators of semilinear stochastic heat equations in spatial dimension \(d\ge2\), extending deep neural operator approximation results to SPDEs with distribution-valued Gaussian forcing. A smooth Fourier cutoff reduces the forcing to finitely many measurements, while Besov and Schauder estimates control the resulting solution error. The finite-dimensional solution map is then approximated by ReLU branch networks together with grid-based trunk functions, yielding explicit bounds on the numbers of forcing measurements and trunk functions and approximation in every finite stochastic moment. For linear equations, the reduction coincides with truncation of the first Wiener chaos. Lastly, as an application, we prove a high-probability local-in-time approximation theorem for stochastic Navier--Stokes equations, with finite measurements of both the initial condition and forcing and an explicit sensing rate.
Recent work characterizes the Adam optimizer as sign descent modulated by gradient uncertainty: the update is the sign of the batch gradient, scaled by how strongly the batch supports that sign. We apply this view directly and introduce ZEMA (Z-score Evidence Momentum Adaptation), which measures that support with a per-coordinate $z$-score computed from the per-example gradients in the current batch, converts it into a sign probability, and smooths the resulting update with a single exponential moving average. Because the probability is formed before smoothing rather than after, ZEMA keeps one state per parameter instead of Adam's two, at a $4.6$ increase in per-step time; the per-example variance it requires can also be estimated from a small fraction of the batch. In fine-tuning and pretraining experiments, ZEMA is competitive with Adam, reduces the variability of the validation loss, and in some settings recovers the sign of the full-dataset gradient more consistently. We close with two directions this opens: extending confidence adaptation beyond the sign geometry, to methods such as Muon, and to differentially private training.
Cluster labels are often used twice: first to discover groups, and then to test whether the discovered groups differ. This double use of the data makes ordinary two-sample $p$-values anti-conservative. Selective inference fixes this by conditioning on the clustering event, but for spectral clustering the event appears difficult because the algorithm combines an affinity map, an eigendecomposition, and a rounding step. We show that, for spectral clustering based on unweighted $\varepsilon$- or $k$-nearest-neighbor graphs, the obstruction is the affinity construction rather than the eigendecomposition. Along the one-dimensional perturbation used by the truncated-$\chi$ selective test, graph changes occur only at finitely many real roots of data-dependent quadratic distance comparisons. Between these breakpoints, the graph, Laplacian spectrum, embedding, rounding, and reported cluster pair are deterministic functions of a finite graph state. Consequently, in real arithmetic and under a known-scale spherical Gaussian model, the selection set is an exact finite union of intervals, yielding finite-sample selective $p$-values for selected spectral-cluster pairs. The same argument gives feature-wise truncated-normal tests. We also record direct extensions to known feature covariance, eigengap-selected $K$, and auxiliary-coordinate graphs. We validate null and partial-null calibration, operational power, numerical sensitivity, and model-scope diagnostics, and give a working-model PBMC algorithmic illustration. Smooth kernels such as RBF affinities fall outside the finite-state guarantee.
Web systems infer user preferences from behavior, ratings, text, and LLM-generated profiles, yet explaining a proxy may reveal its measurement process rather than the preference of interest. We develop an audit framework distinguishing point identification, partial identification, and abstention relative to an explicit reference distribution. Under feature-dependent, arbitrarily signed proxy bias, a strict majority of uncontaminated proxy attributions lets the coordinatewise median recover latent-attribution support and direction and, with externally justified positive scales, magnitude. We prove this threshold sharp using response-valid indistinguishable worlds, characterize sharp minimax error and exact compatible-target sets under approximate cleanliness and scale uncertainty, and propagate clean-proxy attribution error. A frozen known-truth benchmark spans two response families, four proxy counts, all contamination counts, three bias strengths, and eleven methods, with a focal comparison confirmed on twenty unseen seeds. Under correct budgets, a direct attribution-tensor audit achieved 100% coverage, 95.40%-97.05% active-driver sign decisions, and no false signs. In a temporally separated Amazon study, median aggregation reduced comparator-relative conditional sign error for all four targets; six of eight prespecified target-metric contrasts had Bonferroni intervals excluding zero. A controlled real-covariate perturbation favored the median on the prespecified focal metric for all four targets. Compatible sets showed that the current proxy configuration rarely licensed strict-sign conclusions. Weak comparator calibration limits the Amazon evidence to fitted-surface alignment rather than latent-preference recovery.
Pervasive Capability Degradation (PCD) refers to the phenomenon in which supervised fine-tuning (SFT), while adapting a large language model (LLM) to a target task, often erodes the general capabilities acquired during pre-training. Prevailing explanations attribute PCD to macroscopic distribution shift and therefore lack microscopic manipulability, leaving existing mitigation methods without a complete theoretical foundation. In this paper, we first develop a bottom-up geometric theory of PCD grounded in feature interference under superposition: narrow-domain SFT data disrupt the almost-orthogonal feature structure formed during pre-training, and the resulting interference drives activations across decision hyperplanes in the output Laguerre-Voronoi geometry, flipping predictions on general tasks. This theory unifies existing methods into six paradigms and reveals both the source of their efficacy and their theoretical limits. We then prove that cross-entropy under one-hot supervision cannot resolve PCD on its own, and accordingly propose the AEGIS loss, which directly constrains the drift of activation margins along all competing token directions, thereby decoupling interference suppression from task adaptation. Across 3 model families and 14 SFT tasks, AEGIS outperforms representative methods of every paradigm on target tasks and pervasive capabilities simultaneously, at only about 1.3 times the computational cost of standard SFT. Its generality, scalability, and mechanistic interpretability offer a new perspective on fully resolving PCD.
Generalised Additive Models (GAMs) are a cornerstone of interpretable machine learning, offering transparent predictions through structured additive decompositions. Despite advances from classical GAMs to modern neural and spline-based variants, two notable gaps remain. First, comprehensive benchmarking of spline-based GAMs on large-scale tabular learning tasks, the dominant modality in applied settings, is lacking. Second, their robustness under distribution shift, a setting where structured inductive biases should confer advantages, remains unexplored. We propose a multi-resolution, higher-order Composite Generalised Additive Model (CoGA2M) and evaluate it on the TALENT benchmark suite. CoGA2M ranks in the top third across all baselines (29 for regression, 32 for classification), establishing competitive performance with state-of-the-art tabular methods. We further extend TALENT with out-of-distribution (OOD) splits exhibiting train-test distributional shift and benchmark CoGA2M against LightGBM, a consistently top-performing model on TALENT. Under distribution shift, CoGA2M outperforms LightGBM, demonstrating a clear structural advantage. Our results position spline-based GAMs as strong contenders for tabular prediction and reveal a previously unrecognised benefit of additive structure in the OOD regime.
Matrix factorization provides a simple model for studying the training dynamics of neural networks. Recent studies of gradient flow under small initialization have analyzed how trajectories escape from a neighborhood of the origin and approach their first nonzero stationary point. In this paper, we study continuous-time dynamics that applies a finite number of Newton-Schulz iterations to the normalized matrix gradient (Muon). We show that the number of Newton-Schulz steps can change the rank selected during the early stages of training and allow the resulting trajectory to bypass the first nonzero stationary point approached by gradient flow. Under an additional residual-inertia condition, we prove that bypass occurs for all sufficiently large finite Newton-Schulz depths and provide an explicit sufficient threshold. We also construct an example in which the bypass threshold is four.
Continual learning in physical AI differs from conventional continual learning because small per-step policy errors can accumulate over a trajectory. Each deviation may move the agent outside the training distribution, increase subsequent prediction errors, and ultimately cause closed-loop task failure. This work introduces Dynamics-Risk Adaptive Plasticity (DRAP), a replay-free method that estimates the historical importance of adaptable model tensors from sample-level action gradients weighted by local trajectory-amplification risk. A ridge-regularized local linear model is fitted to expert state-action transitions, and the largest singular value of its state-transition matrix provides a policy-independent, Lyapunov-inspired estimate of one-step perturbation amplification. After each task, risk-weighted squared gradients are accumulated in a tensor-level importance memory. During subsequent tasks, DRAP uses normalized cumulative importance to scale the effective AdamW learning rate of each tensor, immediately reducing the plasticity of components that supported previous behavior at dynamically sensitive trajectory points. We instantiate DRAP on a fixed KappaTune-selected subnetwork of OpenVLA and evaluate replay-free sequential adaptation across the LIBERO Spatial, Object, and Goal suites. Compared with standard AdamW fine-tuning and EWC, the results show substantially improved retention of previously learned behaviors while preserving strong new-suite acquisition.
Deep learning in medical imaging is constrained by data scarcity and by privacy regulations that make centralizing patient data across institutions impractical. Federated Learning (FL) addresses these constraints by keeping data local, but its distributed nature exposes training to strong statistical heterogeneity. Multicentric medical data is particularly affected, exhibiting a mixed skew that combines feature, label, concept, and quantity shifts across centers. Personalized FL mitigates heterogeneity by tailoring a model to each client, yet most methods assume training from scratch, which is unrealistic given the high dimensionality of medical features relative to the limited data available. We propose the Personalized Aligner Federated Pretrained (PAFP) framework, a personalized FL approach built on a pre-trained backbone. PAFP inserts client-specific Feature Alignment modules before the fine-tuned blocks to realign hidden feature distributions toward a server-aggregated target. On the FedISIC benchmark, PAFP attains the highest unweighted balanced accuracy while reducing inter-client performance variance roughly six-fold relative to the strongest baseline, yielding accuracy and cross-client equity jointly. Our code is publicly available at https://github.com/Alexandre-Leclercq/PAFP.
Equivariant neural networks incorporate known symmetries of the data, such as permutations, rotations, or translations, directly into their architecture, often improving sample efficiency and generalization. However, strict equivariance can make optimization difficult, for example, when the data only approximately satisfy the assumed symmetry, or when the equivariance constraints induce a challenging loss landscape. Recent studies have shown that relaxing equivariance during training by introducing non-equivariant dense linear layers can ease optimization in such cases and improve the performance of the resulting strictly equivariant model, after removing the relaxation layers at test time. While effective, this approach incurs a significant computational overhead, limiting its applicability in scalable architectures. In this work, we propose a more efficient constraint-relaxation method based on positional encoding (PE). Our framework injects a positional encoding with learnable scale that breaks the model's original symmetry. This positional signal is gradually annealed to zero, recovering the original equivariant model by the end of training. Empirically, we demonstrate significant improvements for translational and rotational equivariant models with only marginal compute and memory overhead.
Direct photons escape heavy-ion collisions with little rescattering, carrying information from the full emission history. Yet describing the measured direct-photon yield and the sizable elliptic flow $v_2$ together remains a long-standing challenge. Here we explore how correlations between emitting partons may affect both observables. The local $2\to2$ photon rate depends on the joint distribution of incoming partons, so a connected correlation between them can change photon emission without changing the one-particle distribution of medium particles. Once folded over the evolving medium in Au+Au and Pb+Pb collisions, the correlation raises the photon yield toward the fitted PHENIX and ALICE data and increases $v_2$ mainly at intermediate $p_T$. The fitted temperature dependence accounts for most of the net flow increase in this region, suggesting that the correlation changes the relative contributions between hotter and cooler regions.
In this paper, we propose Dual-VaDE, an unsupervised generative framework for learning interpretable anatomical prototypes from unlabeled medical images. In contrast to previous variational methods that rely on a single spatial latent code, which often entangles global geometry with local textures and fails to capture their cross-domain dependencies, Dual-VaDE introduces a dual-domain Gaussian mixture prior that explicitly couples spatial anatomy with phase-preserving Fourier spectral representations. We derive an exact, closed-form variational lower bound that fuses this joint prior into the decoder and optimizes the entire network end-to-end via an alternating EM formulation, avoiding heuristic two-stage clustering while ensuring superior training stability and robustness. Evaluated on OASIS-3 brain MRI, Dual-VaDE delivers substantially sharper and more plausible prototypes, achieves superior cluster separation with improvements of up to 94% over spatial-only baselines, exhibits markedly reduced variance across random seeds, and incurs negligible inference overhead (0.28 ms per image).
Predicting affinity between small molecules and their protein targets remains one of the central problems in drug discovery. Due to the substantial cost of wet-lab experiments, machine learning models of drug-target affinity are usually trained on relatively small datasets, ranging from tens to hundreds of molecules. Given the inherent difficulty of generalization between different targets and chemical series, existing datasets of protein-ligand affinities are typically not used for new projects. At the same time, tabular foundation models, including Prior-data Fitted Networks, have recently shown how transformers can be used to solve machine learning problems in a single forward pass of the network, after being trained over a large corpus of, typically, synthetic datasets. Inspired by that line of work, we propose a novel affinity modelling approach, in which the labels of the query molecules are predicted from the labelled support molecules in a single forward pass of a transformer model. The model is trained on thousands of existing affinity datasets, enabling it to capture the general patterns of molecular activity across different targets, chemical series and dataset sizes. For the smallest support sets, our model matches the state of the art on the regression variant of the FS-Mol benchmark, while being faster than existing methods, thanks to the lack of task-specific training. A comparison against a reduced variant of our architecture, which amounts to kernel regression in a meta-learned feature space, shows that the predictive performance of the transformer model is only partially explained by simple molecular similarity.
We give an algorithm for PAC learning intersections of $k$ degree-2 polynomial threshold functions over $R^d$ under a Gaussian distribution. Our algorithm achieves classification error $\epsilon$ using $d^{O(k)}2^{O_k(\log^3(1/\epsilon))}$ time and samples, improving over the $d^{O(k^2/\epsilon^2)}$ running time of the polynomial regression approach. To our knowledge, our result is the first to obtain a dimension exponent independent of the accuracy parameter, even for $k=2$. We complement our algorithmic result with an SQ lower bound showing that the complexity dependence of $d^{\Omega(k)}$ is unavoidable. Lifting the input to polynomial features turns the target into an intersection of $k$ halfspaces. This suggests using the first two labeled moments for subspace recovery. In the lifted space, however, these moments can reveal nothing. Our starting point is a Gaussian comparison theorem showing that, for sufficiently regular defining quadratics, the labeled moments needed for subspace recovery are approximately preserved. To handle arbitrary defining quadratics, we establish a new structural result showing that every direction with significant resampling influence can be detected through degree-$O(k)$ moments. These results yield an initial $d^{O(k)}2^{\mathrm{poly}_k(1/\epsilon)}$-time learner using adaptive restrictions. To improve the exponential dependence on $1/\epsilon$, we develop a two-stage dimension-reduction approach that avoids the costly adaptive partitioning. After identifying a subspace containing all directions with large influence, we use generalized regularity and Gaussian comparison to recover the remaining relevant quadratic features through a new derivative-based moment operator. The recovered features support an intersection close to the target. Combining them with robust hinge-loss minimization gives our quasipolynomial-time learner.
The feed-forward propagation through neural networks is crucial for training and inference. To trade off efficiency and accuracy, it has been proposed to consider intermediate outputs of hidden states as data-dependent regularization during training and, ultimately, as early-exits during inference. In this work, we analyze the joint optimization of loss functions computed for the output layer and intermediate layers using optimal control theory. We prove that ResNets trained in this joint fashion exhibit the so-called turnpike phenomenon, this implies that the number of layers required to reach a prescribed loss level is bounded from above independently of the overall network depth. This induces a similarity property across the internal representations of early-exiting networks of varying depths. We connect the turnpike to the overthinking phenomenon, and provide theoretical insights into the success of early-exit criteria derived from softmax probabilities, specifically entropy and max-probability. In a range of experiments, we demonstrate the existence and effect of turnpikes and the resulting similarity in the early-exit performance for ResNets of varying depths.
Channel estimators at a newly deployed base station usually start from 3GPP stochastic channel models, which describe no particular street. We ask whether a channel prior pooled from many ray-traced cities serves an unseen site better than the 3GPP CDL models, and whether learned estimators make the difference. With Sionna RT on 14 city scenes, LMMSE estimation from covariance priors, leave-one-city-out evaluation, and pass/fail bars written down before each main experiment, we find in simulation that the pooled prior sits about 0.6 dB from a site-trained estimator with no site measurements, about what CDL reaches after 50 measurements. Learned models added little: at most 0.65 dB in estimation, and in channel prediction a simple classical method beat CNN predictors by 1.8-12 dB from 2 ms ahead. On measured channels from the DICHASUS outdoor dataset (1.272 GHz, three 2x8 arrays on one building) the result reversed: the pooled ray-traced prior was 3.2-4.1 dB worse than CDL. The simulated channels were too clean; with users at the real distances, even a scattering level tuned from 50 site measurements left the prior about 2 dB behind CDL. A pre-registered follow-on on two more measured environments found the generic CDL prior 2.1-4.7 dB from a site-trained one: no further in a factory hall than on the street, furthest in a small lab room. Finally, a hallway model calibrated to measurements by differentiable ray tracing reproduced delay spread, but as a prior in an unmeasured part of the hallway it was 4.5 dB worse than CDL on the full array. All measured results come from one sounder.
We study PAC learning when each example comes with a set of at most $\ell$ correct labels and a prediction is correct if it belongs to this set. Training may reveal the full set, only whether a proposed label is correct (bandit feedback), or one correct label chosen by a demonstrator (emonstrations). Our main finding is that demonstrations can be much harder than full sets, and even than bandit feedback, both in which classes can be learned and in how many examples are needed. For finite label spaces, a generalized Natarajan dimension characterizes learnability from full sets and from bandit feedback, in both the realizable and the agnostic setting. From demonstrations, however, a class is realizably learnable if and only if the intersections of at most $\ell+1$ of its hypotheses are learnable from full sets. Already with three labels, this condition can fail for a class learnable from a single full-set example or even a single bandit answer. When the condition holds, we show that $\widetilde O(\varepsilon^{-\ell})$ demonstrations suffice, and the exponent $\ell$ cannot be improved in general, even for classes learnable from $\widetilde O(1/\varepsilon)$ full sets or bandit answers. For countable label spaces and $\ell \geq 2$, our packed dimension, which asks for many disjoint groups of at most $\ell+1$ hypotheses with no common label, characterizes full-set learnability, and demonstrations still reduce to intersections. Agnostic learning from demonstrations, already for finite label spaces, is not determined by the dimensions of the intersection class that characterize realizable learning. Finally, we show that the uniform bound $\ell$ on the number of correct labels is essential. Without this bound, it is impossible to characterize learnability by any dimension determined by finite subclasses.
Synthetic priors have so far served one purpose: pretraining tabular foundation models. We ask whether the same priors can improve conventional tabular models, which are trained from scratch on each dataset. We propose PriorTune, a method that meta-tunes hyperparameter configurations on synthetic tasks alone. Offline, for a given model, we fit a pool of configurations on datasets sampled from a published PFN prior and greedily select the subset with the highest winrate over the random baseline. Online, a new dataset fits only the selected configurations, so the gain comes at no extra cost. Empirically, we observe that scaling both the number of configurations and synthetic datasets has an equal scaling effect on the performance of the final subset of configurations, which we also explain theoretically. On the full BeyondArena benchmark, LightGBM with TabPrep features with our configurations ranks above TabPFN-3 and RealMLP, while being much more efficient at inference. Overall, our results establish existing PFN priors as a scalable source of meta-training data for hyperparameter selection for non-foundational tabular models.
Many scientific simulations are computationally expensive, limiting their use in simulation-based inference, uncertainty quantification, and decision making. We introduce macrocanonical generator networks, a framework for learning fast, data-efficient neural surrogates for amortized simulation of stationary physical processes by matching multiscale statistics of a target process. Inspired by microcanonical maximum-entropy synthesis, which matches feature statistics per sample by inner-loop optimization, our approach instead trains a neural generator to satisfy the same constraints in expectation, preserving physically realistic sample-to-sample fluctuations. We prove that gradient-descent training induces preconditioned gradient-descent dynamics in sample space that inherit the symmetry-preservation properties of the microcanonical gradient descent of Bruna and Mallat, amortizing its per-sample inner-loop optimization into a single training run after which every sample is generated in one forward pass. We further show that residual generator parameterizations initialized near the identity, trained with a small learning rate, preserve entropy locally during early optimization and mitigate mode collapse. Experiments on cosmology and fluid simulations show that macrocanonical generators recover realistic samples from extremely small training sets, in some cases a single training example, while generating new realizations more than $10^6$ times faster per sample than the reference simulator and more than $10^3$ times faster than microcanonical gradient descent. The trained generators can be reused for downstream pretraining, including simulation-based inference, providing a data-efficient route to multi-fidelity scientific machine learning workflows.
Reliable streamflow prediction across time and in ungauged basins requires representations that capture hydrologically relevant basin characteristics. Existing methods rely on either fixed handcrafted attributes or general-purpose Earth observation embeddings not tailored to basin structure. We introduce \textbf{Basin2Vec}, a self-supervised framework that learns compact basin representations from hydroclimatic and physiographic data without observed streamflow. Basin2Vec organizes representation learning around basin scale and relationships across space and time. Across multiple forecasting architectures, it consistently outperforms predefined attributes and general-purpose geospatial embeddings in temporal generalization and zero-shot spatial transfer. Basin2Vec provides average relative gains of up to 8.3% temporally and achieves 3.5x the median predictive performance of the strongest baselines on unseen basins. These results demonstrate the effectiveness of basin-centric pretraining for transferable hydrologic prediction. Code is available at https://github.com/Muhammad-Talha-MT/basin2vec
Mixture-of-Experts (MoE) models are widely deployed in modern machine learning. However, few convergence guarantees exist for gradient-based training of MoEs: the objective is nonconvex, and even simple mixtures trap first-order methods in spurious local optima. We analyze a canonical gated mixture featuring linear experts and a gate induced by Gaussian clusters in the input space, and give, to our knowledge, the first global convergence guarantee for jointly training the gate and the experts of such a model by gradient EM on its likelihood. Given a logarithmic overparameterization relative to the true number of experts $m$, gradient EM converges from randomly initialized gates and least-squares-initialized experts to any prescribed accuracy in polynomially many iterations, under a separation condition on the gate centers. Our analysis establishes a gradient-dominance inequality for the coupled dynamics by decoupling the loss into per-cluster subproblems. Empirical results validate the predicted $m\log m$ phase transition and suggest that the least-squares warm start required by our analysis is not needed in practice: randomly initialized experts perform nearly as well. We also find that leaving the mixing weights unconstrained concentrates routing onto the true number of experts, whereas load-balancing penalties spread it, so that the fitted model can be pruned to a fraction of its width without loss of accuracy.
Two methods are presented that perturb a given learning rate schedule in stochastic gradient descent with cross entropy loss. Both methods use an extra forward pass to estimate various derivatives of the minibatch loss. The first method selects a value $\eta$ by linearising the softmax values and minimizing the squared error $\tfrac 1 2 ||y - q - \eta q'||_F^2$, where $q$ is the matrix representing the softmax values of the minibatch and $y$ is the matrix with one-hot encodings of the labels. The second method considers a quadratic Taylor approximation of the minibatch loss under the outer approximation of the Hessian. Thus, it minimizes $L(0) + \eta L'(0) + \tfrac 1 2 \eta^2 L''(0)$. Here, the derivatives $L'$ and $L''$ are computed from the derivatives of the logits (using the outer approximation of the Hessian of the loss). We observe that both methods give very similar values of $\eta$. Finally, these estimated $\eta$ are used to slightly increase or decrease the learning rate of an existing schedule. The method is evaluated on Cifar10 and Cifar100 datasets for training Resnets, where they decrease the loss faster than the schedule that is perturbed and increase monotonicity. For Cifar10 it also leads to a small improvement of test accuracy, but in Cifar100 test accuracy remains the same.
Distribution shifts, including covariate shifts, are ubiquitous in real-world applications of machine learning. Empirical evidence suggests that, across a variety of models with and without downstream adaptation, ensembles tend to be more robust to covariate shift than single models. We extend the theory of covariate shift severity from single models to ensembles, providing theoretical justification for the empirical trends. We introduce new discrepancy measures of covariate shift severity to control the risk of partition-based kernel estimators. We then show that with sufficient ensemble size, these models can achieve minimax-optimal risk convergence rates under covariate shift. Our results suggest that modelers can expect to need larger ensemble sizes for more severe covariate shifts.
We address the problem of learning generative priors over high-dimensional latent variables from indirect, noisy observations, with a focus on continuous heterogeneity in cryo-electron microscopy (cryo-EM). In this setting, 3D atomic structures $x \in \mathbb{R}^{3N_a}$ are never directly observed and must be inferred from 2D projection images $y$ generated by a complex, non-linear forward operator. We propose Empirical Bayes Flow Matching (EB-FM), a method that learns a flow matching prior $p_\theta(x)$ directly from observations using a Monte Carlo stochastic approximation Expectation-Maximization framework. EB-FM alternates between an E-step that performs approximate posterior sampling from $p_\theta(x \mid y)$ via guided flow-based generation while maintaining persistent latent estimates through stochastic approximation, and an M-step that updates the flow model using these stabilized latent variables as training targets. This approach eliminates the need for pretraining on clean data and naturally accommodates non-linear forward models. We validate our method on controlled inverse problems using two-moons and MNIST data, and demonstrate its effectiveness on simulated cryo-EM data with known poses, showing that EB-FM can recover continuous distributions of protein conformations directly in atomic-coordinate space.
Scientific discovery requires plausible designs spanning useful trade-offs among competing properties: a molecule with strong binding affinity, for example, may be difficult to synthesize. Pretrained generative models, e.g., flow and diffusion generators, provide a foundation for proposing such designs, but steering them toward high-quality, complementary trade-offs remains challenging. Inference-time search incurs high per-sample costs, while fine-tuning typically relies on fixed per-candidate rewards that do not capture how designs complement one another. We introduce Hypervolume Fine-Tuning (HVFT), which adapts a single pretrained generator to maximize the expected hypervolume of its samples while controlling departure from the pretrained distribution. Expected hypervolume jointly rewards candidate quality and collective trade-off coverage. Nonetheless, we show that it cannot be expressed as the expectation of a fixed per-candidate reward and therefore cannot be optimized directly with standard techniques. A key observation of this work is that its first variation corresponds to the expected marginal hypervolume contribution of a candidate, which yields a tractable adaptive reward via an unbiased estimator based on model samples. HVFT uses this reward to decompose the objective into a sequence of KL-proximal scalar-reward subproblems. We establish concavity, relative smoothness, and convergence guarantees for the resulting distributional procedure. Across therapeutic-peptide and molecular design tasks spanning discrete diffusion and continuous flow-matching models, HVFT achieves state-of-the-art expected hypervolume and trade-off coverage.
We provide the first computationally efficient, differentially private algorithms for agnostically learning halfspaces. In particular, we study three different settings of agnostic learning: improper learning under Gaussian and log-concave marginals; proper learning under Gaussian marginals; and semi-agnostic learning under Gaussian marginals. In all three aforementioned settings, the added cost in the sample complexity due to privacy matches the respective non-private sample complexity, implying that we get privacy for free.
Flood peak depth on normally dry land governs damage and risk to life, but is observed only sparsely, by post-event high-water mark (HWM) surveys, and depth derived from satellite flood extent is produced after the event and inherits the vertical error of the digital elevation model (DEM) it is computed from. Deep-learning flood models have in turn been trained on hydrodynamic simulations rather than on observed depths. We ask whether flood peak depth can be learned from field observations and predicted before the peak from spaceborne terrain, geospatial foundation-model embeddings and antecedent meteorological forcing. We assembled 5689 USGS HWM surveys from 39 events (2011-2025) in 24 U.S. states and Puerto Rico, assigned each a peak time from nearby sensors, paired each with a rain-free window at the same site, and linked them to Copernicus GLO-30 terrain, AlphaEarth embeddings and 144 h of hourly forcing masked to the lead time. A cross-modal fusion network trained on these observations predicts depth at the surveyed marks with a mean absolute error of 0.39 m 6 h ahead of the peak. Supervision by surveyed depth corrects the DEM error implicitly: differencing the measured peak water level with the same DEM gives 0.93 m at the same marks, 2.4 times larger, although it uses the measured water level and the model does not. Forcing accounts for 77% of the attributed reduction in depth error; terrain and the embeddings are statistically equivalent to each other and contribute the remaining 23%. On eleven held-out events the model still separates flooded from rain-free conditions but no longer resolves depth at the marks (R2 = 0.03), so the skill reported here is confined to the events and regions sampled. Field-surveyed depths can thus supervise flood depth prediction and correct terrain error that extent-based estimation cannot, while transfer to unseen storms remains the open problem.
Fairness evaluation of clinical prediction models for binary outcomes often involves fairness metrics that compare the area under the receiver operating characteristic curve (AUC, or c-statistic) across subgroups. However, the AUC has been shown to be a non-collapsible metric, meaning that its subgroup specific and overall values are not directly comparable as often implicitly assumed. This raises important questions about how existing AUC-based fairness metrics should be used and interpreted. We examine eight AUC-based fairness metrics proposed in the literature, including: within-group AUC parity; cross-group AUC (xAUC) parity; gap-to-overall AUC; equity-scaled AUC; minimum-based metrics; $\mathrm{xAUC}^1$; $\mathrm{xAUC}^0$; and $\mathrm{xAUC}^1$ or $\mathrm{xAUC}^0$ parity. We characterise their estimands, relate them to the decomposition of overall AUC (into within- and cross-group components), and critically appraise them. We then illustrate their behaviour using simulations based on the EuroSCORE II, a clinical prediction model for in-hospital mortality. We recommend reporting an AUC matrix, containing the overall AUC and all within- and cross-group AUC components, alongside population characteristics. When fairness metrics are needed, we recommend reporting the within-group AUC parity, with cross-group AUC parity providing supplementary information but requiring caution due to its potential conflict with calibration. Metrics involving $\mathrm{xAUC}^1$ and $\mathrm{xAUC}^0$ have merit, summarising within- and cross-group AUCs into single values, but their interpretation is non-trivial, and they should be used with caution. Gap-to-overall AUC, equity-scaled AUC, and minimum-based metrics compare quantities that should not be compared and it is unclear how to operationalise them, making them not fit for purpose and generally not recommended.
End-to-end digital twins can be constructed by fitting separate stage models and connecting them after inference. We investigate the Frequentist prediction properties of a commonly used Monte Carlo simulation scheme that forwards complete predictive draws through such a chain when measured stage outputs contain independent normal errors, but the latent stage means are deterministic. Analytical two-stage examples show persistent asymptotic interval and coverage inflation and, with quadratic propagation, mean bias dependent on the parameters to be estimated. Simulations of four- and eight-stage linear-quadratic univariate chains with 7, 15 or 30 training runs demonstrate pronounced input-dependent distortions. A diagnostic using error-free latent training predictors shows that the effect is mainly not attributable to errors in variables, but to the MCSim approach even showing underestimation of coverage. We propose a model-based bootstrap: an initial chain forwards fitted means, bootstrap chains are refitted on fixed-center bootstrap data while forwarding their mean prediction, and a target-stage future error enters a studentized prediction root. Across the investigated scenarios, relative mean bias is substantially reduced, and coverage is closer to the nominal 90%. At sample size 30, the eight-stage bootstrap has mean width ratios of 1.16-1.42 relative to the known-parameter normal reference, compared with 4.11-4.97 for observed-to-observed MCSim. These results support improved endpoint prediction in the tested setting. Further investigations of the bootstrap scheme in the supplement for a bivariate 6-stage process show similar satisfactory results.
Implicit generative models based on deep neural networks have revolutionized many scientific fields, from weather research to language generation. They successfully model complex random objects arising in real-world problems. In this work, we propose a sample-efficient training mechanism for these models. Our approach leverages regularized inference with shrinkage priors and generalized Bayes. In the large $p$ (parameter dimension) and small $n$ (sample size) setting, sparsity-inducing priors have been successfully used for linear regression when $p \gg n$. However, their application to implicit generative models remains challenging due to the intractability of the likelihood function. To address this, we adapt a scoring rule posterior based on a strictly proper scoring rule, following the generalized Bayes framework. We then propose a block SGMCMC within Gibbs sampling mechanism. This enables efficient exploration of high-dimensional parameter spaces by identifying a sparse representation of neural implicit generative models. We demonstrate the performance of our proposed method through experiments on $p \gg n$ linear regression and an application of neural generative models.
Alternating online mirror descent (AltOMD) is a simple learning rule for zero-sum games, which improves upon simultaneous OMD numerically and theoretically. Previous work obtained $O(T^{1/3})$ and $O(T^{1/5})$ regret bounds for AltOMD, by leveraging modified-energy-function-based analyses. However, settling whether AltOMD achieves near-constant regret has remained an open question. In this paper we resolve the question of whether AltOMD achieves near-constant regret in the affirmative. In particular, we show an $O(log(T))$ regret bound on the sum of the players' external regrets for any two-player zero-sum bilinear game with compact convex strategy sets, under mild regularity assumptions on the regularizers. We verify that our assumptions hold for the negative entropy regularizer, thereby obtaining logarithmic regret for the alternating multiplicative weights updates algorithm. The regret bound also yields an $O(log(T)/T)$ duality gap bound for the averaged strategies.
Byte level language models avoid tokenizer dependence and provide universal coverage, but processing every byte at full resolution is expensive. We introduce DeepMem, a causal multiscale architecture that separates semantic compression from exact byte preservation. A learned invertible subband transform constructs a hierarchy of shorter sequences, while a lossless carrier and explicit detail sidebands preserve byte identity. Fast weight memory blocks operate along the encoder, with iterative shared state computation at the compressed bottleneck, while an attention based super resolution decoder restores byte resolution representations. DeepMem achieves stronger language modeling performance than a matched byte level Transformer on FineWeb-Edu and PG19, and obtains the best macro average performance across the evaluated FLORES languages. It also improves PIQA accuracy while remaining competitive on code modeling, multiple choice reasoning, and byte level corruption robustness. Bottleneck refinement requires approximately $7{,}600\times$ fewer counted FLOPs than equivalent full resolution reasoning, while maintaining competitive evaluation throughput.
De novo peptide sequencing has moved from dynamic programming over hand-designed or generative scores to large transformers that learn the task end to end. We present NovoCRF, a conditional random field (CRF) whose states are prefix masses, so that its paths are exactly the peptides within +/-0.06 Da of the precursor mass. A small network scores each edge from physics-derived ion-match features and is trained with its scores globally normalized over all mass-feasible peptides. To our knowledge, NovoCRF is the first de novo method trained with exact normalization over all mass-feasible peptides. Our software has three modules, each with exactly one job: a precursor mass finder recovers the accurate precursor mass from MS1, the CRF finds the peptides that fit the precursor mass by exact k-best decoding, and two BiGRU rerankers add sequence context and internal-fragment evidence. The model has 310k parameters and trains in about 17 GPU-hours. On seven held-out data-dependent acquisition datasets published in 2026 (126,262 labeled spectra), NovoCRF's mean peptide recall is 42.4%, against 39.1% for Casanovo v5 with five beams, and its amino-acid recall 64.0% against 53.1%, at 1,746 spectra/s on 18 CPU threads, 55x Casanovo's throughput on a GPU. Split by fragment evidence, NovoCRF leads on labels with up to three peptide bonds without a b or y ion, and Casanovo only beyond. NovoCRF can also search modifications it was never trained on at their exact mass: it finds 32.1% of the hydroxyproline peptides of a lion bone sample.
Naive on-policy RL is often bottlenecked by the capabilities of the base model --- insufficient coverage can result in an inability to learn from certain rewards. At the same time, rewards in modern post-training systems often come from rich textual sources, e.g. judge models with a rubric of criteria. Can we utilize such structures to provide a rich directional signal that enables improving on otherwise difficult to optimize rewards? In this work, we examine a framework we refer to as targeted guidance that learns from per-turn textual feedback via an on-policy self-distillation objective. Our focus is not only on directly improving hard-to-optimize behaviors, but on how to do so throughout the training process without degrading baseline RL performance. Over a range of settings across an agentic multi-turn software engineering environment, we find success by combining targeted self-distillation with a per-criteria RL reward, and by limiting feedback to turns where the policy specifically fails. We further examine the design choices that lead to stable entropy and reduce oscillatory behavior throughout training.
Antibody affinity maturation is a natural optimization process in which B cells acquire mutations through somatic hypermutation (SHM), undergo selection for improved antigen binding, and iteratively evolve toward higher affinity. Computationally modeling this process requires capturing both biologically plausible mutations and antigen-specific selection. However, existing mutation models do not directly optimize binding, while affinity optimization often neglects realistic mutation trajectories. We introduce iGC (in silico Germinal Centers), a two-stage reinforcement learning (RL) framework for antigen-specific antibody optimization. Stage 1 learns an SHM mutation prior from parent-child pairs in B-cell lineage trees and adapts it to a target affinity-maturation dataset. Stage 2 initializes an RL policy from this prior and optimizes antigen binding with Kullback-Leibler (KL) regularization toward the frozen prior. To prevent reward exploitation, including repetitive high-reward substitutions and structurally tolerated framework mutations, we introduce a Reward-Tax Meta-Gradient objective. RewardNet smooths the affinity landscape, while TaxNet penalizes reward overexploitation, jointly defining a learned meta-reward for policy optimization. We evaluate iGC on mutation prediction, SHM consistency, affinity optimization, and mutation diversity. The mutation prior more than doubles the strongest baseline in mutation-site prediction. Following RL optimization, iGC maintains SHM sequence structure. It also achieves a strong exploration-exploitation balance, yielding competitive reward-diversity products across candidate-selection cutoffs. These results demonstrate that iGC integrates biologically plausible mutation generation with antigen-specific affinity selection, providing an in silico framework for iterative antibody affinity maturation.
We attempt to alleviate failure modes of classical physics-informed neural networks (PINNs) in scenarios of topological pathologies corresponding to punctures, holes, poles, and voids by introducing de Rham cohomological augmentations, or Hodge-de Rham PINNs. Vanilla PINNs often encounter pitfalls in hole-based settings due to discontinuous interfaces, nonzero closed path integration, path dependence, correspondence to exact forms via the exterior derivative, etc., therefore we absolve traditional failures via parameterizing a 1-form by the Hodge-Morrey-Friedrichs (HMF) decomposition for manifolds with boundary, drawing a connection to de Rham cohomology such as through the HMF isomorphism theorem. 1-forms can act as representations of scalar-valued PDE solutions through integral operator mappings. We propose performing a PINN-based loss on the parameterized 1-form through an associated PDE, which subsequently solves the potential-based PINN problem and moreover remedies the failures of vanilla PINNs via the Sobolev-regular operator under a Fr\'echet differentiation interconnection. The operator relationship often has a nontrivial closed-form and often requires integration or a case-by-case analysis, which we address thoroughly in our experiments. When an integrability condition is not met on our losses, we can equip exhaustion functions to train the networks. We demonstrate our methods on the Aharonov-Bohm effect, which, to our understanding, is largely nonexistent in PINN literature. We develop theoretical guarantees of our methods and demonstrate our pathological safeguards on a broad selection of topologically-adverse problems, including scenarios of inexact forms, and those of nontrivial harmonics and coexact terms.
We prove a broad learning-theoretic generalization of the celebrated hardcore lemma concerning the average-case complexity of computing a Boolean function (Impagliazzo, 1995). Our starting point is the observation that the classic notion of a hardcore measure, which reweights a function's input distribution, is equivalent to an input-dependent masking channel that injects randomness into the function's output distribution until it appears fully pseudorandom. The analogue of density of the measure is the fidelity of the channel. From this viewpoint, we develop a general framework for certifying hardness that allows for non-Boolean functions $f$ and a wide range of notions of hardness besides the classic 0-1 loss, by choosing the right notion of fidelity via convex duality. We show the difficulty of learning $f$ using a class $\mathcal{C}$ to minimize a convex loss is tightly characterized by the maximum fidelity possible for a masking channel, which measures the least amount of noise the channel must add to hide the structure of $f$ and make it $\mathcal{C}$-indistinguishable from random noise. Using our framework, we derive a number of multiclass generalizations of the hardcore lemma, and reveal some unexpected subtleties in going to larger alphabets. We construct a universal masking channel that applies to many natural notions of hardness simultaneously. Such channels are dual to omnipredictors, which optimize multiple losses at once (Gopalan, Kalai, Reingold, Sharan, and Wieder, 2022). We observe a symmetry between masking channels and generalizations of the hardcore lemma using pseudoentropy. We present a tight characterization of pseudoentropy using the MaxEnt algorithm from machine learning. A long line of research has shown that learning and hardness are two sides of the same coin; our framework gives a crisp and general formulation of this principle via convex duality.
Looped transformers are a parameter-efficient alternative to standard transformers for language modeling. By reapplying the same layers across multiple loop steps, recent looped language models match the performance of non-looped models several times their size. While this architecture effectively reduces parameter memory, each loop step still stores its own keys and values, so the key-value (KV) cache grows linearly with the number of loop steps and becomes the main memory bottleneck. We observe that the KV cache of looped models is highly redundant along the loop axis and that the loop steps differ substantially in their sensitivity to quantization. Based on these observations, we propose $\textbf{LoopVQ}$, a KV cache vector quantization method for looped transformers that codes channels of multiple loop steps jointly, and fine-tunes the codebooks to match the output of the full-precision model. At 0.5 bits per element, LoopVQ outperforms existing KV cache vector quantization methods for standard LLMs on general, long chain-of-thought reasoning, and long-context retrieval tasks with Ouro and Huginn models. Its design is orthogonal to these methods, and combined with TaSQ, it retains much of the BF16 accuracy even at 0.25 bits per element. We also implement LoopVQ in SGLang, where it nearly halves the decode latency at a 32K-token context and raises the peak serving throughput by up to $4.3\times$ and $6.1\times$ on the 1.4B and 2.6B Ouro models compared to the BF16 baseline.
Quantifying selection effects in gravitational-wave population studies requires millions of costly signal-to-noise ratio (SNR) calculations. We present $\Phi$asmaNet, a compact neural architecture for structured tabular regression, with staged spectral refinement as an explicit architectural prior. Learned input tokens are refined through ordered attention-based $\Phi$-Stages, each augmented by a $\Phi$-Refractor subspace-mixing module, then revisited through gated top-down $\Phi$-Feedback. Applied to the 11-dimensional parameters of precessing, higher-mode binary black hole signals (IMRPhenomXPHM), it predicts phase-averaged optimal SNR for a single detector at fixed Advanced LIGO design sensitivity. With approximately 732k parameters, it outperforms MLP, 1D-ResNet and Transformer-encoder baselines under matched training across $10^5$--$10^6$ samples. At $10^6$ samples, it achieves $100-\mathrm{MAPE}=98.73\%$ on the held-out test set, $99.29\%$ after rank-preserving monotonic (PCHIP) calibration, and $99.39\%$ ($R^2=0.9998$) in the detection-relevant regime $\rho>8$, with stable accuracy across independent dataset realizations. On a single GPU, it sustains approximately $1.2\times10^5$ SNR evaluations per second and processes a $10^6$-injection catalogue end-to-end in approximately 8.3 s, about two orders of magnitude faster than conventional waveform-and-inner-product calculations on a 16-core CPU. Exact distance rescaling through $\rho\propto1/d_L$ supports repeated catalogue-wide detectability evaluations. The fully documented pipeline is intended as a reproducible benchmark, while results on public tabular datasets indicate competitiveness beyond gravitational-wave applications. Although demonstrated for a fixed single-detector configuration, the construction supports extensions to detector networks and epoch-specific sensitivities. (Abridged)
We study context-aware time-series forecasting, in which a single target variable is predicted from its own history and a set of contextual variables, such as weather and calendar information. As the context becomes more diverse, a single model struggles to capture all of its regimes, while ensemble learning and Mixture-of-Experts approaches rely on manual data partitioning or soft routing, which increases the inference cost and hinders interpretability. We propose CATF, a Context-Aware Time-Series Forecasting framework built on a Manager-Worker architecture: a Transformer-based Manager routes each input to a single Worker based on the context, using hard routing during both training and inference. The Workers thus specialize without manual data partitioning, and any single-model forecaster can serve as a Worker. On five context-rich datasets, four horizons, and two normalization schemes, CATF improves over its standalone counterpart in 162 of 200 comparisons, the best CATF-enhanced model outperforms the best standalone baseline in 19 of 20 settings, and it outperforms the zero-shot Time-MoE foundation model in 15 of 20 settings. A case study on bike-sharing data shows that the Manager routes inputs by the calendar and that each Worker specializes in a different weekly pattern.
Given a class of machine learning models and sample access to a data distribution with binary labels, which computational task is harder: training a discriminative model that distinguishes positive and negative labels as well as the best member of the class, or training a generative model whose outputs are indistinguishable from real data by all members of the class? We prove that these two fundamental tasks in computational learning theory are not just closely related, but actually equivalent to each other in complexity. One direction of this equivalence was previously known. Indeed, the powerful regularity lemma in complexity theory gives an efficient oracle reduction from the indistinguishability task (a.k.a. multiaccuracy) to the distinguishing task (a.k.a. agnostic learning) in the style of boosting algorithms. Our main contribution is an efficient oracle reduction in the other direction. Taking this idea further, we show that a multiaccuracy oracle is a remarkably powerful primitive, achieving not just agnostic learning but also a wide range of seemingly more demanding learning objectives via similarly efficient and direct constructions. Our results crucially use the power of making adaptive queries to the oracle, and stand in stark contrast to prior work, which showed that post-processing a single multiaccurate predictor is insufficient for even weak forms of learning. On a technical level, we draw a connection between our (offline) problem and a version of no-regret learning in the online setting, which allows us to use follow-the-leader type algorithms to modify the label distribution and aggregate the predictions of the multiaccuracy oracle.
Direct Nash Optimization (DNO) aligns a language model to a general preference function by approximating the Nash equilibrium of a two-player preference game. Its practical variant, DNO-Prct, iterates on-policy sampling, LLM-based rating of the samples together with a teacher's gold response, large-margin preference-pair selection and a DPO-style contrastive update. We re-implement DNO-Prct independently and study it on two open-weight instruct models, Llama 3.1 8B Instruct and Mistral 7B Instruct v0.1, with GPT-4o or a larger open-weight model of the same family as the annotator. On AlpacaEval 2.0 with GPT-4o as judge, three iterations raise the length-controlled (LC) win rate against GPT-4o from 31.7% to 45.3% for Llama and from 6.8% to 18.4% for Mistral, and the tuned 8B model scores above Llama 3.1 70B Instruct on every reported win rate. With same-family open-weight annotators, the runs retain 62% (Llama) and 82% (Mistral) of the LC gain achieved with GPT-4o as annotator, so the rating step does not require a proprietary model. In a separate ablation configuration, variants that remove the gold responses, remove the data look-back, or accept only teacher-preferred pairs score 6 to 10 LC points below the full procedure, whereas varying the number of on-policy samples per prompt (1 to 7) moves the LC win rate by at most 2.1 points, and a margin sweep (1 to 5) run in its own pipeline by at most 2.4 points. We also report the cost of the pipeline and where the practical algorithm departs from the theory it approximates. All results are single runs, and GPT-4o is both the teacher and the evaluation judge; we discuss what this implies for the reported gains.
Preserving the privacy of preferences (or rewards) of a sequential decision-making agent when decisions are observable is crucial in many physical and cybersecurity domains. For instance, in wildlife monitoring, agents must allocate patrolling resources without revealing animal locations to poachers. This paper addresses privacy preservation in planning over a sequence of actions in MDPs, where the reward function represents the preference structure to be protected. Observers can use Inverse RL (IRL) to learn these preferences, making this a challenging task. Current research on differential privacy in reward functions fails to ensure guarantee on the minimum expected reward and offers theoretical guarantees that are inadequate against IRL-based observers. To bridge this gap, we propose a novel approach rooted in the theory of deception. Deception includes two models: dissimulation (hiding the truth) and simulation (showing the wrong). Our first contribution theoretically demonstrates significant privacy leaks in existing dissimulation-based methods. Our second contribution is a novel RL-based planning algorithm that uses simulation to effectively address these privacy concerns while ensuring a guarantee on the expected reward. Experiments on multiple benchmark problems show that our approach outperforms previous methods in preserving reward function privacy.
Visuomotor policies trained via behavior cloning are vulnerable to covariate shift, where small deviations from expert trajectories can compound into failure. Common strategies to mitigate this issue involve expanding the training distribution through human-in-the-loop corrections or synthetic data augmentation. However, these approaches are often labor-intensive, rely on strong task assumptions, or compromise the quality of imitation. We introduce Latent Policy Barrier, a framework for robust visuomotor policy learning. Inspired by Control Barrier Functions, LPB treats the latent embeddings of expert demonstrations as an implicit barrier separating safe, in-distribution states from unsafe, out-of-distribution (OOD) ones. Our approach decouples the role of precise expert imitation and OOD recovery into two separate modules: a base diffusion policy solely on expert data, and a dynamics model trained on both expert and suboptimal policy rollout data. At inference time, the dynamics model predicts future latent states and optimizes them to stay within the expert distribution. Both simulated and real-world experiments show that LPB improves both policy robustness and data efficiency, enabling reliable manipulation from limited expert data and without additional human correction or annotation.
Adversarial Imitation Learning (AIL) methods, while effective in settings with limited expert demonstrations, are often considered unstable. These approaches typically decompose into two components: Density Ratio (DR) estimation $\frac{\rho_E}{\rho_{\pi}}$, where a discriminator estimates the relative occupancy of state-action pairs under the policy versus the expert; and Reward Assignment (RA), where this ratio is transformed into a reward signal used to train the policy. While significant research has focused on improving density estimation, the role of reward assignment in influencing training dynamics and final policy performance has been largely overlooked. RA functions in AIL are typically derived from divergence minimization objectives, relying heavily on human design and ingenuity. In this work, we take a different approach: we investigate the discovery of data-driven RA functions, i.e, based directly on the performance of the resulting imitation policy. To this end, we leverage an LLM-guided evolutionary framework that efficiently explores the space of RA functions, yielding \emph{Discovered Adversarial Imitation Learning} (DAIL), the first meta-learnt AIL algorithm. Remarkably, DAIL generalises across unseen environments and policy optimization algorithms, outperforming the current state-of-the-art of \emph{human-designed} baselines. Finally, we analyse why DAIL leads to more stable training, offering novel insights into the role of RA functions in the stability of AIL. Code is publicly available: https://github.com/shshnkreddy/DAIL.
Linear activation steering is a powerful approach for controlling and specializing large language models by adding linear representations of concepts to their activations at inference. While effective, existing methods often steer with a fixed strength, independent of how much steering any given input requires. In this work, we introduce Contextual Linear Activation Steering (CLAS), which decomposes activation steering into a fixed steering direction and a context-dependent steering coefficient. This decomposition preserves the interpretability of the steering direction while enabling the steering strength to vary with context. It also enables CLAS to perform selective steering: steering multiple behaviors across different portions of a single generation, which cannot be achieved using a fixed steering strength. In addition, across a wide range of steering benchmarks and language models, CLAS consistently outperforms standard linear activation steering and matches or exceeds ReFT and LoRA in settings with limited labeled data. Notably, CLAS achieves this using directions that also serve as reliable concept monitors, a property that directions learned via fine-tuning lack. We therefore propose CLAS as an interpretable and accurate method for controlling and specializing large language models.
Large Language Model (LLM) serving has become a critical cloud workload, and realistic traces are essential for motivating and benchmarking serving systems. However, existing LLM serving workload studies remain limited in scale and scope. They often observe short time periods and provide limited visibility into how users interact with models in production. As a result, they do not fully capture how LLM serving workloads evolve over time or how user-model interactions shape production traffic. In this work, we further the understanding of real-world LLM serving workloads through both a global characterization and a longitudinal study of a one-year production trace from CompanyX. Unlike prior studies, our trace captures full production behavior across many models and users, including both popular and long-tail models. We analyze the workload from aggregate, temporal, model-level, and user-level perspectives, revealing workload evolution and user-model structure that are typically hidden behind aggregate views. To support future research, we publicly release the full one-year trace, enabling downstream studies of production behavior without relying on sampled or synthetically generated workloads. The trace is available at https://github.com/HarvardMadSys/chutes_workload.
The urban heat island (UHI) effect amplifies heat stress in cities, increasing heat-related illness, mortality, sleep disruption, and energy demand. Because UHI intensity varies at neighborhood scales, intra-city temperature mapping is essential for assessing public health impacts, yet existing monitoring networks are sparse. Crowdsourced weather stations provide dense, hourly observations but suffer from noise and systematic biases, including preferential sampling driven by where residents install sensors. We introduce two scalable Gaussian process frameworks which adjust for preferential sampling and enable statewide, high-resolution temperature prediction. The first integrates sampling correction directly within a sparse variational Gaussian process (SVGPR+PS). The second uses a two-stage estimator involving a log-Gaussian Cox process (LGCP+GP) that estimates the sampling intensity field and incorporates it into temperature prediction. Both methods rely on inducing-point approximations and variational inference to scale beyond traditional Gaussian processes. Applied to approximately 10,000 citizen stations across North Carolina, the models generate hourly temperature fields within minutes. Validation against independent urban heat campaign data shows that the preferential sampling correction improves predictive performance, increasing average afternoon correlations across four cities from 0.26 to 0.43. This scalable framework enables neighborhood-level UHI assessment and supports studies of heat exposure disparities statewide.
Probabilistic forecasts of day-ahead electricity prices are increasingly used to manage price risk, yet competing models are usually ranked on aggregate scores that average over the price regimes in which uncertainty matters most. This study evaluates probabilistic forecasts of hourly EPEX SPOT GB day-ahead prices over a full held-out calendar year, applying a common pre-auction information boundary and rolling-origin design to linear quantile regression, a quantile regression forest (QRF), a LEAR-QRA benchmark and an encoder-decoder quantile long short-term memory (LSTM) network. Two routes to more reliable intervals are then compared: conformal recalibration and Quantile Regression Averaging (QRA). Aggregate scores separate the LSTM and QRF only weakly, with a non-significant 4.4% difference in continuous ranked probability score (CRPS), but their conditional behaviour differs sharply. Above 200 GBP/MWh the LSTM covered 0.12 of outcomes and QRF 0.71, consistent with the forest retaining realised spikes in its leaf distributions while the network learned a smooth upper tail. Conformal recalibration corrected marginal coverage but did not improve the interval score or materially repair spike coverage. QRA combining the LSTM, QRF and LEAR-QRA forecasts reduced CRPS to 5.14 GBP/MWh and raised coverage above 200 GBP/MWh to 0.65 at unchanged interval width. Recalibration and combination therefore address different failures and are not substitutes.
We develop ML models that predict when cross-factor predictability will erode, enabling prospective risk management. A decay prediction model achieves C-index $= 0.93$ on held-out factor pairs, providing early warning of signal erosion---the key predictor is initial signal strength: stronger signals decay faster (``regression to the mean''). An online regime detector enables real-time classification with 6.6-day average detection delay. Prior work explains why factors decay (arbitrage, publication effects); we predict when decay will occur. Limitation: The primary HML$\to$SMB signal does not replicate in post-break OOS (2013--2024); this null result is consistent with genuine decay but limits confirmatory claims. Motivation: Among 30 directed Fama-French factor pairs (1990--2024), predictability targeting the Size factor (SMB) decays fastest. Of 7 pairs showing exponential decay ($R^2 > 0.3$), 4 target SMB (Mkt-RF$\to$SMB, CMA$\to$SMB, HML$\to$SMB, RMW$\to$SMB; half-lives 2.4--4.1 years). This SMB-concentration pattern motivates the predictive models. Methodology: Using a regime-conditional Granger framework with Student-$t$ HMMs, we document structural breaks clustering around June 1998 ($p < 10^{-10}$). A pre-break OOS test (HMM trained 1990--1995, tested 1996--2007) confirms signals existed: HML$\to$SMB shows $p = 0.013$ (Elevated) and $p < 0.001$ (Crisis). The primary HML$\to$SMB signal does not replicate in post-break OOS (2013--2024), consistent with genuine decay rather than overfitting---after ${\sim}6$ half-lives, the expected signal is $<$1% of its 1990s peak. International analysis finds 2/4 non-US markets with Bonferroni-surviving OOS effects; correction for full specification search was not applied, so results remain exploratory.
Methyl acetate production via direct carbonylation of dimethyl ether and carbon monoxide involves strong nonlinear interactions among reactor operation, recycle structure, and energy consumption, making conventional simulation-based optimization computationally demanding. In this work, a data-driven framework integrating steady-state process simulation, surrogate modeling, and Bayesian optimization was developed for process intensification. A dataset of 2730 steady-state simulations was generated using a Python-based automated process simulation workflow and a structured design of experiments across the operating domain. Surrogate models were trained to predict carbon conversion yield (C-yield) and cost of specific energy consumption (C-SEC), with neural network-based models showing the best predictive performance and robustness. These models were integrated into a multi-objective Bayesian optimization framework to identify optimal operating regions. Results indicate that high C-yield (up to ~95%) can be achieved with reduced energy cost through lower recycle flow rates and compression demand. Pareto analysis revealed a nonlinear trade-off between C-yield and C-SEC, where further improvements near the conversion limit require disproportionately higher energy input. Validation against rigorous simulations showed deviations below 3% for C-yield and 10% for C-SEC.
Molecular design requires coordinated generation, property prediction and iterative editing, with each decision linked to the correct chemical structure and supporting evidence. Integrating these capabilities in practice presents three major challenges: retaining local control of molecular data, connecting specialized models across tasks and verifying that proposed changes satisfy chemical constraints. Here we present MolCrew, a locally deployable system that coordinates four separately trained molecular specialists with a language-model orchestrator and RDKit calculations. The specialists generate structures from descriptions, describe molecular structures, predict biological endpoints and propose edits towards specified goals. Separate model weights allow each capability to be evaluated and updated independently within the workflow. RDKit calculations check structural validity, quantify descriptor changes and assess specified structural constraints, while returned observations guide further tool calls. We evaluated low-rank adaptation of the orchestrator for tool-use decisions and failure recovery while leaving specialist weights unchanged. On 3,300 ChEBI-20 test examples, MolCrew achieved 96.4% exact structure recovery, compared with 38.7% for published Atomas-large results. Across eight MoleculeNet datasets evaluated using scaffold splits, mean area under the receiver operating characteristic curve was 0.851, compared with 0.770 for published Atomas results. In a recorded evaluation using the unadapted orchestrator, 1,554 of 1,699 requests (91.5%) passed every predefined workflow check. By preserving inspectable links between molecular proposals, model outputs and chemical calculations, MolCrew provides a modular framework for computational candidate assessment and refinement under local data control.
Learned temporal models are moving beyond offline forecasting toward interactive environments for decision making, yet logged-trajectory accuracy need not predict closed-loop behavior under policies optimized in the learned model. This forecasting-control gap matters in tokamak magnetic control, where coupled plasma-circuit dynamics and actuator constraints can amplify errors through feedback. Using a calibrated free-boundary Grad-Shafranov solver, we collect four datasets in a 2x2 design over physics-parameter and current-reference coverage. At matched budgets of 500, 1k, 3k, and 5k rollouts, we train 16 action-conditioned Transformers; PPO policies optimized in the 5k surrogates are then evaluated without adaptation on six reference-solver conditions. Only joint-coverage D4 shows monotonically decreasing cross-condition mean error, whereas D1 improves in distribution but degrades out of distribution. Binary completion masks substantial closed-loop differences: D2-D4 each complete 83.3% of conditions, but reach on average 99.0%, 94.7%, and 90.5% of the evaluation horizon, while their worst conditions reach 94.2%, 68.2%, and 43.2%. Policies trained with action noise complete 5/24 matched conditions, compared with 19/24 under nominal training, and have a lower mean horizon-capped episode length. Based on one trained policy per configuration and one evaluation realization per condition, the results support combining prediction error, episode completion, condition-level duration, and tracking quality.
Locating competing reaction pathways is difficult when potential-energy gradients are unavailable. Here we present tReax\_expedition, an energy-only approach combining coupled Monte Carlo exploration, stochastic quench labelling and restrained continuation. Exploration identifies connected states, while continuation resolves candidate channels and records sampled chain maxima. Single-search benchmarks on three two-dimensional potentials recovered an intermediate and competing routes without transition-state guesses. For eight of nine gas-phase reactions on GFN2-xTB, sampled chain maxima differed from independently refined saddle energies by less than 0.76~kJ\,mol$^{-1}$. Alanine dipeptide instead revealed incomplete conformational relaxation, showing that route connectivity can persist despite structural and energetic errors. Neural committor learning and local straddling-pair sampling connect the resulting paths to transition-region sampling, although molecular shooting diagnostics indicate limited committor calibration. These results show how energy-only exploration and continuation recover candidate pathways, while distinguishing route connectivity from geometric localization, barrier resolution and finite-temperature transition statistics.
Before a scientist fits any parameter, they must decide what the model contains: which quantities exist, which are hidden, how they interact, and how they are measured. Physicists call such a trial form an ansatz. Benchmarks for automated equation discovery mostly supply it, by fixing the variables, a library of candidate terms, or a law to be modified. We introduce Unsatz, a benchmark in which an LLM agent must construct the ansatz itself. Each task is a simulated system, generated from a private seed as a program in a small modeling language: ordinary, delay, or partial differential equations, discrete maps, static and causal models, or stochastic processes, possibly with hidden states, distorted sensors, and uninformative measurement channels. The agent buys experiments within a budget, writes candidate programs in the same language, obtains their coefficients from the benchmark's fitter, and is scored only on how its final program predicts experiments it never saw: conditions outside the range it could explore, interventions, and longer runs. Across 80 systems in 20 domains, nine language models solve between 21% and 60%. On 44 of these systems, classical equation-discovery methods almost never succeed, although the same data usually suffice once the true equations are given with unknown coefficients; the difficulty lies in finding the structure, not in estimating it. Because mechanisms recur across systems, we also give agents the program they discovered for a related system; it helped only when the new system shared the mechanism and the first investigation had succeeded. Stronger agents make fewer unscientific errors; beyond these, agents find which variables a system has but not how they act, and prefer revisions that lower the training error to revisions that predict better.
Deep Research (DR) systems extend large language models into autonomous agents that decompose complex information needs, iteratively search the web, and synthesize evidence to answer challenging questions. However, their multi-stage execution may introduce robustness risks, as variations in user questions or intermediate search inputs may propagate across research stages and lead to incorrect answers. Existing evaluations focus on task-solving capabilities or robustness against external evidence disturbances, leaving intrinsic weaknesses in DR execution insufficiently explored. Thus, we conduct an empirical study of 1,932 execution trajectories from three representative DR systems and identify nine recurring defect categories across subquestion formulation, online search, and final reasoning. Guided by these findings, we propose EMBERS, an empirically motivated, defect-guided fuzzing framework for robustness testing of DR systems. Starting from correctly answered seed questions, EMBERS generates stage-targeted test inputs using defect-guided mutation, web-search-assisted fidelity assurance, and diversity-aware seed maintenance. Experiments across 16 framework--model--dataset settings show that EMBERS achieves 3.21 times the failure-discovery throughput of the strongest baseline under identical testing budgets. Human and LLM-assisted evaluations demonstrate that most generated inputs preserve the original information need and necessary contextual constraints. Moreover, model fine-tuning experiments show consistent-answer performance improvement, highlighting their value for robustness testing and model improvement.
Next point-of-interest (POI) recommendation supports personalized place discovery on location-based Web platforms. Mobility routines can recur across users while being realized at different places and adapted to individual preferences. We present DualClusterNet, a causal sequential recommender that composes two population-shared prototype dictionaries. At each observed prefix, soft routers retrieve pattern and spatial mixtures; a factorized multiplicative interaction couples them, feature-wise affine modulation adapts the shared representation, and a separate user-conditioned branch with an additive--projective gate retains individual variation. The prediction head combines this representation with a learned first-order transition term and ranks the complete training POI vocabulary. We evaluate on NYC and TKY using a training-only vocabulary and an original-final-target protocol that neither bridges unknown POIs nor substitutes earlier targets. The available single-seed results show competitive full-candidate ranking, while controlled analyses support the contribution of the compositional and personalized branches. Code is available at https://anonymous.4open.science/r/DualClusterNet-F4EE/.
Metalloprotein docking requires a ligand pose to satisfy both global pocket fit and the identity and geometry of metal-coordinating atoms. General-purpose diffusion docking models do not explicitly condition generation on externally specified complete ligand donor sets, and their predicted poses can exhibit incorrect metal-donor relationships. Metal-aware models address coordination during training and generation; here we ask whether external donor information can correct already-trained general-purpose models without retraining. We propose Complete-Identity Contract Guidance (CICG), which applies a supplied complete donor set through translation, rotation, and torsion updates, and Historical Contract Transfer (HCT), which supplies donor identities from earlier complexes at compatible receptor sites. We compare HCT with trajectory-local controls retaining the historical donor count with or without elemental composition. On 86 assessable HCT-covered queries, HCT-guided CICG increases frozen-confidence-selected joint success from 21.5% to 36.6% on SurfDock and from 25.0% to 34.9% on DiffDock-L. The joint criterion requires correct global ligand placement, complete native donor recovery, coordination geometry, preserved stereochemistry, and absence of fatal receptor clashes. Composition-matched guidance reaches 33.14% and 34.30%, leaving the additional benefit of fixed historical membership unresolved. Our results therefore support donor-set guidance as a conditional correction for frozen docking models, while identifying donor-source reliability and physical-quality preservation as the main barriers to broader use.
We present HyVox, a hybrid neuromorphic--von Neumann system for low-energy open-set speaker identification under tight energy budgets. HyVox splits the workload at the speaker embedding. A neuromorphic processor embeds the speech with sparse, event-driven computation, and a von Neumann microcontroller holds the speaker gallery and makes every decision: it scores queries by cosine similarity, rejects unknown speakers, and enrolls new users by writing a prototype, with no weight update. On the neuromorphic processor, a hierarchical temporal abstraction (HTA) SNN backbone learns a synaptic time constant per leaky integrate-and-fire (LIF) neuron, allocating 832 neurons across fast, medium, and slow timescales. A 32-neuron spiking embedding extractor (SEE), pretrained on episodes that each enroll new speakers and reject outsiders, turns the backbone output into an embedding for speakers unseen in training; its 32 spike counts are the only interface between the two processors. To fit the chip's synapse and integer budgets, we compress the backbone 6.5x to 29.0k synapses, train it for integer bit-shift dynamics, and fine-tune it with chip-accurate spiking training (CAST). CAST trains through a differentiable replica of the chip's integer neuron update, derived from its cycle-accurate simulator, so that gradient descent optimizes the exact weights and thresholds the chip runs; it doubles the top-1 accuracy of the mapped backbone on unseen speakers. On 20 Common Voice households of unseen speakers, HyVox reaches 99.6\% top-1 and 5.5\% rejection EER with 2 members and 94.9\% and 13.2\% with 20, and the network deployed on the neuromorphic chip matches its trained model. Each inference on a 3-s window takes 0.96 s and 1.63 mJ of dynamic energy, 42\% lower latency than the fastest embedded ANN baseline and 13.5x less dynamic and 5.8x less total energy than the most energy-efficient one.
Most malware detection research focuses on identifying whether a file is malicious before it executes. Far less attention goes to what happens after infection. This paper presents an empirical analysis of post-infection malware behavior across 37 families, grounded in dynamic sandbox execution using CAPEv2. Behavioral traces are modeled as heterogeneous graphs with five node types and seven typed edge relationships, preserving relational structure that flat feature vectors discard. A voting ensemble trained on graph-derived structural features achieves 94.44% accuracy and MCC of 0.863, outperforming a flat behavioral baseline by 7.4 percentage points in MCC. A multiclass classifier distinguishes 24 malware families, and a per-family behavioral fingerprint analysis characterizes post-infection activity across registry, network, process, and API dimensions. An LLM explanation layer generates analyst-readable reports with MITRE ATT&CK mappings grounded in observed behavioral evidence. The full pipeline is deployed as HeteroScan at https://ama-andam.github.io/heteroscan.
Mapping the three-dimensional distribution of dark matter is key to tracing the growth of cosmic structure. Two-dimensional weak lensing mass maps recover its projected distribution, but they blend structures at different distances along the line of sight. We reconstruct the three-dimensional matter distribution in the 0.54 deg$^2$ COSMOS-Web field using deep James Webb Space Telescope imaging, with an effective source density of 121 galaxies arcmin$^{-2}$ We combine a physical weak lensing forward model with a diffusion-based Bayesian generative prior learned from cosmological simulations, yielding a posterior distribution of the matter field across 20 lens planes to $z \simeq 2$, doubling the redshift reach of previous Hubble Space Telescope-based 3D reconstructions and approaching cosmic noon, an epoch of rapid galaxy growth. The reconstruction separates structures that overlap in projection and recovers non-Gaussian matter structure across a range of spatial scales, including over-densities associated with independently identified galaxy clusters that were not detected in earlier weak lensing maps. Posterior sampling quantifies the range of three-dimensional matter distributions allowed by the data, enabling a probabilistic visualization of the cosmic web while propagating uncertainties in its inferred structure. Null and simulation-based tests show that the recovered structures are supported by the lensing signal and are consistent with expectations from the fiducial $\Lambda$CDM cosmology. These results extend three-dimensional weak lensing mapping into the high-redshift, non-Gaussian regime and establish a route towards probabilistic mapping of the evolving cosmic web with current wide-field lensing surveys.
Genomic foundation models are increasingly used to propose DNA sequences and edits. But successful design requires capabilities that current benchmarks do not measure. Here we introduce BEND 2.0 (BENchmark for DNA language models), which tests how well models predict the effects of DNA changes and whether these predictions can guide sequence design. Across eight tasks built on five experimental datasets, we find that model choice and how representations are accessed both affect performance. Linear readouts fitted to frozen representations can match or exceed models' native outputs, and functional post-training substantially improves the models' capabilities. Linear readouts transfer well across experimental assays and genes, suggesting that models capture functional patterns beyond the data used to fit individual predictors. At the same time, strong prediction of average effects can conceal weaker prediction of context-specific effects, and accurate prediction of combined mutation effects does not imply accurate prediction of their interactions. In design tasks, model predictions can reduce the number of candidates needed to find a desired response, but still leave substantial room for better design choices, particularly for context-specific and multi-cell objectives. Our benchmark reveals that genomic models now display signs of general utility but have yet to saturate on performance in even simple tasks.
AIRoA-0 (A-0) is a bimanual manipulation system learned from 16,151 hours of handheld demonstrations, with no robot data collected at any stage. The demonstrations are recorded with YUBI, our open-source handheld gripper, which lets people record manipulation data anywhere with no robot present. From a single instruction, A-0 sorts more than a dozen small parts into their numbered cells and changes the mop pads and brush of a robot vacuum; among other tasks, it unzips a pencil case and folds a towel. It performs these tasks using only a camera on each wrist, with no head or scene camera, at a site where none of its training data was recorded. Across the ten tasks of our evaluation, it averages a success rate of 78%, against 27% for the strongest of four other vision-language-action models post-trained on the same task demonstrations. We collect no teleoperation data and use no reinforcement learning or on-robot corrections, and the robot appears only at test time. The system learns which regions of each wrist view matter for the task, predicts the stage the task has reached, and decodes actions as a continuous stream. A-0 is embodiment-agnostic: it runs on robots that mount the YUBI gripper, such as a bimanual Franka, TX-G2, OpenArm and UR. We are releasing the dataset, to our knowledge the largest open dataset of handheld-gripper demonstrations with task and subtask instructions, together with the pre-trained weights and the post-training recipe. Videos of A-0 on every task are on the project page, https://a-0.airoa.io. Anyone can build the gripper, record a task of their own and post-train the model on it in under four hours on a single node of eight H200 GPUs.
Multimodal language models continue to improve on broad music understanding benchmarks, but which musical features can they perceive and reason about? We investigate this question by introducing PianoBench, a benchmark that tasks language models with reasoning about musical performances across congruent acoustic and symbolic input representations. PianoBench is derived from our new multimodal, training-scale dataset comprising more than 20,000 hours of interleaved transcripts of music and spoken musical commentary from instructional piano videos. Open-weight audio-language models achieve at most 52.8% accuracy on PianoBench, and at best only modestly outperform their accuracy on a no-audio control. In contrast, frontier models using chain of thought achieve up to 93.2% accuracy when provided transcripts of music as text, exposing a gap between perceiving musical features from audio and reasoning about musical events when they are given explicitly. We then test whether mid-training on interleaved musical instruction can improve question answering without chain of thought, raising Qwen 3.5 9B Base's zero-shot accuracy on PianoBench with music transcripts from 42.8% to 56.0% after instruction tuning. These experiments suggest that instructional videos may represent an overlooked source of multimodal supervision for music. Available: https://gclef-cmu.org/pianobench.
Identifying faulty components from sensor anomalies is difficult because fault effects can propagate through a plant and different faults can produce similar observations. Detailed component models and labeled fault examples may also be unavailable. We introduce LaplaceDiag, a data-driven method that uses changes in actuator effects to narrow the search for faulty equipment. Actuators act on specific devices and mechanisms, so their changed effects can provide clues to root causes. LaplaceDiag learns a healthy plant model with LaplaceNet, a neural network that operates in Laplace space. It separates nonlinear prediction based on measured history from an explicit matrix that describes how actuator values affect sensor values. Keeping the history predictor fixed allows current actuator effects to be estimated by linear regression without retraining the neural network. Comparison with a healthy reference accounts for estimation uncertainty and normal variation to detect changes and generate actuator-related diagnostic hypotheses. The method supports online and offline diagnosis, providing an intermediate step between anomaly detection and identifying faulty components. We successfully evaluate LaplaceDiag on two datasets.
Retaining predictive gains can require persistent errors in a model's response to interventions. For deterministic predictors with finite second moments under squared-error loss, any observational advantage over the best predictor using only an outcome's direct causes requires dependence on other inputs, regardless of the learning rule. When admissible interventions vary inputs while preserving the worldly outcome, this dependence yields a positive lower bound on root-mean-square intervention-response error. We identify conditions under which population gradient updates introduce this dependence from a parent-only baseline or damage a previously correct mechanism through shared parameters while improving other predictions. These dependencies can also make model behavior vary with how the same intervention is realized. We characterize interventional alignment through a common output mapping across internal states and physical processes realizing the same intervention target and relevant context. Correction is limited by the causal hypotheses currently admitted within inquiry. The relevant hypothesis class comprises these causal alternatives, rather than the function class the computational system could in principle represent. For predictions restricted to these alternatives, interventional mean error cannot fall below the best approximation they permit, whether evidence is observational or interventional. Reducing error below a positive limit requires admitting better hypotheses, which may entail revising the causal variables, mechanism structure, or intervention interpretation. Interventional criticism constrains replacements without determining them. We conjecture that reliably acquiring and maintaining interventional competence when causal commitments fail requires autonomous participation in proposing alternatives, testing their consequences, and revising them in response to error.
We study the numerical constructions of almost complex and complex structures on spheres and products of spheres up to real dimension six. Our general set-up parametrises almost complex structures via a global matrix field in an ambient Euclidean space, compressed by the tangent projector. The projector's algebraic residual measures the almost complex condition, while its Nijenhuis tensor is evaluated in the intrinsic norm of the round product metric. Polynomial features and neural networks are trained without prescribed targets to perform general searches for these structures. The almost complex search passes numerical admissibility checks on all nine examples that admit such structures and rejects $S^4$, where a topological obstruction is known. A matrix correction then enforces the almost complex identity, subject to a novel uniform admissibility condition. For integrability, a free polynomial representation with Levenberg-Marquardt optimisation learns the complex structures on $S^2\times S^2$, $S^3\times S^1$, $S^3\times S^3$ and $S^5\times S^1$, with the largest Nijenhuis norms found by adversarial search all below $10^{-9}$. These general searches were unsuccessful on $S^6$, $S^4\times S^2$ and $S^4\times T^2$. Therefore, for these three cases direct numerical reconstructions were built in torus-fibred coordinates. They are integrable by construction, satisfy their compatibility conditions to better than $10^{-11}$, and corroborate the recent constructions on $S^6$ and $S^4\times S^2$.
Modern high-performance computing (HPC) clusters integrate multi-generation CPUs and GPUs, which generate rich, high-dimensional telemetry streams from out-of-band baseboard management controllers (BMC), job schedulers, and facility power distribution units. Detecting anomalies within this telemetry is critical to prevent energy waste and costly node failures. Existing approaches often rely primarily on statistical deviations and require stronger labels than actual production environments can provide. In addition, these detectors have a limited operational context for interpreting alerts. We present INSIGHT-HPC, a multi-plane anomaly-detection architecture that combines statistical, workload-aware, and physics-guided evidence into operator-interpretable episodes. INSIGHT-HPC integrates heterogeneous, large-scale telemetry streams and transforms noisy minute-resolution observations into de-duplicated fault episodes with operator-facing explanations. This design allows the system to distinguish workload-explained behavior from genuine infrastructure anomalies while producing interpretable alert summaries for operators. On production HPC telemetry with weak operational labels and controlled HPC-relevant fault injections, INSIGHT-HPC achieves a stronger precision-recall balance than all evaluated baselines. The system achieves a precision of 0.913, a recall of 0.817, an F1 score of 0.862, and an AUROC of 0.896. These results demonstrate the value of combining workload context, physical consistency, and explainable episode-level fusion.
Quantum reservoir computing (QRC) processes temporal information through quantum dynamics, but it remains unclear whether small QRC models can forecast real data competitively with classical models and whether their predictions survive the transfer from ideal simulation to noisy hardware. We address both questions for a six-qubit gate-based reservoir that forecasts next-day motor-insurance claim counts. Five reservoir candidates were selected and frozen using hardware-native interaction graphs, Trotterized Ising dynamics, memory and measurement-robustness criteria, circuit cost, and calibrated-noise screening. On 232 held-out test days, the ideal candidates reached root-mean-square errors (RMSEs) of 5.66-6.12 claims/day, compared with 5.36 for ridge regression. For four of the five, the difference from ridge was not significant under a 95% block bootstrap, and their accuracy was comparable with classical recurrent and reservoir baselines. On the 156-qubit IBM Kingston processor, we ran 63 evaluations in 21 paired rounds, comparing hardware-aware, random-native, and Qiskit optimization level 1 layouts. Hardware-aware placement gave lower RMSE than Qiskit in every round, with a maximum transfer distortion of 7.59 claims/day (11.84 for Qiskit), and avoided the severe transfer failures seen for some random layouts. Most of the hardware distortion appeared as a systematic forecast shift, which a short hardware probe on training inputs detected before the test runs. Longer replay windows raised circuit cost without a consistent gain in forecasting accuracy or hardware transfer. Small QRC models can thus reach competitive accuracy on this task, and hardware-aware layout selection makes their transfer to IBM Kingston more consistent.
Low-rank adapters (LoRA) let one base model serve many fine-tuned variants: multi-tenant engines keep a single copy of the base and batch the requests of many tenants, each with its own adapter. As base models move to 8-bit (FP8) and 4-bit formats, their general matrix multiplications (GEMMs) become cheaper, but the adapters do not: engines run them as separate small GPU kernels beside the GEMM, whose latency takes a growing share of every decoding step. An adapter projects its input down to a low rank and back up, and our key observation is that the GEMM needs the down-projection only in its epilogue, the final stage in which it writes its output. EpiLoRA, a plugin for the vLLM engine, therefore launches the down-projection first and the low-precision GEMM as a dependent kernel that streams its weights meanwhile and adds every tenant's up-projection in its epilogue. On NVIDIA Hopper GPUs, EpiLoRA decodes for a single adapter with up to 24% lower time per token than vLLM's fastest configuration, and faster than SGLang and TensorRT-LLM; with many tenants, its adapter kernels deliver more throughput than vLLM's on most GPU-model pairs we tested. The same low-precision base also trains: a LoRA reinforcement-learning trainer that computes on the engine's FP8 weights needs no 16-bit copy, so a 27B model trains on one 48 GB GPU, and an engine that then keeps activations in 16 bits removes most of the remaining mismatch between the model generating training data and the one trained.
Autonomous vehicles require robust perception systems to operate safely under diverse environmental conditions. While cameras provide high-resolution visual data, their performance significantly degrades in low-light and adverse weather. High Definition (HD) radar offers superior reliability and angular resolution under such conditions but lacks rich visual feature recognition. LiDAR, though expensive, delivers highly precise distance measurements and dense point clouds. Each sensor exhibits unique trade-offs, necessitating the development of advanced sensor fusion algorithms to harness complementary strengths while minimizing computational complexity and maintaining accuracy. In this study, the proposed fusion architecture performs early fusion by combining camera images with perspective-view point cloud data from HD radar and LiDAR sensors. The proposed model is trained to detect vehicles and segment free drivable space simultaneously. Experiments are performed on the RADIal dataset where we demonstrate that fusing point cloud data with camera images in front-view surpasses the performances of the most recent standalone and fusion models.
Generative AI tools for research are widely deployed, yet individual productivity gains do not guarantee better team research. We argue that how AI use is structured matters alongside access. Personal AI tools may amplify two failures under deadline pressure: a consumption bottleneck, in which generation outpaces team-visible evaluation, and AI-induced research silos, in which private human-AI loops displace shared decision-making. We motivate this position with a mechanism-generating pilot: a 100-minute research sprint with twelve researchers in three independently operating teams, within a seventeen-person session. Both patterns recurred across teams and evidence sources. Of 33 logged AI interactions, one was accepted outright and 27 lacked a recorded disposition; missing records alone do not establish absent evaluation. A parallel autonomous pipeline produced a complete draft without in-session human evaluation, illustrating production capacity rather than comparative quality. We propose Human-AI-Human collaboration: an interaction topology in which AI connects researchers' contexts and supports inspectable collective evaluation and decision-making. The pilot does not isolate AI-specific effects from time pressure or workflow practices. We close with seven modular, falsifiable design hypotheses, prioritizing shared context and team-visible evaluation checkpoints for controlled follow-up.
A substantial line of work builds rotation invariance directly into point cloud architectures. We benchmark 10 architectures (PointNet, PointNet++, DGCNN, RIConv, RIConv++, RISurConv, PointMLP, PointTransformer, and the equivariant VN-PointNet and VN-DGCNN) across 5 domains under pose, noise, and occlusion, and analyze the results with confirmatory hypothesis tests rather than point estimates. The three rotation-invariant (RI) architectures achieve the best pose robustness but rank among the lowest on noise robustness in 4 of 5 domains. We ask whether this cost is intrinsic to invariance or specific to how invariance is obtained. We retrain all 10 architectures unchanged, with a simple invariance loss on the embedding and three training seeds each, and evaluate all models on the independently designed ModelNet40-C. Invariance training transfers narrowly but reliably. It improves exactly the corruption families its perturbations target wherever the baseline has headroom (mean occlusion +10.5, lidar +13.4 percentage points of accuracy). At matched training budget the best standard backbone and the best RI architecture end on par on unseen corruptions (within 2.5 percentage points in every seed), and invariance-trained RIConv++ is the most seed-stable top model. Once rotation augmentation is widened, standard backbones concede only a flat, angle-independent accuracy offset under full-sphere rotation. We release the benchmark along with a package for analyzing point cloud robustness and reproducing our results.
Large language models (LLMs) read company news at scale, and backtests suggest that their readings predict stock returns. We ask how much of that predictability survives a test that rules out look-ahead, memorization and selection. We propose OLIVIA, a protocol for such tests: point-in-time inputs with company identities masked, selection on validation data only, an analysis plan frozen by cryptographic hash, and a test period unblinded once through a logged lock. We apply it to AI decision models, non-generative language models that return probabilities for typed questions, reading news about S&P 500 stocks: an open-weights model we fine-tuned on 2024-2025 data, three commercial models and standard baselines. The fine-tuned model's signal replicated on two validation quarters (daily information coefficients 0.055 and 0.065) but fell to 0.020 (t = 1.7) on the blinded January-July 2026 test. The commercial models' readings line up with the overnight move to the open (information coefficient 0.074 to 0.077) but not with the move after it, so the news is priced by the time a trader can act on it at the open. No model can tell over- from under-reaction, and no pre-registered strategy survives costs and a deflated Sharpe ratio. Commercial models identify masked companies in 62% to 74% of cases against a 28% baseline: masking hides names, not companies. The market reads the news first.
Long-context inference is bound by the KV cache, which 2-bit codes could shrink about sixfold if their error stays in check. The usual remedy rotates keys before quantizing, assuming rotation always helps. However, we observe that it does not: rotation flattens a spiky piece of a key but can give a flat one a new peak, and both coexist within one key. We therefore propose FlagRot, which stores each 32-channel piece rotated or as is, whichever has the smaller calibrated attention error, and rotates the query instead when reading. We further show that how often to decide follows the quantization axis: per-channel caches such as KIVI need one decision per head, at no write cost, while per-token caches decide per key, with the choice stored in existing sign bits. At 2.5 bits per value, FlagRot lowers the twelve-model KL divergence of KIVI from 18.7 to 12.4 x 10^-3, below RotateKV on 9 of 12 models. Served as a vLLM plugin, it raises RULER at 128k tokens by +2.9 points and the LongBench v1 mean by +0.4 to +1.9 points on the three of four models where 2-bit KIVI loses accuracy. It also lowers the error of 4-bit integer and FP4 caches.
LLM agents struggle to automate complex enterprise procedures because static, pre-built harness cannot capture proprietary policies, undocumented behaviors, or state-dependent recovery paths that surface only during execution. Procedural memory accrued from agent experience offers a training-free path to closing this gap -- yet unlike episodic records or semantic facts, it must consolidate repeated successes and diagnosed failures into reusable actions and recovery strategies. We introduce WASP-Mem (Wiki-based Agent-Synthesized Procedural Memory), a nimble but effective method that distills execution traces and feedback into persistent, interlinked wiki pages, enabling continual agent learning. On TheAgentCompany benchmark, WASP-Mem raises the checkpoint pass rate by 4.6-6.1% over the strongest baseline we compared across three models while requiring 57-61% less storage. We further present a systematic controlled study spanning consolidation method, feedback type, trace ingestion scope, and refinement rounds. Our results reveal that prescriptive knowledge organization such as human defined schemas and explicit deduplication rules hurts performance rather than closing the procedural-knowledge gap. In contrast, human diagnosis converts failed traces into corrective knowledge and drives sustained improvement: a single round of human diagnosis outperforms four rounds of self-judgment by 7.0-10.9%. These findings point to a simple division of labor: let agents organize procedural memory, and invest human effort where it matters most -- diagnosis.
Offline multi-task reinforcement learning (MTRL) aims to learn a unified policy from heterogeneous offline datasets that generalizes across tasks. However, existing offline MTRL methods often struggle with two fundamental issues: severe per-task data scarcity and limited cross-task generalization. In this paper, we propose Language-Model Augmented Decision Transformer (LaMA-DT), a unified sequence-modeling framework that leverages a single large language model (LLM) for both trajectory augmentation and multi-task policy learning. LaMA-DT uses natural-language task descriptions to guide task-aligned trajectory augmentation through a set of supervised generation and alignment objectives, and incorporates a scoring-and-filtering mechanism to retain high-quality augmented trajectories. The same LLM is then repurposed as a language-conditioned Decision Transformer and fine-tuned on a mixture of real and augmented data. Experiments on Meta-World and MuJoCo multi-task benchmarks demonstrate that LaMA-DT consistently outperforms competitive baselines, particularly in low-data regimes, showing that language-guided data augmentation substantially improves cross-task generalization.
Speculative sampling can accelerate decoding without changing the target distribution. Its standard guarantee assumes that logits processors, which enforce constraints or modify token scores, receive the appropriate token prefix and state at each verification position. We formalize this processor-prefix contract and present BADC (branch-aware differential conformance testing). BADC combines enumerated references on tiny random models, checks of processor inputs and state, and simultaneous confidence bounds on output-event probabilities. Preregistered, source-informed experiments on selected Transformers and llama-cpp-python releases confirm two causes: processors read unverified drafts or advance state without rollback. For one pilot-selected prompt on a trained checkpoint, fresh samples give a 60.26-percentage-point lower bound on the difference in the probability of a predefined first-token event between ordinary and prompt-lookup decoding, with nominal simultaneous 95% coverage under within-arm i.i.d. sampling. Controlled comparisons establish no cause-recovery advantage of directed selection over random selection, while permutation testing detects each tested family at least as often as BADC. These findings connect processor-interface failures to observable decoding differences without establishing diagnostic superiority or workload prevalence.
The uncertainty of a statistical model is most commonly factorized into an aleatoric and an epistemic part. This factorization changes how predictions are interpreted in downstream decision tasks. Importantly, except for the idealistic scenario with no model mismatch, the quantification is a characteristic of the model and not the data generating process. In this paper, we propose Forgetting to Improve a method that reduces this discrepancy by incorporating the task into the modeling framework. Our key insight is to acknowledge that in scenarios of model mismatch, data can have a detrimental effect on the modeling for a specific task. Based on this insight, we propose an influence function for Gaussian process models that allows for principled removal of detrimental data samples. We showcase the flexibility of this approach by demonstrating significant improvements across a range of tasks, including Bayesian optimization, model-based reinforcement learning, and transductive learning.
In speculative decoding, parallel drafters such as DFlash make drafting cheap, so target verification dominates each round. A central choice is the verification length, the number of tokens that the target processes in each round. Verifying more draft tokens can accept more of them, but work after the first rejection is wasted. For hybrid target models that combine Gated DeltaNet (GDN) and mixture-of-experts (MoE) layers, verification time grows much faster with this length than for dense models. In a single-request profile of Qwen3.6-35B-A3B with DFlash, it grew by 71% from length 4 to 16, almost entirely in GDN and MoE layers, compared with 4%-6% for a dense Qwen3-4B target. Each GDN layer must update a recurrent state at every token and recover it at the end of the accepted prefix. The MoE cost depends on the experts that the tokens activate, which are known only once verification runs. We present HybridSpec, which selects the verification length on the GPU from draft-side acceptance estimates and measured target latency. Verification then runs this length plus one extra token up to the first MoE router, before any expert runs. There, a lightweight routing correction uses the observed expert assignments to decide whether to keep that token. The rest of verification runs only on the selected tokens. Block-parallel GDN kernels avoid storing a state after every token and reconstruct the accepted state of every GDN layer in one launch. In isolated benchmarks at verification lengths 4 to 16 with every draft token accepted, these kernels, including accepted-state reconstruction, are $1.54\times$-$4.04\times$ faster than the default kernels. With three GDN-MoE target models in single-request greedy decoding, HybridSpec is $1.75\times$-$1.88\times$ faster than autoregressive decoding and $1.30\times$-$1.33\times$ faster than DFlash.
We ask how hard it is to build a neural network when its parameter values are given in advance. An instance supplies a multiset of rational numbers, a target behavior, and a bound on the number of neurons, output included. The question is whether one can choose the graph of a ReLU network within the bound, and place the numbers on its edges and biases so that it meets the target. This separates the numbers from the architecture; permutation training instead fixes the hidden units and only rearranges the output weights. Matching two labeled examples exactly is already $\mathsf{NP}$-complete; we ask what changes when the target must hold on every input or on average. Agreeing with a Boolean circuit on every Boolean input is $\Sigma_2^{\mathsf P}$-complete, already for four neurons. Agreeing with a given ReLU network on every real input lies in $\mathsf{DP}$ for every fixed neuron bound and is $\mathsf{DP}$-complete for three neurons. Approximating a ReLU network on the unit cube within absolute error one everywhere is $\Sigma_2^{\mathsf P}$-complete, even with no hidden neurons and all numbers equal and positive. Deciding whether some network has expected squared loss at most a threshold is $\mathsf{NP}^{\mathsf{PP}}$-complete for uniform Boolean inputs. For uniform inputs on the unit cube and for standard Gaussian inputs with a positive threshold, the same holds for two neurons when a positive risk gap, given in binary and possibly exponentially small, separates yes from no instances. Without a gap, the cube version is complete for guessing a polynomial-length string, then comparing a sum of exponentially many rationals with a threshold, all computable in polynomial time from the input and the guess. Since the graph is chosen too, the reductions must rule out every other network built from the same numbers.
Real-world problems span diverse domains, including program synthesis, symbolic reasoning, planning, and optimization, and often require fundamentally different solution paradigms. A central challenge for LLM-based reasoning is twofold: identifying an appropriate problem-solving approach and formulating it correctly. In this paper, we propose AlgoPilot, a cross-paradigm reasoning framework that enables adaptive selection of the solving approaches and structured problem formulation prior to execution. The AlgoPilot framework contains three components: 1) a SteerLM, trained via supervised fine-tuning (SFT) and reinforcement learning (RL), that adaptively selects the solving strategy (paradigm and modes) based on input problems, 2) the corresponding expert GuideLM, trained via SFT, that generates structured formulation guidance tailored to the selected strategy, and 3) an ExecutionLM that generates solutions conditioned on both the selected strategy and its formulation. Unlike prompting-based methods that rely on fixed reasoning templates, our approach learns to adaptively choose and structure problem-solving strategies based on the input. We build AlgoPilot on Qwen3-8B and evaluate across 5 problem categories and 13 benchmarks with 66 unique domains, comparing against both the latest similarly scaled small models and large LLMs. AlgoPilot significantly improves average accuracy for the same model from 38.4% to 72.9%, outperforming the best large LLM prompting baseline GLM-5 by 11.7%. Ablation studies confirm the effectiveness of each component. Notably, we show that the learned steering and formulation modules can transfer across execution models, suggesting their potential to improve other LLMs. Project page: https://sites.google.com/view/algopilot.
Vision-language models (VLMs) increasingly support construction hazard monitoring, yet their frequent false alarms remain poorly understood. This study investigates whether false struck-by hazard warnings arise from failures in object recognition, spatial relation encoding, or hazard decisions dominated by learned co-occurrence priors. We evaluated Qwen2.5-VL-7B and InternVL3.5-8B using 355 counterfactual worker-excavator scene pairs with controlled spatial proximity and worker-removal variants, combining behavioral analysis, internal representation probing, causal interventions, and decision correction. Both models recognized the entities and encoded their spatial relations (out-of-fold AUROC: 0.82 and 0.91), assigning higher hazard scores to near than far scenes in 79% and 73% of pairs. Nevertheless, 74% and 100% of far scenes were falsely classified as hazardous. Removing the worker reduced Qwen's hazard rate to 11%, while worker presence exerted 2.4- and 3.2-times greater influence on hazard scores than spatial proximity in the respective models. Attention-head interventions, neuron-level interventions, and activation steering failed to eliminate this co-occurrence bias. However, conditioning hazard decisions on the model's own spatial-relation representation improved Qwen's specificity from 0.36 to 0.58 at matched sensitivity, outperforming threshold recalibration. These findings reveal that VLM struck-by false alarms stem primarily from decision-making failures rather than deficiencies in object recognition or spatial encoding, suggesting that future efforts should prioritize conditioning hazard decisions on encoded spatial evidence rather than further improving visual recognition alone.
Speech watermarking is becoming a key defense for digital copyright and content authentication. Existing attacks treat evasion as a removal problem: they erase the watermark, relying on detector access or on speech data that is hard to collect, such as paired clean-watermarked or clean target speaker speech. Commercial platforms are moving to close off these assumptions. TrustedAudio, for instance, restricts detection to authenticated users and denies meaningful detection feedback to unauthorized parties. Adversaries thus cannot tell whether a sample is watermarked or which watermark it carries, nor reliably remove it without degrading speech quality. In this work, we show that evading a speech watermark does not require erasing it: detection relies on the structural consistency of the embedded pattern, and disrupting this consistency is enough to defeat it. Building on this observation, we identify harmonic structure as an effective attack surface, as perturbing it breaks watermark consistency while preserving perceptual quality. We therefore propose ScrambleMark, a zero-query, watermark-agnostic, black-box attack that evades detection through harmonic scrambling. ScrambleMark combines a deterministic engine for discovering transferable harmonic perturbations with a generative model that learns randomized scrambling patterns, avoiding reliance on fixed attack rules. Across three commercial platforms (Google SynthID, Resemble AI, and TrustedAudio), ScrambleMark is the only attack to both evade detection and keep speech natural, achieving nearly 97% evasion and a 4.2/5 human rating, far above the 1.1 to 1.7 of prior attacks. Beyond evasion, ScrambleMark withstands adaptive recovery designed to reverse it, evidence that structural evasion is hard to mitigate in practice.
Small students degrade on large reasoning models' (LRMs') long solutions, so recent work shortens them by interpolating an LRM with a short-answer model at ratio $\lambda$. We ask what student weights retain when this teacher stops thinking, comparing LoRA adapters of three students from three teacher pairs at each $\lambda$. In each pair, the adapters' update subspace shifts most where the teacher stops thinking and solutions shorten most, locating the change without attributing it. We identify a shared adapter subspace tied to thinking-like behavior (reflective markers, monologue format and long solutions). Shared by adapters before that point, it keeps about half (post hoc) while one pair's teacher halves its solutions but keeps its think tag, and largely empties after thinking stops. Removing it largely turns this behavior off for every pair, student and benchmark. Only in that pair and one other student does it leave clearly fewer markers than removing the update's top components; elsewhere, same-energy random directions often nearly match it. Grafted into a short-answer student at twice its weight, it turns early markers (a post hoc count) on, like the whole update, but the student rarely finishes. Our work identifies a weight subspace tied to distilled thinking-like behavior.
We study nonparametric estimation and forecasting for GARCH-type processes and propose an iterated neural estimator that alternates between estimating the recursion map and updating the unobserved conditional variances. Under mild conditions, we establish finite-sample bounds for the time-averaged in-sample variance-path error and show that the estimator is minimax rate-optimal up to logarithmic factors. Under stationarity and mixing conditions, we establish the same rate for recursion-map estimation and derive finite-sample bounds for out-of-sample variance forecasts and expected excess Gaussian quasi-likelihood (QLIKE) loss. The convergence rates depend on the true lag orders and remain unchanged when the fitted orders are overspecified. Simulation studies illustrate the effectiveness of our estimators under nonlinear and time-varying volatility dynamics. Real-data applications suggest that our neural methods perform competitively and can incorporate option-price and high-frequency information.
Long-horizon video generation is increasingly moving toward autoregressive rollout, where each newly generated segment becomes part of the evidence used for future prediction. Recent causal decoders, long-context training, and KV-cache mechanisms have extended feasible duration, but most inference pipelines still decide the visible past implicitly through recent windows, compressed caches, or external memory. This raises a direct question: \textbf{can the choice of which past frames or segments to show to the decoder be treated as an explicit test-time decision?} We answer this question with \textsc{ReCR}, a training-free method that formulates autoregressive video decoding as representation-guided context rewriting. At each step, \textsc{ReCR} builds a selected visible context from the generated history by using internal decoder representations to favor context units that match the global rollout state, avoid redundancy with the recent continuation, preserve temporal boundary coverage, and widen the range of spatial layouts in view. The selected units are then mapped onto a compact causal axis before decoding the next block. Across long-horizon text-to-video benchmarks, \textsc{ReCR} consistently improves multiple autoregressive backbones under matched visible-token budgets. Empirical analyses such as representation-space diagnostics and transition-count prompt switching further show that improving what remains visible is an effective path toward more stable long-video decoding.
We study the minimization of the average of $n$ convex $(L_0,L_1)$-smooth functions: the Hessian norm of each function is bounded by an affine function of its gradient norm. We assume known upper bounds $R$ on the distance from the starting point to a minimizer and $\delta$ on the difference between the minimum of the average and the average of the infima of the components. We derive upper and lower bounds on the oracle complexity of reaching accuracy $\varepsilon$ in function value in expectation. Up to logarithmic factors, both bounds are of order $n(1+(L_1R)^{2/3}) +\sqrt{\frac{n(L_0+L_1^2\delta)R^2}{\varepsilon}+n(L_1R)^2}$. The upper bound is attained by the proposed method YETI. The method combines ideas of accelerated proximal methods and variance reduction methods with a heuristic for choosing additional points. The number of oracle calls made by the heuristic is bounded, which preserves the upper bound on the total number of queries needed to reach the target accuracy. In the experiments, YETI is compared with first-order methods on synthetic problems and Poisson regression problems. We also examine the stopping criterion and the effect of the heuristic separately. The lower bound is obtained by an explicit construction of families of instances that are hard for any first-order method.
Deep-learning-based mass detectors usually downsample mammograms, trading fine image detail for lower computational cost. However, published studies cannot reveal how much accuracy this trade-off sacrifices, since they also differ in architecture, preprocessing, and evaluation protocol. In this work, we isolated the effect of input resolution in a controlled study on the VinDr-Mammo dataset, training 18 detectors at six input resolutions ranging from 512 x 416 to 1600 x 1280 pixels, for 108 model-resolution combinations. Whole mammograms were resized without cropping or tiling. For each model, preprocessing, model initialization, augmentation, and evaluation procedures were kept fixed across resolutions. Detection accuracy was evaluated by mAP50 on the 219 mass-positive images of the official test split, and 57 combinations with test mAP50 of at least 0.60 were analyzed. From the lowest to the highest resolution, median mAP50 rose from 0.626 to 0.660 and the best mAP50 from 0.670 to 0.709. The best result, achieved by DINO with a ConvNeXt-S backbone, is the highest mAP50 reported to date on mass-positive VinDr-Mammo images. Gains were concentrated in small masses: for the smallest third of lesions, median mAP50 rose by 0.074 (from 0.367 to 0.441), compared with 0.007 and 0.032 for medium and large masses. Median training time increased from 0.45 to 1.40 h, and median inference time from 23.6 to 113.7 ms per image. Therefore, higher input resolution mainly benefits the small masses that are hardest to detect, at a computational cost that remains practical.
Motivated by the expressive power of neural ordinary differential equations in flow-based generative modeling, we study the approximation of probability measures by pushforwards generated by parameterized families of Lipschitz vector fields. For Wasserstein distances, we establish a quantitative relationship between finite-sample point interpolation and distributional approximation, and derive universality criteria based on $L^p$-approximation of transport maps. In particular, affine-invariant families containing a non-affine vector field are universal from compactly supported absolutely continuous base measures. For total variation, we show that coordinate-wise dynamics are insufficient in dimensions $d\geq 2$ and establish sufficient conditions based on one-dimensional Sobolev approximation and shear-type mixing. We further examine the role of point interpolation, establishing when different regimes relate to one another and when approximate interpolation can be made exact.
Set separation by continuous linear functionals is fundamental in functional analysis and convex optimization and underlies binary classification in machine learning. Support vector machines provide maximum-margin affine separators for linearly separable data, while kernel methods represent nonlinear decision rules through linear separation in a reproducing-kernel Hilbert space. We study when two classes, not necessarily convex or linearly separable in the input space, can be separated with positive margin by finite ReLU networks and by random ReLU features. At the deterministic level, we prove that positive-margin separation by a finite ReLU network is equivalent to the existence of finite closed convex covers whose opposite-label members have positive pairwise distances, and exactly two hidden layers suffice. For random separation, suppose each class is represented as a finite union of nonempty closed convex cells, possibly unbounded. If every opposite-label cell pair has positive distance and its closed difference cone is pointed, then two independent random ReLU layers yield a common positive-margin affine separator with high probability. The first layer separates the feature image of each negative cell from the feature image of the positive class by a hyperplane through the origin. It also controls all cross-class distances uniformly. The second layer combines cellwise scores through Chebyshev polynomials and realizes the resulting polynomial using Gaussian Hermite features, followed by a uniform finite-width approximation. With the remaining geometric, bias, and confidence parameters fixed, each sufficient width is quadratic in the largest Gaussian width of a spherical difference-cone section, hence at most linear in its span dimension. An example shows that this squared Gaussian-width dependence is sharp for both layers.
Tokenization is a fundamental component of modern language models: before a model sees data, its tokenizer has already determined how the data appears. We identify and challenge three myths in the tokenization research literature. Myth #1: Bits-Per-Byte (BPB) is agnostic to the tokenizer and/or the vocabulary size: BPB is tokens per byte times bits per token, and the two factors do not cancel; using untrained models, we show that BPB prefers some tokenizers over others: the model has learned nothing, but the tokenizer--model pair has. Myth #2: lower BPB means a better model: On untrained models, BPB strongly prefers some vocabulary sizes, even though accuracy remains at chance, and on real models trained on the same data, BPB correlates weakly with downstream performance. Myth #3: different "tokenizer-agnostic" metrics yield consistent results: The "unigram-normalized loss" is not tokenizer-agnostic; its reported correlation of 0.989 with BPB comes from a confounder and falls to 0.057 once fixed; from the same data the two metrics select different vocabularies at each model size. We present a tokenizer trilemma: no metric can score only the true distribution best, give the same predictions the same score, and tie all untrained models, so choosing a metric is choosing a research question. We conclude with a caution for reading the tokenization literature and a conjecture for future work.
Under expert parallelism, mixture-of-experts (MoE) layers pay all-to-all token dispatch, grouped GEMMs, and a combine step in every layer. In decode each rank holds only 1-256 tokens: dispatch latency and weight traffic dominate while tensor-core utilization is inherently low. In fused operators such as DeepGEMM's Mega MoE, dispatch and computation overlap only by sharing every SM: dispatch warps compete with TMA/MMA and epilogue warps for registers, shared memory, and issue slots; every phase shares one CTA resource budget, synchronization spans the whole SM array, and every replay ends with a workspace-clean pass and a cross-rank barrier. We present GC-MoE, an open-source FP8/FP4 MoE inference operator for NVIDIA Hopper and Blackwell GPUs that uses CUDA Green Contexts to split the device into a small communication partition and a large compute partition. Each role gets its own kernel, resource budget, and synchronization domain, so dispatch and GEMM tune independently. A three-node CUDA Graph (four on Hopper with fused shared experts) launches the dispatch and grouped-GEMM kernels as concurrent roots; a token-ring handshake starts computation on the first filled ring blocks, and double-buffered receive counters remove the trailing cross-rank barrier between replays. The compute partition runs operand-swapped tcgen05 block-scaled MMA on Blackwell and register-A WGMMA with in-register FP4 decoding on Hopper, with fused SwiGLU, requantization, and optional shared experts. On 8 H20 GPUs GC-MoE accelerates DeepSeek-V4.1-Flash decode by 1.188x on average on both the routed and shared-expert suites (peaks 1.296x and 1.353x); on 8 B300 GPUs it accelerates GLM-5.2 and DeepSeek-V4.1-Flash by 1.106x (1.123x with fused shared experts, peak 1.209x), beating Mega MoE at all 54 measured points at matched accuracy. GC-MoE reserves only 20 of 148 SMs on B300 for dispatch.
Digital Twins are evolving from passive virtual representations into closed-loop systems that continuously synchronize with physical assets and drive real-time decisions. Yet predictive modeling, machine learning (ML), uncertainty quantification (UQ), and decision-making have been studied as separate capabilities, leaving limited understanding of their integration and maintenance during online operation. This narrative review, scoped by the functional requirements of real-time decision-making rather than by Digital Twin terminology alone, examines how such decisions are generated and sustained in engineering Digital Twins. We first examine foundational decision-making architectures and representative engineering applications, followed by methods that address their limitations across the decision loop. We identify optimization-based, policy-based, and knowledge-based approaches as the major pathways from virtual representation to action. Across them, ML plays three roles in decision-making: enabling efficient surrogate modeling, transforming heterogeneous physical feedback into decision-relevant information, and accelerating or directly generating decisions. UQ supports uncertainty-aware decisions and model adaptation, while continual validation, drift detection, and adaptation help maintain reliable operation as the physical system evolves. Together, these findings underscore the need for a shift from a model-centric to a system-centric perspective on Digital Twins, where its performance depends on the interactions with sensing, state estimation, computation, decision-making, and adaptation rather than on their standalone performance. Therefore, future research should shift from isolated algorithm development toward system-level co-design, scalable decision-making, and approaches to design, and data-efficient continual adaptation and validation for AI-enabled autonomous systems.
Data-driven algorithms are ubiquitous across industries, and there is an extensive literature on analyzing their \emph{average} performance. The risk profile of such algorithms, while critical, is much less understood. This paper quantifies the risk of data-driven policies and isolates its sources. We evaluate a policy through the conditional value-at-risk (CVaR) of its loss over the training sample, any policy randomization, and the future outcome, at a fixed unknown distribution. Regret against an oracle that knows the distribution decomposes into expected post-sample regret (PS), the average outcome-only CVaR gap of the selected action, and a sample-risk premium (SR), the additional tail risk omitted by averaging those evaluations. The analogous premium vanishes under expected loss. We characterize when the premium vanishes and how large it is relative to post-sample regret. For losses with a continuous density near the tail cutoff, we identify conditions under which both components decay at the same rate, yet the limiting ratio of the premium to post-sample regret can reach the thousands. When a probability mass at the cutoff moves with the action, the premium can decay more slowly and dominate post-sample regret. We establish conditions under which sample average approximation learns the oracle action at rate $n^{-1/2}$ in both settings, but its combined-risk regret is of order $n^{-1}$ in the first and $n^{-1/2}$ in the second. In finite-sample numerical examples of SAA in both settings, the premium accounts for approximately $74\%$--$97\%$ of regret.
We study online learning with directed weakly observable feedback graphs and propose a Follow-the-Perturbed-Leader (FTPL) policy that achieves best-of-both-worlds guarantee with nearly linear expected time complexity. To preserve the simple structure of FTPL, we develop a novel graph-based loss estimator that combines forced exploration, conditional geometric resampling and adaptive learning rates with surrogate probabilities. Our policy achieves $\tilde{\mathcal{O}}(\delta_*^{1/3} T^{2/3})$ regret for adversarial regimes as well as $\tilde{\mathcal{O}}(\delta_* \log T / \Delta_{\min}^2)$ regret for stochastic regimes, where $\delta_*$ and $\Delta$, respectively, denote the fractional weak domination number and the suboptimality gap. These results show that refined guarantees for problems with minimax regret of $\Theta(T^{2/3})$ can be obtained through appropriately designed adaptive learning rates and loss estimators without sacrificing the simplicity of FTPL.
In mission-critical networks, a missed jamming event is costlier than a false alarm, so detectors should favor recall, operate causally, and work without labeled attack data. Retrospective Key Performance Indicator (KPI) detectors meet the label-free requirement but rely on future context and cannot run in streaming operation. We present CORA, a teacher-guided reinforcement learning (RL) framework that transfers the hindsight of a retrospective detector to a causal one without labels during training or deployment. Offline, a retrospective KPI detector evaluated on unlabeled historical logs defines a recall-weighted reward for a masked controller that adapts the sensitivity, persistence, and grace of a causal KPI detector. At deployment, the teacher is removed and the frozen controller acts causally, with no ground-truth labels and no target-specific calibration. The design maps onto offline training in the Non-Real-Time RIC and causal control in the Near-Real-Time RIC of an Open Radio Access Network (O-RAN). Under leave-one-out transfer across ten 5G O-RAN testbed configurations, CORA reaches recall-oriented operating points on the deployable frontier that none of the evaluated baselines covers. A Proximal Policy Optimization (PPO) controller raises mean recall from 0.687 to 0.771 over a fixed default detector on all ten configurations, while a Deep Q-Network (DQN) achieves a comparable F2-score to the best teacher-selected static configuration, with 61% fewer false alarms. Removing the teacher lowers recall on every configuration.
Language models are increasingly used in text generation, decision support, and automated interaction, creating a need for localized, reversible, and selective control of their behavior. Activation steering demonstrates that internal interventions can alter the behavior of frozen models, but how to select and evaluate compact, multi-site control interfaces remains an open question. I introduce sparse internal control, a framework for steering target behavior by applying inference-time interventions to a small set of internal model nodes. A local target-controllability formulation characterizes candidate site--direction pairs by their behavioral effects and specifies constraints on target recovery, off-target preservation, and intervention energy. Two complementary selectors instantiate this formulation: Driver-OMP selects nodes for sparse reconstruction of a desired behavioral displacement, while Coverage-Driver favors stable coverage across prompts. I propose a six-axis control-evaluation protocol measuring reachability, signed reversal, dose response, feedback controllability, off-target preservation, and held-out reuse. Exploratory experiments on indirect object identification, MMLU multiple-choice question answering, sentiment steering, and refusal steering provide evidence for the feasibility of steering task-defined behavior through sparse internal interventions. The six-axis evaluation characterizes these interventions beyond target movement, revealing distinct control properties that depend on the task and available direction information. Together, the formulation, reference algorithms, and evaluation protocol provide a foundation for the systematic study and development of internal control methods for language models.
Radio Frequency (RF) jamming threatens fifth-generation (5G) networks. Open Radio Access Network (Open RAN) telemetry enables jamming detection without dedicated spectrum monitoring equipment, but labeled attack data are rarely available across deployments, motivating unsupervised detectors. Comparing such detectors across algorithmic families, however, remains difficult without ground-truth labels. This paper benchmarks statistical and reconstruction-based unsupervised jamming detectors using a symmetric evaluation interface that combines detection confidence and temporal interval coherence into a detector-independent composite score. We compare a Statistical Ensemble Method (SEM) detector against Vanilla, Long Short-Term Memory (LSTM), and Variational Autoencoder (VAE) baselines on Key Performance Indicator (KPI) telemetry from ten heterogeneous 5G environments with fixed and variable jamming bursts and dynamic jammer gains. Under leave-one-out (LOO) evaluation, SEM achieves a mean composite score of 0.906, compared with 0.688 for the best autoencoder (AE) per fold. Ground-truth validation yields a mean F1-score of 0.918 for SEM, compared with 0.521 for the best AE variant. The composite ranks SEM first in all environments, consistent with F1 rankings. However, in both within-trace and cross-trace evaluation, it fails for detectors that flag far fewer samples than the jamming prevalence, indicating that rankings are reliable only when detection rate matches the expected prevalence.
Many fields of physical interest on the sphere, such as wind, the polarisation of the cosmic microwave background and the shear of weak gravitational lensing, are spin fields, whose values change phase when the local frame rotates. We extend spherical Fourier neural operators (SFNOs) to fields of any spin, through lifted spherical convolutions that are exactly rotationally equivariant. Lifting spin fields to Mackey functions on the rotation group provides the conceptual framework, while computation is performed exactly and efficiently through their harmonic representation and by appealing to sampling theorems on the sphere and rotation group. We construct spin SFNOs (S$^2$FNOs) from these building blocks. We demonstrate empirically that rotational equivariance holds to machine precision for these layers. We achieve state-of-the-art performance on existing benchmarks, and introduce new benchmarks on which spin-aware models outperform their scalar counterparts.
Vision-Language-Action (VLA) models generalize across tasks, but their autoregressive decoders need sequential forward passes in proportion to the action horizon, each bounded below on embedded aerial platforms by streaming billions of parameters from memory. We present Aero-VPA, a lightweight Vision-Predictor-Action architecture for 3D navigation on edge hardware. A compact RGB-D Vision Transformer and a five-head ensemble predictor, trained without pixel reconstruction, forecast the next latent state; a Conditional Flow Matching solver generates a velocity and yaw-rate chunk in two Euler steps, so each decision takes five sequential network evaluations; and the ensemble dispersion drives an uncertainty fallback. Version 1 specified the architecture, deployment and protocol without results. This version reports the first closed-loop evaluation: PX4 software-in-the-loop simulation of quadrotor missions in procedurally generated forests. Behaviour cloning failed on all 60 missions despite open-loop errors of 0.02-0.04 m/s: the policy copies its last command, a copycat form of causal confusion. DAgger with a privileged, safety-gated supervisor gave the first successes but plateaued at 11.1-21.1% across six checkpoints. A map-free runtime shield with a 15 s depth memory, barrier constraints, a cap on motion outside the 73-degree field of view and backoff along the flown path raised the first DAgger policy's success on 180 paired missions from 21.1% to 52.2% (95% CI [45.0, 59.4]; exact McNemar p = 1.2 x 10^-10) and cut its crashes from 72.2% to 22.2%, while timeouts rose from 2.8% to 23.3%. All six checkpoints gained 30.0-34.4 percentage points, at mean speeds of 1.16-1.36 m/s. The code (https://github.com/ylee-sling/Aero-VPA), the trained policy (72.6 MB, https://huggingface.co/ylee-sling/aero-vpa) and the full benchmark dataset are public.
LLM agents increasingly act on information retrieved from public sources they treat as credible, which turns a manipulated recommendation into a transaction, a dependency, or a decision with physical consequences. The security literature models this threat as instruction injection, in which an adversary writes a command into retrieved content and deployed defenses learn to recognize command syntax. We show that an attacker can instead steer an agent through the agent's own harm-avoidance. A single fabricated hazard, written as an ordinary review, leads the agent to disqualify the legitimate options and select the attacker's target by its own reasoning, with no command for a detector to recognize. We call this voluntary servitude. These payloads steer agents about as often as instruction injection (60.4% vs. 66.0%), and deployed injection detectors score them as benign, so behind a filter they are 1.7-4.4x more effective. We measure this at the generation stage across six models and ten consumer-service domains, for three attack families, and show that it transfers to production browser agents, outperforming instruction injection, and to a retrieval pipeline (1.0-57.0% end-to-end success, depending on the retriever). We then show that a corroboration requirement added to the agent's instructions removes roughly 95% of the exposure against a non-adaptive attacker without retraining. With the requirement in place, the agent still reads the planted hazard and no longer acts on it.
Visualizing multivariate scientific simulation data is essential for understanding complex physical phenomena, yet the massive scale of such data imposes a severe storage bottleneck in high-performance computing environments. Implicit neural representations (INRs) have emerged as a promising continuous alternative to traditional lossy compressors, but existing methods struggle to explicitly capture cross-variable interactions and to adapt to the diverse spectral characteristics of physical variables. To address these limitations, we propose SWAP-INR, an architecture for high-fidelity compression of multivariate, time-varying simulation data. SWAP-INR introduces two key components: (1) The switchable attention path (SWAP) module explicitly models cross-variable interactions through a cross-variable attention layer, then routes each variable to an attention or non-attention path by a deterministic, loss-driven criterion, and prunes the inactive layers for structural compression. (2) The adaptability amplification SIREN (A2-SIREN) replaces rigid frequency scaling with channel-wise learnable modulation, adapting to heterogeneous physical features with negligible overhead. Combined with post-training 16-bit quantization, SWAP-INR achieves state-of-the-art performance on small modular reactor (SMR) datasets and remains competitive on public physics simulation datasets, while maintaining a constrained memory footprint. Our code is https://github.com/hvcl/SWAP-INR.
The integration of unmanned aerial vehicles (UAVs), mobile edge computing (MEC) and non-orthogonal multiple access (NOMA) provides an effective solution for computation-intensive and latency-sensitive services in infrastructure-deficient environments. However, the joint optimization of UAV trajectories, task offloading, communication and computing resources under heterogeneous task priorities results in a highly coupled mixed-integer non-convex problem. To address these challenges, this paper proposes a Hierarchical Intelligent Evolutionary Algorithm (HIEA) for priority-aware multi-UAV cooperative NOMA-MEC networks, which integrates multi-head deep Q-network (MH-DQN)-based intelligent decision generation, communication-aware initialization, adaptive evolutionary exploration and successive convex approximation (SCA)-based precision refinement. Specifically, a novel communication-aware initialization mechanism, termed Channel Quality Indicator (CQI)-guided Load-Aware Trajectory Initialization (CQI-LATI), is designed to incorporate CQI and task workload characteristics into evolutionary population construction, generating high-quality and diversity-preserving initial populations. Theoretical analysis characterizes HIEA's search-space reduction, the quality-diversity trade-off of CQI-guided initialization, and the accuracy-efficiency trade-off of SCA refinement, and extensive simulations demonstrate that HIEA outperforms benchmark schemes, with gains of up to 6.5% in throughput, 13.6% in task completion rate and 15.3% in energy efficiency.
Audio-based piano transcription is hampered by the sustain pedal and by noisy or missing soundtracks, problems that video-only transcription avoids, since a top-view video shows directly when each key is pressed and released. Most visual transcription networks, however, describe the whole keyboard with one global feature vector and must learn where each key is. In this paper, we propose KeyWise, a lightweight per-key architecture that exploits the known geometry of the keyboard instead of learning it. The rectified keyboard is divided into 88 equal tiles that serve as regions of interest (ROIs), and one small recurrent network, shared by all keys, follows each key over time. On PianoVAM, KeyWise reaches the onset accuracy of the state-of-the-art V2N (95.0% vs. 94.7% onset F1 at 50 ms and 97.8% vs. 97.0% at 100 ms) with 1.3 M instead of 28.2 M parameters, and it runs 4.7 times faster on the same GPU. At 100 ms, it is also significantly more accurate when the key releases (97.0% vs. 95.5% F1) or the velocities (89.4% vs. 84.5% F1) must be correct as well. On PianoYT, KeyWise pre-trained on PianoVAM reaches 69.4% onset F1 at 100 ms, comparable to the best published result, although it needs fine-tuning for new recording conditions.
Vision-language models such as CLIP enable zero-shot classification from arbitrary class names, but under test-time corruptions, their visual encoder projects distorted features onto incorrect textual anchors with high confidence: a phenomenon we term modality bias. Standard entropy minimization reinforces these overconfident errors, which creates a confirmation-bias loop and collapses the predicted label distribution. We propose an online test-time adaptation (TTA) method that breaks this loop at three levels. At the representation level, a frozen self-supervised backbone (e.g., DINO v2) is combined with CLIP through a cross-covariance logit ensemble. CLIP text prototypes are projected into the feature space of this text-free backbone, so the self-supervised branch classifies native image features and is less affected by the text bias of CLIP. At the optimization level, a gated test-time information-maximization objective pairs per-sample sharpening with batch-level diversity, while an adaptive gate excludes unreliable samples from the gradient. At the prediction level, an online logit adjustment corrects the residual class-distribution skew. CDIM outperforms the compared CLIP-based online TTA baselines on ImageNet-C, CIFAR10-C, and CIFAR100-C, and remains competitive in continual and mixed-domain TTA. Extensive analyses and ablation studies verify the contribution of each of its components. Our code is available at https://github.com/B1ueC1oud/CDIM
Predicting how the world evolves in response to robot actions is central to policy learning, evaluation, and planning. However, many action-conditioned world models predict future observations in 2D image or video space without explicitly modeling the underlying 3D scene, making spatial consistency across viewpoints and over long-horizon interactions challenging. We introduce 3D World Simulator, an action-conditioned multimodal world modeling framework that jointly predicts appearance, geometry, and tactile observations from a unified 3D representation. Our approach combines a hybrid particle-voxel autoencoder that compresses point clouds into compact 3D voxel latents with an action-conditioned latent dynamics model that predicts their evolution. The shared 3D representation supports geometrically consistent rendering from arbitrary viewpoints and flexible integration of spatially aligned tactile signals. Our experiments demonstrate interactive, long-horizon, multimodal prediction at 50 Hz. We then show how our model enables scalable data generation for policy training and evaluation using tactile and multi-view visual observations, including wrist-mounted and novel camera views. Furthermore, we scale our method to large datasets containing diverse objects and demonstrate model-based planning with unseen objects and robot configurations. Together, these results highlight the potential of hybrid particle-voxel dynamics as a foundation for multimodal 3D world simulation.
Electroencephalography (EEG)-based visual decoding and reconstruction aims to recover visual information from non-invasive brain signals. But there are challenges due to the complex temporal-frequency characteristics of EEG signals and the difficulty of discrete denoising framework in mapping neural activity to meaningful visual representations. Here, we propose E2I-Flow, an EEG-to-image decoding and reconstruction framework that combines a novel EEG encoder for time-frequency representation learning from EEG features, and conditional flow matching for visual decoding. A stationary-wavelet frequency branch and a temporal-mixing branch are integrated through bidirectional cross-attention and adaptive fusion, with the resulting EEG representations aligned to ground-truth image embeddings using contrastive and mean-squared-error loss supervisions. The aligned EEG representation is then used as a condition for a Transformer-based flow-matching model to predict the corresponding visual embedding for image synthesis. Experimental results show strong zero-shot EEG-to-image retrieval and improved visual reconstruction performance, outperforming several state-of-the-art baselines.
Radar-based human sensing typically relies on end-to-end task-specific models trained on hand-chosen signal projections, which generalize poorly across unseen users, tasks, and hardware. We argue that the dominant signal processing choice, the micro-Doppler spectrogram, is a poor starting point for transfer because it keeps the Doppler axis, which carrier frequency rescales, and discards the angular axis, which carrier frequency leaves unchanged. We instead pretrain a convolutional beta-VAE without labels on clutter-removed raw radar measurements from four platforms spanning 60 and 77 GHz and 4 to 12 virtual antennas, aligned by a unified spectral interpolation grid. For every downstream benchmark the encoder stays frozen, and only a small supervised temporal classifier is trained on its embeddings. We compare against supervised models trained end to end on each benchmark and a pretrained radar foundation model across ten settings covering activity, gesture, sign-language, and gait recognition. When the test users and the radar are seen during training, our encoder is best on all five benchmarks. For users never seen in training, it leads on three of four benchmarks, by 4.8 to 23.2 points. On subject-disjoint gait recognition, it is best for each of the four held-out subjects and 8.3 points above the strongest baseline on average, even though its pretraining excluded every recording of this radar. Excluding the test radar from pretraining costs at most 5.3 points on any benchmark. Code is available at https://github.com/jonasmueler/UniversalRadarEncoder_WACV.
Discovering predictive alpha factors from noisy financial data is a fundamental yet challenging problem in quantitative finance. LLM-driven formulaic alpha mining offers an attractive paradigm by combining interpretable symbolic expressions with iterative generation and refinement guided by quantitative feedback. However, existing generation-centric and tree-based search strategies primarily organize factors through generation lineage, leaving structural and statistical relationships across search branches underutilized. Consequently, information remains localized along parent--child trajectories, which can confine exploration to redundant regions. Moreover, effective factor discovery must account for multiple distinct objectives that cannot be adequately captured by a single search criterion. We introduce CosmosAlpha, a graph-based LLM factor-mining framework that represents discovered factors as a multi-relational graph encoding generation lineage, expression-structure similarity, and cross-sectional exposure correlation. A Graph Flow mechanism propagates search utilities across the graph, allowing each factor to be assessed using relational evidence from nearby factors rather than in isolation. Pareto-guided selection jointly considers quality, fertility and diversity, explicitly modeling their trade-offs within a multi-objective search framework. Across CSI 300 and CSI 500, CosmosAlpha-Graph achieves the best mean rank over seven predictive and portfolio metrics and outperforms its controlled CosmosAlpha-MCTS counterpart on every reported metric, showing a consistent advantage for relational graph search in LLM-driven factor mining. The source code is available at https://github.com/wzhmtldswz/CosmosAlpha
Implicit neural representations (INRs) model signals as continuous functions from coordinates to values, but their performance depends strongly on how spectral capacity is allocated. Yet most Quantum INR (QINR) architectures treat spectral components as independent outputs of separate quantum blocks. We introduce the Spectral Quantum Memory Model (SQMM), a memory coupled quantum INR in which sequential hybrid data reuploading layers (bands) communicate through a shared coherent quantum memory register. Each band reads the memory, computes on its own coordinate upload, and writes part of its state back, so later bands can build on the computation of earlier ones. When memory read and write layers are input independent, we prove that the memory state forms an operator valued Fourier series whose support is contained in the Minkowski sum of the initial memory spectrum and the local band spectra. Downstream bands then condition their Fourier coefficients on the spectral content stored by earlier bands, and the memory coupling creates no new primitive frequencies. This theory motivates a spectral memory alignment loss that encourages memory to encode the spectrum expressed by the deployed prefix of the model. Across audio and image reconstruction benchmarks, SQMM improves over compact QINR and INR baselines.
Ranking models define probability distributions over rankings and are fundamental in applications such as preference learning and decision-making. Without structural assumptions, these models are intractable due to their factorial complexity. Although independence assumptions are key to reducing complexity and enabling interpretability, standard notions of conditional independence must account for the mutual exclusivity constraints inherent to rankings. We introduce relative rank models, a probabilistic framework based on structured coarsenings of the ranking space that relax these constraints while preserving relative order information. This representation enables a well-defined notion of conditional independence for rankings and allows rankings to be modeled using Bayesian networks. As a result, standard methods for learning and probabilistic inference can be applied to rankings, with interpretable representations based on conditional independence. We further show that two prominent ranking models based on independence assumptions tailored to rankings arise as special cases, providing a unifying framework for independence in ranking models.
We propose a new estimator of low-frequency skewness (monthly) that exploits high-frequency data (5 min) through a direct functional mapping consisting of layers of Convolutional Neural Networks followed by Multi Layer Perceptrons. Monte Carlo experiments involving stochastic volatility models, with and without jumps, show that the estimator reduces finite-sample bias relative to existing realized-skewness estimators and remains accurate under model misspecification. Empirically, our estimator exhibits temporal stability and delivers cross-sectional performance in actual skewness-sorted portfolios. We characterize the finite-sample bias of conventional plug-in skewness estimators and establish risk consistency properties of the proposed NN estimator under standard regularity conditions.
Distributed learning has drawn considerable attention due to its capability of learning from distributed data without aggregating raw observations. However, frequent exchange of model updates can become the main bottleneck of communication. Therefore, one-shot approaches where each worker fits a local model and sends it to the server to aggregate emerge. It is known that averaging can achieve the so-called pooled-data mean squared error (MSE) order of a hypothetical node that has access to data from the server and all the workers, and performs empirical risk minimization (ERM) on it. However, when the number of workers increases too fast compared to the local sample size, nonzero local estimation bias prevents averaging from achieving this order. In this paper, we study whether independent server data can overcome this limitation. We consider $m$ workers with $n$ observations each and a server with $p$ observations, all i.i.d., giving $N=mn+p$ observations in total. Under certain regularity conditions, a pooled-data ERM achieves MSE of order $\mathcal O(N^{-1})$. We analyze deterministic nonnegative averaging of convex-loss M-estimators, with equal worker weights and the server's own estimator included. Under our assumptions, including nonzero local bias of order $n^{-\alpha}$, $\alpha\geq1$, sample-size weighted averaging achieves the pooled-data MSE order whenever $m=\mathcal O(n^{2\alpha-1})$. Above this worker scale, it does so if and only if $p=\Omega(nm\cdot m/n^{2\alpha-1})$. Hence, server-only estimation, whose MSE is of order $\mathcal{O}(p^{-1})$, can achieve pooled-data order while sample-size weighting fails. Analytical examples and synthetic experiments illustrate when averaging matches the pooled-data MSE order, complemented by CIFAR-10 experiments comparing aggregation strategies with linear and nonlinear classifiers.
Multimodal perception is central to many computer vision applications, including robotics and autonomous driving, where different sensing modalities provide complementary cues. Self-supervised learning offers a scalable solution for exploiting large amounts of unlabeled multimodal data. However, existing methods either neglect modality-specific information by enforcing feature-space alignment between modalities or learn only global representations without preserving spatial structure. We propose DenseTwins, a multimodal self-supervised framework that learns dense representations by explicitly decoupling the dense feature space into shared and modality-specific components. The proposed approach combines dense cross-modal correspondences with redundancy-reduction objectives, enabling the model to capture both shared semantics and modality-specific cues within a structured dense embedding space. DenseTwins jointly optimizes an intra-modal objective across augmented views and a cross-modal objective that aligns shared features while decorrelating modality-specific information. We evaluate DenseTwins across multiple multimodal benchmarks and sensing setups, achieving state-of-the-art or competitive performance across multimodal self-supervised baselines. Code available at: https://github.com/CarloSgaravatti/DenseTwins.
Small multirotor drones produce a characteristic acoustic signature, and microphones offer a low-cost way to detect them. High accuracies have been reported on public drone audio benchmarks, often with random clip-level splits of datasets built from a few long recordings and pieces of existing sound collections, so pieces of one source can appear in both training and test data. We study the binary task of the public DroneAudioDataset (11,704 one-second clips). Its 1,332 drone clips come from 257 five-second chunks of 7 recordings, and 96.4% of its non-drone clips are pieces of the 2,000 clips of the ESC-50 sound dataset. A grouped, source-disjoint protocol (five folds, three repeats) is compared with a random split. Each clip is described by 57 hand-crafted acoustic descriptors, reduced by a hybrid minimum-redundancy maximum-relevance and ReliefF selection or by NSGA-II multi-objective search, and by a 1024-dimensional embedding from the pretrained YAMNet network. Random Forest and XGBoost learners on both views form a stacking ensemble. Under the grouped protocol, the two-view stack reached a mean accuracy of 95.29% (SD 1.62 points over 15 test folds), a specificity of 98.29% and an area under the ROC curve of 0.990, and it correctly classified 98.78% of unseen ESC-50 source clips as non-drone. The random split overstated accuracy by 1.89 to 3.22 points for the four models. Adding YAMNet embeddings to the hand-crafted view raised accuracy by 1.11 points (higher in 12 of 15 folds). Hybrid selection matched the accuracy of all 44 baseline descriptors with 16 to 18 features, and NSGA-II retained 4 to 6 features at lower accuracy. The stack took 51.0 ms per clip on one CPU thread. Source-disjoint evaluation is needed for a fair benchmark on this dataset, and fusing hand-crafted descriptors with pretrained embeddings gives a small but consistent gain at modest computational cost.
In single-pixel (ghost) imaging, known illumination patterns probe an object and a detector with no spatial resolution returns one number per pattern. Each measurement costs acquisition time and photons and, for ionizing probes, contributes to delivered dose. Motivated by automated radiographic defect screening, where a class label suffices, we introduce Learning Illumination Patterns for Single-Pixel Classification (LIPS-PC). The patterns are learned end-to-end while the measurement operator remains one fixed linear matrix, identical for every input. We parameterize it as a residual correction to a two-dimensional discrete cosine transform (2D-DCT), $D=D_{\mathrm{base}}+\Delta$, and jointly train the correction and a compact classifier using cross-entropy, without a reconstruction objective. This yields a task-driven linear embedding through optimization. We evaluate LIPS-PC numerically on MNIST and Fashion-MNIST under a controlled protocol, then on RIAWELC, a real industrial X-ray weld-radiograph dataset. Across five budgets, $M\in{8,16,32,64,128}$, corresponding to sampling ratios from below 1% to 12.5%, the learned operator outperforms the best tested fixed transform by 9-15 percentage points at $M=8$ on all three datasets. Its eight-measurement accuracy requires roughly 16-32 measurements with the best fixed transform, indicating a $2\times$ to $4\times$ saving at comparable accuracy. The advantage generally shrinks as $M$ increases, with a small reversal on RIAWELC at higher budgets. These results show that learning what to measure is especially valuable when measurements are scarce.
Expert offloading enables large mixture-of-experts (MoE) language models to run locally, but conventional batching stalls the GPU while missing experts are fetched. A popular remedy, known as prefetching, is to predict what experts will be needed soon and fetch them using idle link bandwidth. Imprecise prediction, however, wastes bandwidth and GPU memory on fetching unnecessary experts. Our key finding is that prefetching can be harmful when there is an expert cache in GPU memory. This is because as the cache includes more necessary experts, expert prediction becomes less precise. Our key insight is that proper request scheduling beats prefetching. Building on this insight, we present Bevy, an MoE inference system based on runahead inference for concurrent requests. A request whose experts are available can run ahead to later layers while other requests wait for fetches, revealing the experts it needs next. Bevy can then fetch the missing experts so identified while computing ready requests, overlapping computation and fetching without predicting expert selections. Running requests separately, however, repeats reads of shared weights in GPU memory and increases compute time. To limit this cost, Bevy's self-balancing scheduler batches ready requests at each layer and prioritizes lagging requests, adapting their grouping as expert availability changes. Across two local platforms, Bevy delivers up to 3.46$\times$ the decode throughput of llama.cpp and up to 1.55$\times$ that of the fastest prefetching baseline.
Post-training updates change what large language models (LLMs) can answer, but may leave their answerability judgments misaligned. We study this knowledge-judgment desynchronization through functional feeling of knowing (FoK), which supports selective generation, retrieval triggering, and model routing without implying subjective experience. We propose two trainable channels: a knowledge channel that changes answerability and a FoK channel that judges the current knowledge state. Three stages train them: on-policy supervised fine-tuning (OP-SFT) updates knowledge from corrected on-policy answers; knowledge-delta FoK training (K$\Delta$-FT) adapts judgments across intermediate knowledge states; and FoK-guided policy optimization (FGPO) trains the subsequent answer or explanation. Factual-QA experiments, cross-dataset evaluation, and three sequential updates show that explicit FoK alignment improves judgment quality on updated knowledge. MMLU and ARC-Easy separately assess capability preservation.
Reinforcement learning (RL) post-training increasingly targets low-bit deployment, with NVFP4 weight-and activation quantization (W4A4) becoming the state-of-the-practice format. However, integrating NVFP4 into RL degrades rollout quality and amplifies cross-engine training-inference mismatch (TIM), destabilizing training. We show that these failures are concentrated rather than uniform - severely mismatched tail tokens concentrate in a few outlier sequences, while policy weights remain largely intact and quantization error concentrates in a few high-variance activation channels. Guided by this, we present FourSight, an algorithm-system co-design with two targeted solutions: TIM-aware sequence filtering that excludes drifted trajectories from policy updates, and sparse precision compensation that keeps only top-variance channels in BF16 while retaining native NVFP4 elsewhere. We further design a fused SpliceGEMM kernel with specialized prefill and decode datapaths to minimize overhead. FourSight sustains near-BF16 accuracy and strong FP4 deployment quality across models and engines, while improving step-to-accuracy by 1.76-2.60x and staying within 2.3% of FP4 step time up to 120B-scale models.
Reliable multi-sensor perception in autonomous vehicles requires accurate calibration between cameras and LiDAR. However, calibration drift is often unavoidable in long-term deployments: vibrations, thermal effects, and minor mechanical shocks may alter the relative pose between sensors, leading to systematic errors in downstream fusion tasks. We propose MiCaDe (Miscalibration Change Detection), an online solution detecting when the extrinsic LiDAR-camera calibration parameters are no longer accurate by means of sequential change detection. MiCaDe requires no calibration targets, or modification of the perception stack, providing low false-alarm rates with negligible computational overhead. Our key insight is that 3D and 2D bounding boxes produced by standard object detectors should remain geometrically consistent under correct calibration. By matching cross-modal detections, we derive geometric consistency metrics that serve as sensitive indicators of extrinsic drift. These metrics are aggregated into fixed-size batches and scored against a reference distribution, producing a likelihood stream that is monitored by a Change Point Model (CPM). Crucially, the CPM accumulates statistical evidence over time, enabling detection of subtle drifts that would otherwise go unnoticed. Experiments on multiple long-sequence datasets and on a real-world dataset with physically induced miscalibration confirm reliable detection under diverse conditions. Code available at: https://github.com/CarloSgaravatti/MiCaDe.
Saliency maps are offered as a safeguard for medical image classifiers, but whether they reveal individual errors at a new hospital remains largely untested. We trained a convolutional network and a vision transformer on NIH ChestX-ray14 and evaluated 303,768 decisions per model at the training hospital and at two external sites, in the USA and Vietnam, with radiologist-consensus labels. We scored each explanation by its anatomical enrichment, the area-normalized concentration of attribution in a segmented thoracic region, and tested whether it detects errors beyond the model's confidence. Grad-CAM localized findings less often than radiologists (hit rate at most 0.575 versus 0.888). Adding enrichment to a confidence-based error detector calibrated at the training hospital never raised its area under the curve by more than 0.011, and for the convolutional network the relation between enrichment and correctness reversed between hospitals, degrading the transferred detector. Model confidence, not saliency, should anchor deployment monitoring.
Port-Hamiltonian models describe physical systems by separating stored energy, dissipation and external inputs. Learning such models from data requires a choice of representation and fitting method. We compare using neural networks, analytic expressions and combinations of the two for representing the different components of the port-Hamiltonian system, and compare fitting the models with gradient-based optimisation or sparse regression. Testing on simulations of mechanical and electrical systems, with and without added noise, we present prediction accuracy and computational cost of the different approaches.
Generative models of protein conformational ensembles inherit much of their design from structure prediction for folded proteins. Ordered and disordered proteins, however, occupy distinct structural regimes, and we investigate how design choices hold up across them. Using a minimal C$\alpha$ flow-matching model, we ablate choices of architecture, sequence conditioning, rigid-body alignment and training source, and evaluate each variant against experimental SAXS and NMR chemical-shift data partitioned into ordered, disordered, multi-domain and transitional proteins. Mixture-of-experts routing, a static pair representation and a disorder probe have no detectable effect, a smaller protein language model performs worse on ordered proteins but not on disordered ones, and alignment in the loss degrades agreement with SAXS. Replacing coarse-grained simulations of disordered proteins with fragment-assembled ensembles improves agreement with chemical shifts at a cost to SAXS, while an equal mixture of the two keeps most of the gain at a smaller cost. This points to the training data, rather than the architecture, as the main constraint.
Choosing a variational quantum circuit (VQC) usually means searching over ansatze and depths, and training each candidate. We give three rules that fix these choices from the data and the gate set alone, before any training. First, a circuit without entangling gates has a rank-one Fourier coefficient tensor on every subset of qubits. A singular value decomposition of a dataset's Fourier coefficients therefore indicates whether entanglement is needed, which qubits it connects, and a lower bound on the error of omitting it. Second, the rank of the quantum Fisher information matrix grows by $p_0$ per layer, after a first-layer deficit $d_0$, until it reaches a ceiling set by the dynamical Lie algebra. This law fixes the depth. Third, reading out all $2^N - 1$ Pauli-Z products from one batch of shots resolves half of the state's directions, the most any single-basis readout can. A single-qubit readout resolves an exponentially small fraction. For the Madelon dataset, the rules prescribe 7 blocks of 3 qubits at depth 3: a 239-parameter classifier with 78.5% test accuracy, found without a grid search. Ablations confirm three predicted effects: shallower circuits, smaller blocks and dropping reuploading all lower accuracy.
Reinforcement learning (RL) is promising for adapting robot policies to unstructured real-world environments. However, standard RL pipelines rely on episodic training with frequent scene resets, which is impractical for real-world deployment due to the need for substantial human intervention. While prior work has studied non-episodic RL in relatively small policies trained from scratch, its behavior when fine-tuning modern large-scale pre-trained vision-language-action (VLA) models remains underexplored. In this work, we introduce CRONOS, a simulation benchmark for RL fine-tuning of VLAs under non-episodic scene resets. CRONOS leverages high-fidelity physics simulators and adopts shared-scene multi-task settings. Using CRONOS, we show that naively fine-tuning pre-trained policies fails in non-episodic settings due to shifts in object configurations and end-effector poses; however, these challenges can be mitigated through targeted reset allocation and by addressing biases in pre-trained models. Moreover, we find that non-episodic training enhances long-horizon manipulation and improves generalization to previously unseen object configurations and task sequences. Project page: https://embodiedai-ntu.github.io/cronos/
Digital twins (DTs) require timely updates from Internet of Things (IoT) devices to accurately represent the state of the corresponding physical systems, but transmitting large data volumes can compromise freshness. Moreover, different task classes have distinct computational demands and freshness requirements, while stochastic task generation and sizes complicate resource allocation. In this paper, we propose a freshness-aware framework in which multi-access edge computing (MEC) servers preprocess IoT data, reducing transmission volume and DT-server processing load. The framework establishes mutual support between DT and MEC servers. The DTs determine feasible CPU-allocation strategies guaranteeing class-specific age-of-information (AoI) requirements under specified workload, delay, and update-generation bounds. MEC servers collaboratively identify the feasible strategy minimizing expected freshness-normalized preprocessing delay. We formulate CPU-strategy selection as a fixed-confidence distributed best-arm identification problem in linear bandits, with MEC servers as agents and CPU allocations across task classes as arms. We employ DistLinGapE, a distributed collaborative learning method, to coordinate exploration and share observations. Numerical results show that the method identifies the optimal feasible strategy and reduces preprocessing delay when deployed under stationary task-class assignments and workload statistics. Collaboration yields an identification speedup nearly proportional to the number of participating MEC servers in the evaluated settings.
Modeling the nonlinear acoustic wave propagation in biological tissues needs accuracy, especially in ultrasound imaging and therapeutic applications. The 1D Westervelt equation, \( p_{xx} - p_{tt} + \gamma\, p_{ttt} + \eta\,(p^{2})_{tt}=0, \) could be used to describe a finite-amplitude ultrasound propagation in these tissues. This equation accounts for nonlinearity, \(\eta\), and attenuation, \(\gamma\). To solve this equation, we observe that traditional solvers may struggle to do so correctly. This is due to the difficulties arising from the nonlinear nature of the equation. We solve the equation by defining three synthetic cases - linear (\(\eta=0.001\), \(\gamma=10^{-6}\)), shock-dominated (\(\eta=0.02\), \(\gamma=10^{-6}\)), and dissipation-dominated (\(\eta=0.01\), \(\gamma=5\times10^{-4}\)) - and using Physics-Informed Neural Networks (PINNs) to showcase the solution. A sinusoidal activation function is used to capture the oscillatory behavior of the problem. The results of our study, with PDE residual converging to at least \(10^{-6}\), boundary condition loss to \(10^{-5}\), and the PINN loss converging to \(10^{-5}\) for each case at \(30,000\) epochs, show that \(\gamma\) scales the rate of amplitude decay, while \(\eta\) governs frequency-dependent phase speeds.
We study a fundamental problem in machine learning: partial recovery of a planted signal $\boldsymbol{v} \in \mathbb{S}^{n-1}$ from a spiked tensor $\boldsymbol{\mathsf{X}} = \lambda \, \boldsymbol{v}^{\otimes p} + \boldsymbol{\mathsf{Z}}$, where $\boldsymbol{\mathsf{Z}}$ has symmetric Gaussian entries of variance $\asymp n^{-1}$. While previous works replace $\boldsymbol{\mathsf{Z}}$ by an i.i.d. standard Gaussian tensor to circumvent technical difficulties caused by dependencies, this paper directly analyzes the symmetric model via random matrix theory. We show that the square unfolding algorithm is subject to a sharp phase transition for partial recovery at $\lambda \asymp n^{(p-2) / 4}$, validating an open conjecture from Montanari and Richard (2014). We get precise expressions for the limiting spectral distribution, deterministic equivalent, phase transition threshold and overlap. The main message is that unfolded symmetric tensors spectrally behave like standard Wigner matrices, up to simple transformations accounting for high-order symmetries. We consider extensions to multiple orthogonal spikes, recursive unfolding and non-square unfolding via disproportionate random matrix theory, leading to an interesting conjectured statistical-computational tradeoff regarding the choice of unfolding: square versus long.
Vision-language verifiers judge whether an image faithfully matches a textual description, yet existing training pipelines typically rely on static supervision from stronger external judges or curated hard-negative data. Such supervision requires external generation or annotation, is teacher-bounded, and is difficult to adapt as model failures evolve. We introduce PROVE (Perturb-and-Repair Optimization for Vision-Language Evaluation), a self-play framework that trains verifiers through a closed perturb-verify-repair-verify loop. Starting from a known-correct image-caption pair, a Mutator introduces a typed perturbation over objects, attributes, counts, or relations. The verifier then judges the perturbed caption and provides textual feedback, which a Repairer uses to recover a faithful caption for re-verification. These two verification rounds produce structured rewards for informative mismatches and visually grounded repairs. Since perturbation, verification, and repair are role-conditioned behaviors of the same evolving multimodal backbone, updating the shared parameters also improves the verifier itself. Experiments show that PROVE improves fine-grained compositional verification across model scales and families and strengthens visual-detail robustness on the reference backbone, with limited, benchmark-dependent changes on broader multimodal tasks.
Multimodal neural data can enable a more complete understanding of brain dynamics underlying behavior. Modalities such as spiking activity, local field potentials (LFPs), and behavior capture diverse spatiotemporal aspects of brain processes, yet prior neural-behavioral models that separate sources of variability rely on a single neural modality. By leveraging these complementary strengths, multimodal neural fusion can provide a unified, rich representation of brain-behavior processes and address the limitations of single-modality analyses. Here we develop BREM-NET, a nonlinear multimodal dynamical model that integrates behavior with multiple neural modalities having distinct statistical characteristics and temporal resolutions. BREM-NET performs multimodal neural fusion during inference while separating behaviorally relevant and residual neural dynamics via a structured learning framework. In a public multimodal non-human primate dataset with simultaneous spiking and local field potential recordings during a 2D reaching task, BREM-NET separates these two types of dynamics and learns them more accurately, as reflected in better behavior decoding, better neural prediction, and stronger latent alignment with behavior, and it retains these capabilities when neural modalities are asynchronous or suffer dropouts, which are major challenges in real-world recordings. This framework provides a new tool for studying behaviorally relevant neural computations across spatiotemporal scales of brain activity.
Model reduction via the Mori-Zwanzig (MZ) formalism is well understood for coarse-grained representations constructed from standard basis functions such as Fourier functions. We show that the same formalism can be used to build reduced-order models in bases discovered by physics-informed machine learning. The fluctuation-dissipation theorem (FDT) is used to close the resulting non-Markovian evolution equations for the expansion coefficients of the learned basis. We further introduce a complementary, generative closure model. Relative to the Markovian dynamics alone, both the MZ and generative closures substantially improve predictions of the filtered states.
We introduce PLANTEUR, a benchmark for crop-type mapping from satellite images in the French overseas territories. Building on the PASTIS design, PLANTEUR provides georeferenced and aligned Sentinel-2, Sentinel-1, and Landsat-8 sequences with parcel-derived labels spanning six tropical regions and an 8-class label space. We define both a standard test split and a leave-one-island-out protocol to measure cross-territory generalization, and evaluate fully supervised models, linear probes, and fine-tuned foundation models. The benchmark code and data are open source and available on GitHub and Hugging Face.
A general mathematical model of experience formation is considered, in which the learning level is defined as the probability that the next value of the uncertainty factor has already been mastered and not forgotten. Within this model the shape of the learning curve is derived, and it turns out to be determined primarily by the distribution of situation occurrence: a uniform distribution yields an exponential curve, whereas a power-law distribution yields a power-law curve whose exponent equals the ratio of the distribution exponent to that same exponent increased by one. Both mechanisms generating power-law curves, established earlier in the analysis of debugging, assembly and repair activities, are shown to have been independently rediscovered in deep learning theory as the "frequency" and "spectral" mechanisms of scaling laws for artificial neural networks, while the exponential curves of machine learning (contraction operators, strongly convex optimization, stationary reinforcement) are generated by the same constancy of the contraction coefficient that was revealed half a century earlier in models of radioactive decay, the first-order lag element, and others. A solution is given for the problem of choosing a training distribution that differs from the deployment distribution. A methodological consequence is discussed: three distinct mechanisms generate curves indistinguishable in shape, so an observed learning curve does not allow one to judge the "internal" structure of the learning system.
A policy can earn its performance by where it locates on average or by reacting to information. For any Markov policy $\pi$ in a finite-horizon MDP, let $\nu$ be its open-loop marginal, which plays $\pi$'s average action distribution at each step. We show that $J(\pi)-J(\nu)$ equals $\sum_t tr Cov(Q^\nu_t,\pi_t)$ under endogenous dynamics. Audits based on immediate costs alone can have the wrong sign. Writing the covariance as a joint-versus-product gap gives the bound $|J(\pi)-J(\nu)|\le\sum_t span_t\sqrt{I(s_t;a_t)/2}$, with a sharp constant. Hence, an MDP's adaptivity gap is bounded by the information that its optimal policy uses. Estimating $J(\pi)-J(\nu)$ from $\pi$'s own bandit-feedback logs takes $\Theta(\bar\chi^2 \log(1/\eta)/\delta^2)$ rounds, where $\bar\chi^2$ is an average $\chi^2$-divergence. At the same time, deterministic non-constant policies cannot be audited at all. We also provide an importance-sampling auditor for episodic MDPs.
Importance sampling is a building block of modern statistical computing, but the weighted estimate is prone to high or infinite variance. Moreover, its empirical standard error is often unreliable, leaving importance sampling without practical uncertainty quantification. Motivated by multilevel regression and poststratification (MRP) in survey sampling, we propose post-stratified importance sampling (PIS) for estimating target expectations from proposal draws. PIS reformulates arbitrarily high-dimensional weighting into two one-dimensional probabilistic modeling problems: regression for the conditional mean given the importance ratio and marginal density estimation for the importance ratio. Different modeling choices yield a family of PIS estimators, encompassing ordinary importance sampling as a special case. We develop a model-based implementation that combines one-dimensional Gaussian process regression for the conditional mean and score-informed heavy-tailed density estimation for the marginal law, using both the observed weights and the available density evaluations. We study the oracle gains of PIS and establish consistency under general and model-specific conditions. Simulations and applications to cross-validation demonstrate improvements in estimation accuracy and uncertainty quantification.
We introduce Gaussian partitioning, an algorithmic primitive for distributions smooth with respect to a Gaussian. It partitions space into polyhedral regions so that almost all mass lies in thin strips restricting a few orthogonal directions. Within each strip, the conditional distribution approximately matches low-degree Gaussian moments in the unrestricted directions. Our construction builds on logconcave partitioning of Diakonikolas, Kane, and Ma (FOCS 2026). Under input marginals smooth with respect to the standard Gaussian, we obtain agnostic learners with error $(1+c)\mathrm{opt}+\epsilon$ for broad classes of Boolean multi-index models, including functions of halfspaces, low-dimensional convex concepts, and low-dimensional polynomial threshold functions. For functions of $k$ halfspaces, both agnostic and testable agnostic learners achieve this error with dimension dependence $d^{\mathrm{poly}(k/c)}$, improving the previous exponential dependence on $k/c$ in the agnostic exponent. For learning from positive examples with imperfect unlabeled data (PUI), we give algorithms and Statistical Query (SQ) lower bounds that qualitatively characterize complexity through Gaussian moment matching. For fixed smoothness and positive-mass parameters, we learn intersections of $k$ halfspaces with dimension dependence $d^{O(\log(k/\epsilon))}$ and matching SQ lower bounds, improving the previous $d^{\mathrm{poly}(1/\epsilon)}$ upper bound even for one halfspace. A general reduction from agnostic reliable learning to PUI yields negative reliable learners for these intersections with the same dimension dependence, for fixed smoothness. Finally, for PQ learning and testable learning with distribution shift, we prove qualitatively optimal SQ lower bounds of $d^{\Omega(1/\epsilon^2)}$, even for intersections of two halfspaces with Gaussian training data and arbitrary test distributions.
Conformal prediction gives finite-sample coverage when calibration and test points are exchangeable, but exchangeability can fail when deployment changes the data distribution. We study conformal prediction in performative settings, where the deployed predictor affects the data distribution, and seek prediction sets valid at performative equilibrium: coverage must hold under $D(f^\star)$ for a performatively stable predictor $f^\star$. Under total-variation $L$-sensitivity of the distribution map, we prove a phase transition at $L=1$. If $L \ge 1$, no finite-sample procedure can uniformly guarantee both equilibrium-valid coverage and nontrivial prediction sets. If $L<1$, the minimax excess miscoverage over informative procedures is $\widetilde{\Omega}(1/((1-L)\sqrt{n}))$, implying a sample-complexity inflation of order $1/(1-L)^2$ for fixed excess miscoverage. Our main technical tool is a transfer theorem: any procedure returning a predictor within $\varepsilon$ of equilibrium yields equilibrium-valid coverage up to an additive $L\varepsilon$ term after split-conformal calibration. We instantiate the theorem with two retraining algorithms: a model-free method for discretizable classes and a parametric tail-averaged method that matches the $n$- and $L$-dependence of this rate up to logarithmic factors. Experiments, including semi-synthetic credit scoring, demonstrate the predicted threshold behavior and scaling laws.
Fine-tuning a large language model (LLM) on harmful data in a narrow domain, such as bad medical advice, can make it broadly misaligned. How realistic this risk is remains unclear, since models are very rarely fine-tuned on explicitly harmful data in practice. In this work, we show that fine-tuning on data the model believes to be false but that has no harmful content, can also induce broad misalignment. Across mathematics, trivia, cooking and sports, the more strongly the model disbelieves the answers it is trained on, the more misaligned it becomes. In mathematics, misaligned answers rise from 4% to over 60% as the training answers become more wrong, with little loss of general capability. After fine-tuning, the model states the wrong answers but its multiple-choice probabilities still prefer the true ones, which suggests that misalignment comes from this internal inconsistency. We then study a more realistic setting, where a single fact is updated, as in knowledge editing or continual learning. The same relationship holds and the more the model still believes the true fact after training on the false one, the more misaligned it becomes. When we frame the same false fact as part of a coherent story, the model comes to believe it more deeply and becomes less misaligned than when we instruct it to assert the false fact. Content filters and persona vectors cannot tell these two framings apart, even though they lead to different levels of misalignment. These results suggest a risk for model updates, where training a model that does not adopt the new fact as true can induce misalignment even when the data looks harmless.
Regularizing online reinforcement learning (RL) toward a pretrained base policy can improve sample efficiency and learning stability, but regularization that is too restrictive limits how much the policy can improve. Base policies -- typically diffusion or flow-matching models trained by imitation learning on large, diverse datasets -- capture many behaviors that are reasonable but suboptimal. Existing regularizers using such base policies are too restrictive in one of two ways: behavioral cloning regularizers pull the actor toward the full base-policy distribution, penalizing it for discarding suboptimal modes, while hard support constraints forbid improvements that lie just outside the base policy's support. We propose Geometric Support Behavior Regularization (GSBR), which penalizes actions by their distance from the base policy's support. Actions within the support incur no penalty, so the actor can concentrate on high-value modes; outside the support, where value estimates are less reliable, the penalty grows smoothly with distance while still permitting improvement when the value gained outweighs the penalty. GSBR estimates this distance from a finite set of base-policy samples, so it requires only sample access to the base policy. Across 19 tasks on four robotic benchmarks, GSBR raises the average success rate from 33% for the base policy to 95%, achieving the highest average success rate and throughput among all compared methods, with about 23% higher normalized throughput than the next-best RL method. Project website and open-source code at https://charlesxu0124.github.io/gsbr/ .
Efficient preconditioning plays an important role in large-scale first-order optimization. In this paper, we propose an adaptive accelerated optimistic dual averaging ($\text{A}^2$ODA) method for unconstrained convex optimization that adopts low-cost preconditioning based on $p$-norm geometry for $p\in[2,\infty]$. The method is line-search-free and employs a new stepsize selection technique based on an aggregated affine lower model. Its update structure is simple and unified across $p\in[2,\infty]$, with the case $p>2$ requiring no more elaborate updates than the Euclidean case $p=2$. Under a generalized local smoothness condition, we derive a convergence bound that separates the effects of constant and gap-dependent curvature, with the latter potentially yielding faster-decaying terms. In the special case of local smoothness, this bound recovers a nearly optimal convergence rate.
In-context learning (ICL) lets a pretrained language model perform a task specified by demonstrations in its prompt, with no parameter updates. With arbitrary labels the model must acquire a mapping it cannot have memorized, so the learning happens within a single forward pass. Prior work found candidate parts but not how they compose. We give the composition for few-shot sentence classification with arbitrary labels in three 3-4B-parameter transformers. The model implements a key-value memory over label-token positions: each label token's residual stream encodes the label's identity, and middle-layer attention heads write into it the class of the preceding demonstration as its key. At the query, upper-layer heads match the query's class against the stored keys and copy out the matching labels. An early-layer signal encoding whether any label has repeated gates this readout; in its absence the readers attend to the start of the prompt and no label is favored. A separate set of heads suppresses the most recently seen label. Each step is established causally by activation patching, head ablations against random-head baselines, and steering with the model's own signals. A softmax model of the readout with two constants per head, fit on one prompt distribution, predicts another. The computation is conserved across the three models, the named heads alone reproduce the behaviour, and a transformer trained from scratch grows the same parts. The circuit transfers to semantic concepts and to labels with meaning, and the same parts, with a causal gate, appear from 0.6B to 8B in one family and survive instruction tuning. ICL in this setting is a gated associative lookup built from a few identifiable component types, traced causally from demonstrations to prediction.
When a language model answers, we can usually only estimate why. We build a transformer in which every step is a Born measurement - attention $|\langle q|k\rangle|^4$ normalised over each row, a measured product in place of the feed-forward nonlinearity, and a Born readout. With its denominators held at their values for the input, each class probability is a polynomial in the word amplitudes, so its integrated-gradient evidence is computed by a finite quadrature with no quadrature error: the words' contributions sum to the class probability above its no-information share, to $10^{-11}$ or better. On e-SNLI natural language inference, with about 98k parameters outside the embedding, the model reaches $76.98 \pm 0.17$% test accuracy, level with a softmax transformer of the same size ($76.62 \pm 0.36$%). Deleting the word its evidence ranks first changes the answer in 45% of test pairs, out of the 53% any single deletion can change, against 32% for the softmax model explained the same way and 40% by its attention x gradient. Attention can look more like the words annotators highlighted, yet accounts for the answer less well: deleting the word it ranks first changes the Born Transformer's answer in 32% of pairs. The same weights compile exactly to optoelectronic optics and to qubits, and a coherent attention layer becomes one circuit equal to the model (852 CNOTs on 23 qubits). On IBM's ibm_kingston, reading only the attention overlaps by Hadamard tests, the device gave the model's answers on two short test pairs and kept or changed them under deletion as the model does; the evidence itself did not yet survive the device, and we work out the shots and calibration that measuring it on hardware would need.
Undesired model behaviors can emerge or re-emerge at any stage of training. Once observed, how effectively can we suppress them? To study this question, we first construct two model organisms--language models deliberately trained to exhibit a specific misalignment--that differ in what triggers the behavior. The narrow-trigger organism misbehaves only when the prompt contains a specific cue: it cheats on multiple-choice questions by following a planted hint that points to a wrong answer, and behaves normally otherwise. The broad-trigger organism has no such cue: after fine-tuning on insecure code, it gives harmful responses across a wide range of open-ended prompts. We then formalize post-hoc behavioral suppression and evaluate methods along three criteria: immediate suppression of the target behavior, durability under further training, and preservation of general capability. We compare standard methods--SFT, GRPO, and gradient-ascent unlearning--along these dimensions. Building on prior work on alignment shallowness, we hypothesize that more durable suppression requires the ability to recover from a wider range of misalignment states. We introduce Prefix-Expanded Adversarial Reinforcement Learning (PEARL), a GRPO variant that augments rollouts with continuations from intermediate states sampled along cached organism trajectories. Across both organism settings and two model families (Qwen3-4B, GPT-OSS-20B), PEARL achieves strong immediate suppression while preserving task accuracy, and remains durable under further fine-tuning on either the organism's data or unrelated data. It obtains the greatest benefit when the undesirable behavior is broad and cannot be addressed through targeted SFT. Together, our framework and method offer a step toward more durable post-hoc removal of undesired behaviors.
Imitation learning policies in robotics are trained on datasets collected through costly real-world teleoperation. Often, the data-collection process is unguided, relying on the teleoperator's intuition, resulting in data that contributes little to overall policy performance. Recent work has shown that curating such unguided datasets can substantially improve policy performance, but these approaches operate by filtering existing datasets after collection and therefore do not prevent the acquisition of uninformative data. In this work, we propose GUILD, a novel Guided Data Collection (GDC) method that actively reasons about which data should be collected to maximize downstream policy performance. We introduce an iterative framework that alternates between 1) task proposal and data collection and 2) data evaluation to close the loop. For task proposal, we leverage semantic priors from vision-language models to identify relevant environments, object configurations, and goals for data collection. However, since semantic relevance does not imply policy improvement, we perform data evaluation. By leveraging predictive data attribution we approximate how the collected trajectories influence policy performance on a target deployment setting. This signal is then used as feedback to guide subsequent data collection, closing the data collection loop. We evaluate our approach in simulation and real robot settings, demonstrating that guided data collection substantially improves sample efficiency and deployment performance compared to passive data collection strategies, multiple baselines and experienced roboticists. More information at https://robin-lab.cs.utexas.edu/guild2026/
Inverse kinematics (IK) converts end-effector pose commands into joint motion, and many vision-language-action (VLA) policies rely on it to act on a given arm and to transfer across embodiments. Precise IK remains laborious: redundancy and singularities complicate the inverse mapping, execution depends on hand-tuned controllers, and training data on physical robots are costly to collect. We present PolicyIK, a goal-conditioned reinforcement learning policy for high-precision IK. It updates joint-position references from pose feedback and observes its accumulated reference relative to the measured joints, so it can retain the offset that the position controller needs at rest. Two-scale pose errors with an error-dependent action gain resolve millimeter residuals while preserving large motions. Progress and dwell rewards teach it to both reach the tolerance and stay within it. In simulation, a single attempt by the trained PPO policy keeps the mean errors over the final second within 1 mm and 0.2{\deg} on 93.5% of held-out tasks, and a deployment protocol raises success to 98.9%. When trained directly on a physical arm with an off-policy RL learner, PolicyIK achieves higher reaching and holding success than all four classical IK controllers under the same time budgets. In simulated pick-and-place, PolicyIK executes the end-effector commands of the $\pi_{0.5}$ VLA policy with 95.7% success. These results show that reinforcement learning can provide high-precision inverse kinematics for robot arms. Code is available at https://github.com/Tinaliu0123/PolicyIK.