CS Paper Radar

Daily listing

2026-11-17 · cs.LG

149 papers first submitted this day (UTC) in cs.LG — Machine Learning. arXiv announces a day's submissions the next evening, so the newest day fills up over ~24 h.

Score = weighted mean ofCitations 20%Upvotes 10%Code stars 10%Venue 15%Code/data 10%Novelty 15%Momentum 15%Substance 5%how it works

AI & Machine Learning 109

Weight-Space Programming Doesn't Need Transformers

cs.LG cs.AIcs.CL ·2026-11-17 ·Hyunjong Ok, Jaeho Lee
code/datacommunity pick
Abstract · Preprint (Work in progress)

Standard neural architectures such as the Transformer were chosen because they train well with gradient descent and run fast on GPUs. In weight-space programming, a large language model (LLM) writes target weights by hand without gradient-based training, so those reasons no longer apply. Should we keep the Transformer? Given minimal guidance, ten labeled examples, and a sandbox, mainly on multiple-choice questions, we task three frontier LLMs with either generating the weights of a fixed architecture or designing the architecture themselves. When free to design, all three compile plain-text facts into a sparse table queried by a fixed procedure, reaching up to 54.4% test accuracy. An LLM-designed model achieves higher test accuracy than the most accurate released Transformer that the same LLM wrote while using about 1,800x fewer parameters. Without gradients, they choose a table, in which the LLM can state what each weight means and can find and rewrite a wrong one. Each wrong answer points to one row, and removing a row changes about one other answer. From this, we propose an associative-memory-based model up to 17x faster on GPUs than a same-sized Transformer. We also test pretrained weights, other benchmarks, and tasks where models generate answers. Pretrained LLMs remain more accurate; the claim concerns which structure to build when writing weights.

0.68

The Kernel Reality Check: Benchmarking and Distilling Efficient Attention at Scale

cs.LG ·2026-11-17 ·Fırat Öncel, Cem Subakan, Mirco Ravanelli, Çağatay Yıldız
top venuecode/data
Abstract · Accepted at NeurIPS 2026. 20 pages

Kernelized attention methods, which aim to replace the $O(N^2)$ complexity of softmax attention with $O(N)$ alternatives, have been the focus of intense research in the last five years. Despite growing interest, systematic evaluations and fair comparisons across approaches and scales are still missing. We analyze kernelized attention in two halves and highlight prominent limitations on each. On efficiency, FlashAttention achieves lower memory consumption and faster wall-clock time than SOTA kernelized methods during training and in vision transformers across most sequence lengths (with inference-time memory in causal language models the only clear exception), contradicting the core assumption for pure kernelized attention used as a drop-in replacement for softmax. We also discover that reported efficiency gains often fail to reproduce against modern baselines. On quality, current kernelized methods leave gaps of 2-10 points relative to softmax under a unified distillation protocol, and model scala alone does not close these gaps. We show this gap is not intrinsic to kernelization: LARA++, a principled extension of LARA that approximates softmax via importance sampling with adaptive proposal distributions, recovers up to 99% of softmax accuracy. We establish that the gap is closable; however, the computational overhead remains substantial, leaving the efficiency wall intact. Code is available at https://github.com/firatoncel/KernelRealityCheckNeurIPS2026

0.65

DDSL: Iterative Self-Labeling for Financial Market States

cs.LG ·2026-11-17 ·Xianbin Gu, Qingyue Zhu, Yuhong Geng
publishedcode/dataunusual
Abstract · 17 pages, 3 figures, 5 tables. Conference version published in ICAIF 2026. Code available at https://github.com/hogwild/DDSL

Financial market states lack a unique ground truth and must therefore be constructed from price data. Conventional labeling methods typically define states independently of the downstream classifier, leaving unresolved whether the resulting labels can be recovered from information available at prediction time. We propose DDSL, an iterative self-labeling framework that couples a price- and grammar-guided Viterbi labeler with a Transformer classifier through model feedback. DDSL constructs a nine-state representation that jointly encodes current and future market trends while retaining structured price emissions and temporal constraints throughout refinement. Experiments on SSE50, CSI300, CSI500, and NASDAQ100 show that DDSL increases held-out joint-state macro F1 by 78%-96% over single-pass dynamic-programming labeling. An independent GRU classifier that never participates in the labeling loop reproduces the same trend, while label entropy and class coverage remain stable, indicating that the gain transfers across architectures and is not caused by label collapse. We further show that learnability is controlled by both iteration and the model-feedback weight $β_{\mathrm{model}}$. Iterative gains diminish as successive label systems converge, whereas stronger feedback increasingly widens the performance gap between the feedback Transformer and the independent GRU. Increased learnability is accompanied by declining segment-level price-feature separation after an initial improvement. Finally, a non-anticipative event study provides partial external validation: refinement strengthens a conditional BEAR$\rightarrow$BULL reversal signal on the three A-share datasets, while the corresponding effect has the opposite sign on NASDAQ100. These results position DDSL as a structured framework for constructing market-state labels whose learnability can be explicitly controlled and evaluated.

0.60

Fair Bubble Sort: Provably Optimal Fair Ranking with Continuous Sensitive Attributes

cs.LG cs.CYcs.DS ·2026-11-17 ·Antonio Ferrara, Fabio Vitale, Atsushi Miyauchi, Francesco Bonchi
top venuetheorycode/data
Abstract · 35 pages, 9 figures, 4 tables. Accepted at NeurIPS 2026. Code: https://github.com/Ambress92/FairBubbleSort

Most fair-ranking methods assume categorical protected attributes, requiring continuous attributes such as age or income to be discretized into groups. This introduces arbitrary thresholds, loses fine-grained information, and ignores the natural ordering of the attribute. In this paper, we study the problem of fair ranking with continuous protected attributes without relying on thresholding. We introduce Continuous Pairwise Parity (CPP), a fairness notion based on Kendall correlation between ranking positions and the protected attribute, and seek rankings that satisfy a prescribed CPP constraint while minimizing their Kemeny distance from an initial score-based ranking. To solve this problem, we propose Fair Bubble Sort (FBS), an efficient post-processing algorithm based on targeted adjacent swaps. We prove that FBS is exactly optimal for the unweighted Kemeny distance and provide approximation guarantees for the score-weighted setting. FBS is output-sensitive, with running time determined by the number of swaps required to meet the fairness constraint. Experiments on synthetic and real-world datasets show that FBS is scalable and consistently outperforms discretization-based baselines, achieving a superior fairness-accuracy Pareto front.

0.54

Virtual Double Oracle: Faster Convergence by Leveraging Strategies That Were Watching All Along

cs.LG cs.GTcs.MA ·2026-11-17 ·Ruilan Wang, Francisco Aristi Reina, Katerina Papadaki
top venuetheory
Abstract · Accepted at NeurIPS 2026. 25 pages

Double Oracle (DO) and PSRO approximate Nash equilibria in two-player zero-sum games by iteratively expanding a strategy population with best responses supplied by an oracle. Online Double Oracle (ODO) interleaves multiplicative weights updates (MWU) with population discovery and obtains an $O(\sqrt{k\ln k/T})$ rate, where $T$ is the number of learning rounds and $k$ is the larger of the two players' final population sizes. We identify the reinitialisation step as an underexplored design choice: at each expansion, MWU-based methods reset newly discovered strategies to a uniform prior. We introduce Virtual Double Oracle (VDO), which exploits a bilinear identity (a strategy's full hindsight loss history collapses to its utility against the time-averaged opponent) to initialise each new strategy with the MWU weight it would have had from the start. We prove that VDO achieves $O(\sqrt{k/T})$ exploitability over the discovered population as it grows, provided the cumulative cost of adding strategies is at most linear in $k$, a condition supported by our empirical measurements. Discounted VDO (dVDO) moderates historical preferences to reduce over-concentration at reinitialisation. Experiments across matrix games, poker, and Goofspiel show that, with a suitable discount, dVDO can substantially reduce finite-horizon exploitability relative to both VDO and uniform-reset ODO, and can improve the discovered strategy populations.

0.54

FrameRouter: Frame Budget Routing for Long-Video Understanding on Video-MME-v2 and Beyond

cs.AI cs.CVcs.LG ·2026-11-17 ·Bingjun Luo, Jialin Guo, Siqi Li
top venuecode/data
Abstract · Accepted to NeurIPS 2026

Long-video question answering benchmarks like Video-MME-v2 typically evaluate systems under a uniform frame budget for every question, despite stark variation in the evidence each query actually demands. We analyze this mismatch on Video-MME-v2 and find that different question subsets prefer different frozen frame-budget policies, while uniformly increasing the frame count does not reliably improve all metrics. These findings motivate fixed-mean frame-budget routing, a controlled inference setting in which each question may receive a different frame budget while the average budget matches a uniform reference. We introduce FrameRouter, a training-free test-time framework that combines an evidence-demand router, a budget-constrained allocation rule, and a frozen bank of frame-sampling policies. By controlling how much visual evidence each question receives while leaving the underlying samplers unchanged, FrameRouter is orthogonal to query-aware frame selection. Across Video-MME-v2 and several long-video understanding benchmarks, FrameRouter improves long-video question answering by using the same average visual budget more selectively across questions. Code is available at https://github.com/bingjunluo/FrameRouter.

0.51

Balancing Exploration and Exploitation in Constrained Combinatorial Optimization with U(1)-Symmetric Tensor Trains

cs.LG math.OC ·2026-11-17 ·Sergei Iudin, Mikhail Podobrii, Dmitry Zheltkov, Stanislav Moiseev
publishedcode/data
Abstract · 39 pages, including supplementary material. Code: https://github.com/antientropic-guy/OhMyU1_public. Extends an earlier version presented at…

Tensor Train (TT) decomposition has emerged as a powerful tool for mitigating the curse of dimensionality in a variety of machine learning and combinatorial optimization tasks. Recent advances show that TT-based generative models can efficiently represent and explore high-dimensional discrete spaces, even when the cost function structure is unknown. However, incorporating hard linear constraints into TT-based optimization remains challenging, as existing interpolation and sampling approaches often fail when feasible points are rare. In this work, we develop a general TT-based framework for constrained combinatorial optimization built upon U(1)-symmetric tensor networks. Our method introduces an expressibility-enhancement mechanism for constructing sparse symmetric tensors and a graph-based diversity measure that guides sample selection during training. On assignment benchmarks, expressibility enhancement and diversity-aware selection improve mean solution quality over their respective baselines. The assignment-selected diversifying strategy also extends to nonpolynomial Set Partitioning, with larger relative gains observed in the higher-tensor-rank data families. Separate runtime benchmarks show substantial speedups over ConstrainTNet and PROTES on non-degenerate assignment workloads; comparisons with simulated annealing depend on constraint information and computational budget, but, overall, support the practicality of the current implementation.

0.50

Learned versus Kinematic Trajectory Prediction under Reduced Observation Frequency

cs.LG cs.RO ·2026-11-17 ·Nukhbala Gadzhiev, Alexander Morochko
code/data
Abstract · 8 pages, 1 figure, 2 tables. Code and per-scenario results: https://github.com/nleaxy/learned-vs-kinematic-trajectory-prediction

Trajectory prediction models for automated driving are usually trained on data with a fixed sampling frequency, while positions received via vehicle-to-everything (V2X) communication arrive at 1-10 Hz. We study whether a learned predictor that uses only the observation history keeps its advantage over constant velocity (CV) and constant acceleration (CA) models when the history frequency drops. On the Argoverse 2 motion forecasting dataset, the 5 s history of the focal agent was downsampled from 10 Hz to 5, 2, and 1 Hz, and the 6 s future was predicted without maps or neighboring agents. An LSTM-based recurrent neural network (RNN) trained for each frequency was compared with the same network trained at 10 Hz and applied to sparse histories after linear interpolation. Hypotheses and thresholds were fixed before training, the main RNNs were trained with ten seeds each, and results on 24,984 test scenarios have bootstrap confidence intervals over scenarios and seeds. The average displacement error of the frequency-specific RNN grew by 3.1% at 5 Hz, 10.1% at 2 Hz, and 20.3% at 1 Hz. Contrary to our hypothesis, its advantage over the better kinematic model did not vanish on straight motion and was 24-30% on both straight motion and turns at every frequency, while for pedestrians CV was more accurate. The 10 Hz model with interpolation was at most 1.3% less accurate than the frequency-specific model down to 2 Hz but 11.5% less accurate at 1 Hz. Each hypothesis was only partly supported.

0.49

When does deep reinforcement learning pay off in HVAC supervisory control? Four small Modelica case studies trained directly on FMUs

cs.LG cs.AIeess.SY ·2026-11-17 ·Bingnan Zhao, Harshala Gammulle, Tharindu Fernando
code/data
Abstract · 16 pages, 4 figures, 2 tables. Code, models and trained agents: https://github.com/Bingnan-Zhao/hvac-drl-lab

Deep reinforcement learning (DRL) is often proposed for the supervisory control of heating, ventilation and air-conditioning (HVAC) systems. The reported gains are, however, difficult to interpret, because the baselines are often weak and the attainable optimum is seldom known. This paper examines four small, reproducible case studies based on an office building in Brisbane, Australia: pre-cooling under a time-of-use tariff; chiller staging with condenser water reset; chilled water supply temperature reset with a wet cooling coil; and chilled water storage under time-of-use prices and a demand charge. The plants were modelled with the Modelica Buildings library, and in three of the cases a soft actor-critic agent was trained directly on the exported functional mock-up unit (FMU). Each agent was compared with tuned rule-based controllers and, where practicable, with a near-optimal reference from hourly exhaustive search or dynamic programming with perfect foresight. In the two cases without storage the agent reached the near-optimal reference, but the margin above a well-chosen rule was small: 1.3% for condenser water reset and 0.18% for chilled water reset. In the latter case two common reset rules consumed more energy than a fixed set point. With a chilled water storage tank the margin was considerably larger. The agent outperformed both rules (-7.8%, compared with -3.9% and -6.3%) and adopted midday charging during the solar off-peak period, yet it obtained only about 45% of the saving achieved by the perfect-foresight schedule and did little to limit the demand peak. In the pre-cooling case a hand-tuned rule remained superior (-12.0% compared with -5.6%). These results suggest that studies of DRL for HVAC control should state the margin between the best rule and a near-optimal reference. Some practical difficulties encountered in training on co-simulation FMUs are also reported.

0.48

WATERFALL: Workflow for Adaptive Training with Evolutionary Reward Formulation and Automated Learning Loops

cs.LG cs.AI ·2026-11-17 ·Eleftherios Triantafyllidis, Filippos Christianos, Zhibin Li, Bernd Bickel
publishedbenchmarkcode/data
Abstract · Accepted at NeurIPS 2026 (poster). 78 pages, 13 figures, 19 tables. Code: https://github.com/etriantafyllidis/waterfall OpenReview: https://…

Reward design remains a central bottleneck in reinforcement learning (RL), particularly in sparse-reward, long-horizon and partially observable settings requiring dependency chaining, memory, and deceptive affordance disambiguation. While recent foundation-model approaches reduce manual reward engineering, most assume goal-conditioning, task descriptions, expert demonstrations, or ground-truth metrics. These assumptions are problematic in partially observable environments, where task-relevant objectives, affordances, and dynamics must be discovered through interaction rather than disclosed a priori. We formalise this problem setting by introducing WATERFALL, an iterative population-based workflow for automated reward discovery under the strict observability constraints inherent in POMDP environments. WATERFALL achieves this by (i) synthesising diverse programmatic reward candidates via persona-conditioned generation, (ii) evaluating candidate behaviour from raw visual rollouts utilising a Swiss-system evolutionary tournament judged by vision-language models to identify elites, and (iii) leveraging longitudinal assessment histories to iteratively mutate, escalate, or prune candidates. We evaluate WATERFALL across fully observable and reveal-gated partially observable environments ranging from discrete to continuous, comparing against sparse-reward RL, intrinsic-motivation and foundation-model reward-synthesis baselines. Our results indicate that WATERFALL discovers and refines reward programs without access to privileged information, outperforming baselines that retain their native information advantages, most clearly in the POMDP regime. Extensive ablations substantiate that visual evidence, tournament-based selection, and iterative refinement are jointly important. Enforcing this contract is what separates autonomous reward discovery from benchmark-privileged reward shaping.

0.46

Every Measurement, Every Direction, All at Once: Multimodal Flow Matching for Molecules and Spectra

cs.LG physics.chem-ph ·2026-11-17 ·Thorben Prein, Elton Pan, Rafael Gómez-Bombarelli, Santiago Miret
top venueposition
Abstract · 30 pages, 12 figures. Accepted at NeurIPS 2026

Generative modeling in science is often data-limited, where experiments are expensive. However, each experiment may yield several coupled measurements of the same physical entity. Standard generative models often train only on a single direction, discarding supervision from the rest. We argue that each experiment contains more training signal than this asymmetry allows the model to exploit: random masking enables measurements to serve as either conditioning inputs or prediction targets, yielding multiple supervised denoising tasks from a single experiment. We investigate this premise with MOSAIC, a multimodal flow matching transformer over molecular 3D structure and UV, IR, and Raman spectra. MOSAIC is a single model that is capable of any-to-any generation: generating any subset of modalities from any disjoint subset of the others. On QM9S, MOSAIC achieves 88.8% molecular graph recovery, outperforming the strongest spectra-to-structure baseline by 23.2% in structure generation, while simultaneously outperforming a structure-to-spectra specialist model in spectral generation. Ablations show that the any-to-any objective improves structure generation over structure-only training across model sizes. Furthermore, MOSAIC supports arbitrary subsets of available measurements and enables molecular identification by matching generated spectra to experimental references.

0.44

Right Labels, Wrong Reasons: Signed Feature Attribution for Multi-Label Explainability

cs.LG cs.AI ·2026-11-17 ·Kenton White
Abstract · 7 pages, 1 figure, 5 tables. Accepted at the 2027 IEEE Symposium Series on Computational Intelligence (IEEE SSCI 2027), Gold Coast, Australi…

When a multi-label classifier assigns the right labels, it may do so for the wrong reasons. Diffuse feature combinations may vary across classifier families and lack a dominant feature. This paper proposes FLEX (Fisher LDA Explainer), a global surrogate that fits a single direct multi-label Fisher LDA over the full binary label matrix, producing signed feature attribution lists for each label in a single coherent coordinate system. Across the evaluated benchmarks and classifier families, FLEX reproduces the black box's predictions with consistently high fidelity. Pooling covariance across labels outperforms a per-label alternative on most datasets, with the advantage growing largest where labels are numerous and positive examples are scarce. FLEX's fidelity-complexity Pareto curve is monotonically increasing, giving practitioners a predictable path to higher-fidelity explanations as more features are used, and its in-sample fidelity matches or leads Binary Relevance alternatives once a modest feature budget is allowed. The magnitude of each discriminant weight vector bounds its label's separability, flagging labels that may be predicted correctly for the wrong reasons. The evaluated classifiers largely agree on which predictions to make, but reach those decisions through different feature subsets when no single feature dominates.

0.43

Addressing Exogenous Variability in Cooperative Multi-Agent Reinforcement Learning

cs.MA cs.LG ·2026-11-17 ·Seongmin Kim, Woohyeon Byeon, Jiwon Jeon, Seungyul Han, Youngchul Sung
top venue
Abstract · 28 pages. Includes supplementary material. Accepted at NeurIPS 2026

Cooperative multi-agent reinforcement learning (MARL) often fails under exogenous variability such as shifts in opponents or environment regimes that cannot be controlled by the team but alter the transition dynamics. We formalize this challenge as Exogenous Dec-POMDP (ED-POMDP), in which an episode-level exogenous regime, selected independently of the team, indexes the dynamics, and the global state is partitioned into components that are directly responsive to the current team action and components that are not directly action-responsive at the current step. This formulation exposes mediated influence: while team actions do not directly update the latter at the current step, they can shape their later evolution through induced changes in the former. Based on this, we propose LEICA, a CTDE-compatible algorithm that learns history-conditioned context representations for both components and shapes policy updates using agent-wise, influence-weighted counterfactual intrinsic rewards. Across SMAX benchmarks with opponent strategy shifts, LEICA consistently improves both training performance and generalization to unseen opponents' strategies over existing baselines, supporting the usefulness of influence-weighted shaping under exogenous train-test regime shifts.

0.43

BQ-LoRA: Binary-Quantized Low-Rank Adapters as Implicit Regularizers for Parameter-Efficient Fine-Tuning

cs.LG ·2026-11-17 ·Juyoung Park, Heejun Lee
top venuebenchmark
Abstract · Accepted at NeurIPS 2026. 20 pages, 4 figures, 21 tables

Parameter-efficient fine-tuning of quantized large language models is dominated by QLoRA, which combines 4-bit weight quantization with FP16 low-rank adapters. We ask whether the FP16 adapter precision is necessary. Replacing it with a single binary bit per adapter parameter, combined with a learned per-layer scale $s$ and a Scaled Straight-Through Estimator, reduces adapter storage by 13-16x. We introduce two variants -- NormScale, which binarizes under a norm-preserving scale with a clipped STE, and LearnThresh, which makes the binarization threshold learnable -- unified under a framework we call BQ-LoRA. Our main theoretical contribution is a tight characterization of the BQ-LoRA hypothesis class: it has bounded entrywise norm, a typical-set spectral norm bound that is $Θ(r)$ times sharper than the worst case, a full-rank, near-isotropic update, and consequently a Rademacher complexity that yields an explicit $\tilde{O}(s_{\max}α\sqrt{d_{out}/(rn)})$ generalization bound. FP16 LoRA has no comparable a priori norm or isotropy control. On Qwen2.5 (1.5B-72B), BQ-LoRA exceeds QLoRA's benchmark averages at every scale by 0.8-3.5 points; on LLaMA-2 (7B/70B) it remains competitive at the same storage reduction but does not uniformly lead the strongest FP16-adapter baselines (LoftQ at 7B, QLoRA at 70B). On Qwen2.5-7B the binary adapter preserves the 4-bit base WikiText-2 PPL within 0.06 (9.86 vs. 9.80) while QLoRA worsens it by 0.59. The reduction applies to inference-time adapter storage and serving; training memory matches QLoRA and step time rises by under 4%. We translate the storage saving into a deployment study, BQ-SGMV, by extending Punica's SGMV kernel with a binary inner-product primitive: the same A100-80GB holds 15.6x more adapters in HBM, and our research prototype eliminates cache misses under a skewed synthetic workload.

0.43

Discovery of new antiviral agents through artificial intelligence: in vitro and in vivo results

cs.LG q-bio.BM ·2026-11-17 ·Roza Izmailyan, Mher Matevosyan, Hamlet Khachatryan, Anastasiya Shavina, Smbat Gevorgyan +7 more
publishedunusual
Abstract · Accepted manuscript. 13 pages, 5 figures; supplementary material included (supplementary text, Tables S1-S2, Figures S1-S5). Published in An…

In this research, we employed a deep reinforcement learning (RL)-based molecule design platform to generate a diverse set of compounds targeting the neuraminidase (NA) of influenza A and B viruses. A total of 60291 compounds were generated, of which 86.5% displayed superior physicochemical properties compared to oseltamivir. After narrowing down the selection through computational filters, nine compounds with non-sialic acid-like structures were selected for in vitro experiments. We identified two compounds, DS-22-inf-009 and DS-22-inf-021 that effectively inhibited the NAs of both influenza A and B viruses (IAV and IBV), including H275Y mutant strains at low micromolar concentrations. Molecular dynamics simulations revealed a similar pattern of interaction with amino acid residues as oseltamivir. In cell-based assays, DS-22-inf-009 and DS-22-inf-021 inhibited IAV and IBV in a dose-dependent manner with EC50 values ranging from 0.29 uM to 2.31 uM. Furthermore, animal experiments showed that both DS-22-inf-009 and DS-22-inf-021 exerted antiviral activity in mice, conferring 65% and 85% protection from IAV (H1N1 pdm09), and 65% and 100% protection from IBV (Yamagata lineage), respectively. Thus, these findings demonstrate the potential of RL to generate compounds with promising antiviral properties.

0.42

Flocking Time and Graph Volume in Nonlinear Cucker-Smale Models with Dynamic Networks

cs.LG math.DSnlin.AO ·2026-11-17 ·Sooyeon Song, Jea-Hyun Park
publishedtheory
Abstract · 14 pages, 3 figures. Accepted manuscript of the article published in J. Inf. Commun. Converg. Eng. 24(2), 155-164 (2026)

This study proposes a Cucker-Smale model incorporating dynamic networks and the discrete $p$-Laplacian ($p \in (1,2)$) and analyzes the relationship between flocking time constraints and graph connectivity from both theoretical and experimental perspectives. We establish that this model achieves flocking within finite time and derive a quantitative relationship between the lower bound of flocking time and graph volume. Based on these results, we observe that imposing sufficiently small flocking time induces an increase in network volume, and that larger initial velocity variance requires greater volume under the same time constraint. In our experiments, we trained a transformer-based PPO policy to select network structures at each time step. Under the most stringent time constraint, the maximum average number of connections increased sharply, confirming a threshold effect. The negative correlation between convergence time and average number of connections reflects the theoretical results well. This study provides quantitative evidence for the phenomenon that time constraints induce connection mobilization.

0.41

PANO: Parallel Latent Space Neural Operator for Solving PDEs

cs.LG cs.AI ·2026-11-17 ·Zijiang Yang, Xiaomeng Wu, Chen Kuang, Dongmei Fu
code/data
Abstract · 39 pages, 24 figures, 23 tables

Aggregating observations into a limited number of latent tokens and learning physical interactions in the latent space has emerged as an effective approach for Transformer-based neural operators to solve Partial Differential Equation (PDE) problems involving large-scale observations. However, existing methods typically compress observations using a single aggregation pattern, making physical information not retained by that pattern invisible to the latent space and thereby constraining physical interactions within it. To address this limitation, we propose PArallel Latent Space Neural Operator (PANO), a novel neural operator for solving PDEs that formulates physical interactions among observations as an additive combination of interaction results from parallel latent spaces with shared parameters. Specifically, PANO transforms observations into multiple distinct compressed representations through independent mappings, learns physical interactions in parallel based on these representations, and finally fuses the decoded results to predict the target physical field. Within PANO, we further propose Coupled Assignment Projection (CAP) to construct bidirectionally consistent encoding and decoding mappings, and Pairwise Latent Space Attention Layer (PaLA) to jointly update latent representations of each latent space through adaptive cross-space interactions. Extensive experiments on nine benchmarks spanning diverse physical systems, geometries, and observation scales demonstrate that PANO achieves state-of-the-art performance. Notably, PANO reduces equivalent plastic strain prediction errors on Unilateral Stamping by 28.3% (or more) and lift-coefficient prediction errors on AirCraft by 21.4% (or more) compared with existing methods. Code is available at https://github.com/ZijiangY1116/PANO.

0.40

WoRD: Diffusion-Based Posterior Inference for Online Goal Recognition

cs.AI cs.LGcs.RO ·2026-11-17 ·Almog Anschel, Sarah Keren
top venue
Abstract · Accepted at the 40th Annual Conference on Neural Information Processing Systems (NeurIPS 2026)

Goal recognition is the task of inferring an agent's intended goal from its observed behavior. Most existing work focuses on high-level symbolic recognition, abstracting away the continuous motion through which agents reach their goal. Methods that do reason over continuous trajectories rely on an explicit dynamics model that is hard to acquire and rarely transfers across settings. Neither approach explicitly learns how agents move toward each candidate goal, limiting accuracy on ambiguous trajectory prefixes where such knowledge is most needed. We present WoRD (Waypoint Recognition with Diffusion), which uses a learned trajectory prior to infer the agent's goal from partial observations. At each step, WoRD runs one warm-started denoising chain per candidate goal, steers each chain toward the observed trajectory through a closed-form log-likelihood gradient, and reweights the resulting samples by Bayes' rule into a calibrated posterior over goals. We evaluate WoRD on simulated robot settings, pedestrian trajectories, and vehicle trajectories using four metrics targeting distinct failure modes. Across all three settings, the same inference procedure preserves calibrated multimodal beliefs through ambiguous prefixes where filter- and classifier-based observers collapse to a single mode or grow overconfident.

0.39

The Illusion of Grounding: Do Vision-Language Models Weigh Retrieved Text and Images Like an Ideal Observer?

cs.LG cs.CLcs.CV ·2026-11-17 ·Basavaraj R M, Harshal Chandel
code/data
Abstract · 10 pages, 5 figures, 11 tables. Code and data: https://github.com/Basavarajearlybird/GIB-200K (archived at https://doi.org/10.5281/zenodo.24…

Retrieval-augmented vision-language models (VLMs) answer from two sources of evidence that can each fail: the retrieved text can be wrong or noisy, and the image can be degraded or contradicted. We introduce GIB-200K, 200,000 image-question items built from COCO and Visual Genome, each expanded into controlled stress conditions: hard-negative retrieval noise (1-20 documents), visual ambiguity (Gaussian and motion blur, low-resource crops) and cross-modal conflict with counterfactual flip pairs. Instead of only measuring degradation, we ask how models combine the two modalities. From ideal Bayesian cue combination we derive a testable prediction that yields formal definitions of Knowledge Inertia (text over-weighted relative to the model's own image evidence) and Sensory Capture (the reverse), and we estimate the weights with a measurement-error-corrected regression whose recovery of known weights we verify. A counterfactual flip decomposition separates modality bias from concept priors, and a Grounding Illusion Score measures how much clean accuracy is not grounded in the image. We evaluate 4 open models on one frozen, seeded evaluation split of 2,000 GIB-200K items. All models are accurate on clean inputs (92.5%-94.2%) but adverse accuracy falls to 41.3%-59.9%. The two LLaVA models over-weight the text (Knowledge Inertia, w_T/w_I = 1.229 and 1.322), whereas InternVL3 favours the image (w_T/w_I = 0.680); between 34.8% and 71.4% of the clean answers that are correct are abandoned when the document lies. A training-free consistency check that compares the answer with and without retrieval trades coverage for accuracy under misleading documents, while falling back to the image fails when the image is heavily degraded. All per-instance outputs, code, and an audit that recomputes every reported open-model number are released.

0.39

ReMoE: An Interpretable Mixture-of-Experts Framework for Human Mobility Recovery Forecasting Across Disasters

cs.LG cs.CY ·2026-11-17 ·Yilin Lyu, Kang Liu, Wei Wang, Qifan Zhang, Ling Yin +1 more
top venue
Abstract · Accepted by KDD 2027

Human mobility often declines sharply following disasters and subsequently exhibits complex recovery dynamics. Effective forecasting of this recovery is critical for timely assessment and resource allocation. However, existing approaches are often developed for individual events and operate as black boxes, limiting their transferability and interpretability beyond the events they were trained on. Recent studies have identified recurring post-disaster mobility recovery types across disasters, suggesting that recovery knowledge may be transferable. Motivated by this, we propose ReMoE, an interpretable mixture-of-experts framework for cross-disaster mobility recovery forecasting. ReMoE uses a feature-guided router to adaptively combine specialized experts that learn these heterogeneous yet reusable recovery types, producing a region-specific recovery pattern. It further refines this pattern through residual correction informed by forecasts of future hazard conditions. We evaluate ReMoE on a large-scale mobility dataset covering 18 disaster events spanning hurricanes, floods, winter storms, and wildfires. ReMoE is competitive in-distribution and achieves the strongest performance on unseen events and hazard types, with larger gains under greater distribution shifts. Further analyses show that ReMoE yields interpretable learning outcomes and maintains robust performance under uncertainty. Broadly, ReMoE offers a transferable and traceable forecaster that can support anticipatory post-disaster assessment.

0.39

When the tail is earned by its companions: measuring label dependence in long-tailed multi-label classification

cs.LG cs.CLcs.IR ·2026-11-17 ·Christoffer Löffler
Abstract · 27 pages, 3 figures, 13 tables. Submitted to Information Sciences

A rare label seldom occurs alone, and a classifier can recover it from the frequent labels it co-occurs with, its companions, without having learned the label itself. Existing diagnostics of this need a trained model: none says before training whether a corpus admits the failure, or what the standard aggregate hides. Macro-F1, like any measure built from per-label confusion counts, is unchanged by how a label's recall divides between occurrences with and without a companion. We introduce a within-label estimand of that division, a reading rule under which a fall in dependence counts only if companion-free recall rises, and a statistic of training labels that bounds, before training, how far reported recall can exceed companion-free recall. Across 23 corpora, 7 training objectives and five families of label-correlation mechanisms, companion-free recall falls short on all six corpora where the estimand is defined. Up to 18.4% of the tail true positives macro-F1 credits would vanish at the companion-free rate, within the training-label bound on every corpus. No long-tail objective reduces the dependence, but architecture moves it. Models whose tail score is computed from other labels' predictions are strongly dependent.

0.39

Fine-Tuning Gemma 4 Edge Models for MATLAB Signal-Processing Code with SFT and QLoRA on a Laptop iGPU

cs.LG cs.AIcs.SEeess.SP ·2026-11-17 ·Mehmet Altan Toksöz
Abstract · 18 pages, 3 figures, 4 tables

Small language models that run on a laptop are attractive for writing signal-processing code offline, but they make systematic errors on digital signal processing (DSP) conventions. We apply supervised fine-tuning (SFT) to the Gemma 4 edge models E2B and E4B with 66 checked examples covering 11 signal-processing MATLAB subtasks, using low-rank adaptation (LoRA) and quantized LoRA (QLoRA) on the integrated GPU (iGPU) of a laptop, and deploy the 8-bit integer (int8) models with the adapters through OpenVINO. On held-out variants scored by deterministic checkers, the pass rate rises from 18/66 to 37/66 (LoRA) and 37/66 (QLoRA) for E2B and from 32/66 to 47/66 for E4B (QLoRA). Reading all 159 failing programs shows that the adapters remove most errors of convention, such as the sign of a lag, units and array shapes, but few errors in building signals and noise. Applied at run time, the adapter costs 17-19% of the generation speed.

0.38

BernBRS: Certified Backward Reachability for Deep Bernstein Neural Networks

cs.LG ·2026-11-17 ·Ibrahim Albool, Yasser Shoukry
Abstract · 18 pages, 6 figures, 7 tables

Ensuring the safety of complex neural network systems requires mathematically rigorous verification. While backward reachability analysis is a critical tool for identifying safe initial states, computing tight, formal bounds remains notoriously difficult to scale. Recent advancements have demonstrated that DeepBern-Nets neural networks utilizing Bernstein polynomial activation functions possess advantageous mathematical properties that enable exceptionally fast and tight forward reachability analysis. Building upon this efficiency, this paper introduces BernBRS, a highly accurate and scalable framework for the backward reachability analysis of DeepBern-Nets. BernBRS simultaneously computes certified inner and outer approximations of the Backward Reachable Set (BRS). Rather than relying on single convex bounding, our method intelligently refines a cover of the input space by analyzing activation patterns, splitting the input domain precisely where an exact decomposition of the relaxation gap predicts the most significant reduction in error. This targeted refinement enables high levels of scalability to highly complex and massive architectures that remain intractable for existing certified methods. We demonstrate the efficacy of our framework on high-dimensional models, including a linearized closed-loop approximation of the Half-Cheetah controller and a distilled Bernstein student of a Transformer block and we empirically study how these model-level guarantees relate to the original Half-Cheetah and TinyBERT classifier, respectively.

0.37

Pareto-Informed Smart "Predict, then Optimize" for Multi-Objective Combinatorial Problems

cs.LG ·2026-11-17 ·Aabhash Dhakal, Tim Lachner
published
Abstract · 8 pages, 5 figures, 1 table. Accepted after peer review at the Multi-Objective Decision Making (MODeM) Workshop at ECAI 2025

Many real-world planning applications, from logistics, routing to resource allocation, give rise to optimization problems that need to balance multiple, conflicting objectives. While machine learning is often used to predict the unknown parameters or costs of these problems, standard models trained on prediction accuracy can lead to suboptimal plans. Decision-Focused Learning (DFL) addresses this by integrating the downstream optimization task into the training objective. However, existing DFL methods for multi-objective problems rely on simple scalarization, which fails to capture the rich trade-offs of the Pareto Frontier. To address this gap, we propose a Pareto-aware extension to single-objective Smart "Predict, then Optimize" (SPO+) for multi-objective combinatorial optimization problems. Our loss function leverages the true Pareto Frontier, forming a target decision that aligns with a sampled preference vector, and minimizes regret against this target. Experiments on multi-objective shortest path and multi-objective knapsack problems show that our approach outperforms traditional two-stage Mean Squared Error (MSE) baselines. It also surpasses scalarization-based SPO+ approaches.

0.37

Distortion-Free Semantic Watermarks for Black-Box Generative Models

cs.LG cs.CRcs.CV ·2026-11-17 ·Aviv Yossef, Or Patashnik, Or Zamir
code/data
Abstract · Project page: https://avivyossef.github.io/distortion-free-semantic-watermark/ Code: https://github.com/avivyossef/distortion-free-semantic-…

We present a provably distortion-free semantic watermarking algorithm that requires only black-box sampling access to a generative model and black-box access to a semantic embedding. A watermark is called distortion-free if, averaged over its secret detection key, it leaves the generator's output distribution unchanged; this provides a formal guarantee that watermarking does not reduce output quality. Conventional such watermarks are often brittle to transformations. This is especially problematic for images, where benign operations such as compression, cropping, or resizing change the entire pixel representation. Semantic watermarks therefore require detection to be robust to content-preserving transformations. Previous methods combining these properties rely heavily on specific white-box properties of the underlying model: either to autoregressive text generation or to diffusion-specific access, such as control of the initial noise or diffusion inversion. In contrast, our construction is fully generic: It only interacts with the generative model and the semantic embedding, a function that determines the notion of "preserving content" under which detection should be robust, via sample and query access. We then instantiate this framework for image generation. We characterize the two competing properties required of the semantic embedding: it should remain stable under transformations that preserve an image's content, but should distinguish independent generations from the same prompt. We train an image encoder directly for this tradeoff. The resulting encoder improves separation while maintaining stability over off-the-shelf alternatives and enables strong aggregate detection while remaining robust to various transformations. We demonstrate the effectiveness of our watermark scheme on both open-source and closed models, which is possible only due to restriction to black-box access.

0.37

Who Pays for Byzantine Robustness? Validity, Data-Share Dictatorships, and Fairness When Only a Few Hospitals Collaborate

cs.LG cs.CR ·2026-11-17 ·Adnan Malkani
benchmarkcode/data
Abstract · 12 pages, 7 figures, 6 tables. Code: https://github.com/adnanmalkani/byzantine-fl-bench

Clinical federated learning (FL) collaborations typically connect only three to six hospitals, yet Byzantine-robust aggregation rules are designed and benchmarked on synthetic partitions across dozens of equal-sized clients. We conduct an empirical audit of eight standard aggregation rules: FedAvg, coordinate-wise median, trimmed mean, Krum, Multi-Krum, Bulyan, geometric median (GeoMed), and Centered Clipping (CC). We evaluate both unweighted and sample-size-weighted variants across three cross-silo medical datasets from the FLamby benchmark: Fed-Heart-Disease ($n = 4$ hospitals), Fed-TCGA-BRCA ($n = 6$ regions), and Fed-ISIC2019 ($n = 6$ centres). We execute 54,214 training runs across enumerated attacker coalitions ($f \in \{1, 2\}$) paired with exact-twin leave-out controls that decompose performance drops into honest data loss versus adversarial attack damage. We report three findings. First, client-count validity checks fail to capture breakdown in skewed medical consortia: while $n = 6, f = 1$ satisfies every count condition except Bulyan's, assigning sample-size weights turns median and GeoMed into a clean-model dictatorship whenever a single hospital holds more than half the training data, and causes weighted median, GeoMed, and CC to collapse when that dominant centre attacks. Second, exact-twin leave-out controls reveal that when the largest hospital attacks, the apparent failure of count-based robust rules is predominantly caused by the loss of honest training data (7.5 to 10.4 points on ISIC) rather than adversarial perturbation (2.3 to 3.8 points on ISIC). Conversely, weight-dependent rules suffer large true attack damage (16.1 to 25.1 points on ISIC). Third, unweighted robust rules consistently discount whichever institution generates the most atypical parameter updates, penalizing the smallest hospital on Heart but penalizing larger institutions on TCGA and ISIC.

0.36

LoopPrune: Recurrent Pruning for Looped Diffusion Transformers

cs.LG ·2026-11-17 ·Weizhong Huang, Huaijia Kou, Yan Zhang, Fei Chao, Xiawu Zheng +1 more
benchmarkcode/data
Abstract · 10 pages, 11 figures

Looped diffusion Transformers reuse shared blocks to deepen computation and match larger models with fewer parameters, reducing storage and deployment burdens. Yet repeated attention and MLP computations remain costly. Pruning reduces storage and computation by removing redundant parameters or units. Training-free methods struggle to resolve recurrent contributions, while training-based methods lack explicit constraints on position-specific errors affecting later inputs and final outputs. We propose LoopPrune, a pruning and recovery framework addressing these requirements. Recurrent importance estimation combines shared-removal effects and position variation to select a shared structure; shared reconstruction pools recurrent input statistics to compensate weights. Loop-position recovery compares complete denoising predictions at identical student states and matches text guidance, updating all student parameters against a fixed dense teacher. Extensive experiments across four open-source models, six benchmarks, and three pruning paradigms demonstrate effectiveness. At 30% structured parameter pruning, Looped-DiT B/16 retains 93.14% of dense six-benchmark average accuracy, exceeding the best baseline by 9.49%. Across PyTorch, vLLM-Omni, and SGLang Diffusion, pruned SDXSv3 reduces peak allocated GPU memory by 34.36-44.64% and accelerates denoising 1.46-1.74x over dense models.

0.36

NVIDIA Kumo Tabular: Pushing the Accuracy$\unicode{8211}$Efficiency Frontier for Tabular Prediction

cs.LG ·2026-11-17 ·Jingang Qu, Valter Hudovernik, Martin Jurkovič, Cedric Lorenz, David Holzmüller +10 more
code/data
Abstract · 25 pages, 13 figures

We introduce Kumo Tabular, a family of open tabular foundation models for classification and regression, in three sizes: Kumo Tabular-S (27M parameters), Kumo Tabular-M (61M), and Kumo Tabular-L (214M). Pretrained only on artificial tables from the Kumo prior, they can make predictions on new tables without any training. In our approach, a table encoder compresses each row into an embedding, and an in-context learning transformer predicts the unlabeled rows from the labeled ones. Kumo Tabular introduces architectural innovations that improve both training stability and predictive performance, enabling the model to scale up to 150M cells on modern GPUs. Kumo Tabular-L ranks first on eight benchmarks: TabArena, TALENT, BeyondArena, ScoringBench, STRABLE, MulTaBench, RamanBench, and TabBench-Bio. These cover probabilistic regression, tables with over 1M rows or 100K columns, and tables with strings, text, or images. Accuracy improves with model size in all tested comparisons. All three sizes lie on the accuracy-speed frontiers of TabArena and BeyondArena. Kumo Tabular-L predicts faster than the runners-up, LimiX-2 on TabArena (36 times) and TabPFN-3.5 on BeyondArena (1.7 times). Weights and pretraining code are released under OpenMDW-1.1 and Apache 2.0, respectively.

0.35

Symmetries of Counterfactuals: Identification and Testability of Counterfactual Maps from Covariate Heterogeneity

cs.LG stat.ML ·2026-11-17 ·Marc Braun, Adel Daoud, Jose M. Peña
code/data
Abstract

In general, counterfactuals are not identified from data alone, because many causal models produce the same observed distributions but different counterfactuals. A common approach to circumvent this problem is to restrict the mechanism to a class in which it becomes unique, such as triangular monotone maps or optimal transport maps. However, whether the assumed class is correct is usually not examined. In this work, we study what the data can tell us when covariates are available that influence the outcome only through its latent cause. First, we characterise all counterfactual maps that are consistent with the data. They differ from the true map by a group of symmetries that preserve the conditional outcome distributions given the covariates. The more these distributions vary across covariate values, the smaller the group. Second, we show that for the classes above, consistency with the data is equivalent to a single conditional independence, which yields a test of the assumed class. We illustrate both results on synthetic data and on the ACTG 175 HIV trial.

0.34

Towards Optimism-Pessimism Trade-off in Model-based Offline-to-Online Reinforcement Learning

cs.LG ·2026-11-17 ·Guochen Zhou, Yijun Yang, Qiqi Duan, Qing Su, Weiming Ou +4 more
top venue
Abstract · Accepted at NeurIPS 2026

Model-based offline-to-online reinforcement learning (RL) enables sample-efficient adaptation by leveraging offline pre-training for online fine-tuning. However, the distribution shifts between offline and online stages often hinder fine-tuning performance. Many existing methods encode the optimism-pessimism trade-off through a single-objective formulation. Selecting its trade-off coefficient through online fine-tuning can require costly bi-level optimization. We identify the balance established during offline training as a key challenge: in our motivating experiments, optimistic policies adapt better to novel tasks, whereas pessimistic policies perform better on the same online tasks. To address this challenge, we propose a bi-objective formulation that captures this trade-off, yielding a Pareto policy pool during offline training. This policy pool enables adaptive selection during online fine-tuning across diverse online tasks. To generate the pool, we introduce Multi-Objective Soft Actor-critIC (MOSAIC), which solves bi-objective problems and constructs diverse Pareto policies. After offline training, a contextual bandit algorithm hierarchically selects the most suitable policy for fine-tuning at each online interaction step. Empirically, our pipeline, Hierarchical Pareto Policy Pool (HiP3), achieves the strongest aggregate performance across both same and novel online tasks and maintains its advantage under stronger distribution shifts. Comprehensive ablations further clarify the roles of the policy pool, MOSAIC-based pool construction, and the online selection mechanism.

0.34

Path-Aligned Constraint Equivariance

cs.LG ·2026-11-17 ·Yuhao Wang, Yuanxi Peng
Abstract

Learned super-resolution of physical fields must satisfy structural constraints---such as divergence-free flow ($\nabla\cdot u=0$), curl-free fields ($\nabla\times u=0$), and pointwise sum-to-one normalization ($1^\top p=1$)---that standard training objectives do not enforce. Soft penalties trade fidelity against satisfaction via a scalar hyperparameter $λ$ but cannot drive the residual to zero; hard-constrained regression eliminates the residual by construction yet converges to the conditional mean, losing small-scale spectral structure. We observe that in Conditional Flow Matching with linear paths, a linear constraint $Ax=c$ is preserved along the entire transport trajectory whenever the velocity field is compatible ($Av_θ=0$) and the initial noise is consistent ($Ax_0=c$); along the trajectory, the latter two conditions are also necessary. We term this the path-equivariance principle and instantiate it in PACE (Path-Aligned Constraint Equivariance), a framework that enforces these conditions through a null-space parameterization of the velocity field and a constraint-respecting noise sampler, with no modification to the loss function or the ODE integrator. Across three constraint families, PACE reduces the mean constraint residual by one to five orders of magnitude relative to unconstrained baselines, matching the residual level of the data-generation pipeline itself. The proposed framework requires only a closed-form parameterization onto the null space of $A$; within this scope, it offers exact constraint enforcement while recovering the spectral content and predictive diversity that conditional-mean regression systematically loses, all under a single unified recipe.

0.33

Perception into Imagination: Executable Memory for World Models

cs.LG cs.AIcs.CVcs.RO ·2026-11-17 ·Xianyao Li, Ruitong Tian, Rui Min, Fang Xu, Jing Du
hot direction
Abstract · 18 pages, 6 figures, 15 tables

Physical agents must use what they previously saw or touched after visibility or contact is lost. Their actions change the implications of these remembered facts. We introduce executable perceptual memory (EPM), a world-model framework that advances explicit perceptual state alongside neural latent dynamics during imagination. Its evolving read-out informs reward prediction, value estimation and policy learning. We align reward inputs with the transitions that produced them and supervise the reward head on open-loop rollouts of recorded actions. On Memory Maze $13\times13$, EPM reaches 30.37 return at 1M interactions, compared with 23.37 when the same evolving read-out informs only policy and value on a common perceptual front end: a gain of 7.00 [3.80, 10.21]. The complete method also improves control on $9\times9$ and in MiniWorld. Probes of remembered, out-of-view arrivals show stronger reward anticipation in both visual implementations. Concentration, tactile recovery and real tactile frames demonstrate complementary roles of remembered facts in imagined-policy training, motion-dependent decisions and sensory-content retention. EPM connects persistent perception to the predicted consequences and learning signals of a world model for Physical AI.

0.33

SilverSpoon-KD: A General-Purpose Toolkit for Knowledge Distillation

cs.LG ·2026-11-17 ·Xaver R. Davey, David Broman
code/data
Abstract · 28 pages, 10 figures, 11 tables. Code: https://github.com/silverspoon-dev/silverspoon-kd

Knowledge distillation transfers deep learning representations from a teacher model to a student model, typically reducing model size or inference cost. However, knowledge distillation tooling has not kept pace with the breadth of research methods: existing toolkits typically support only a limited subset of techniques and use cases. We present SilverSpoon-KD, an open-source PyTorch toolkit that unifies different distillation paradigms and architectural relationships across workloads and applications behind a single API. Moreover, SilverSpoon-KD offers the first general-purpose blockwise knowledge distillation implementation with unique scalability features. We validate its versatility across vision, masked language model, and causal language models on tasks spanning homogeneous, heterogeneous, and quantization distillation across paradigms. We also evaluate SilverSpoon-KD by comparing its performance, including throughput and GPU memory usage, with existing toolkits. Despite extensive support for convenience, metric calculation, and tracking, we have found SilverSpoon-KD's performance to be on par with or better than existing toolkits.

0.33

COTTI: Contrastive Training and Time-Aware Inference for Robust Industrial Recommendation under Promotional Drift

cs.AI cs.LG ·2026-11-17 ·Maolei Huang, Jing Tian, Dexuan Duan, Fangfang Lin, Huawei Cao +1 more
Abstract

Promotional events shift traffic composition, purchase behavior, and the scale of conversion probabilities on different time horizons. We present \textbf{COTTI} (\textbf{CO}ntrastive \textbf{T}raining and \textbf{T}ime-aware \textbf{I}nference), an industrial recommendation framework that separates historical representation learning from event-time score adjustment. Cross-Period Contrastive Alignment (CPC) regularizes paired histories of the same user with adaptive consistency weights. Active Re-weighting (RW) controls the influence of promotional examples through historical-interest distance. Adaptive Boundary Alignment (ABA) combines stochastic prediction variability with a temporal event kernel to adjust logits while keeping predictor parameters fixed. We analyze the weighted objective and bound ABA-induced probability and ranking changes. On Fliggy data, COTTI attains an offline AUC of 0.8127, compared with 0.8101 for TEN and 0.8123 for TEN with ABA. Cross-category analysis reveals improved discrimination alongside declines in paid-item recall for novel categories. End-to-end measurements give a p99 latency of 38.9\,ms and throughput of 14,700 requests/s; enabling ABA adds 2.2\,ms to p99. A ten-day online experiment against the serving HDR model yields relative gains of 3.14\% in conversion rate and 2.68\% in revenue per mille. Together, these results characterize the trade-off between aggregate prediction, novel-category discovery, and serving cost in promotional recommendation.

0.33

A Transformer-Derived Iterative Preconditioner

cs.LG ·2026-11-17 ·Patrick Lutz, Themistoklis Haris, Aditya Gangrade, Venkatesh Saligrama
top venuetheory
Abstract · Accepted at NeurIPS 2026

Transformers trained on in-context prediction problems can learn known numerical algorithms, e.g., gradient descent. We go further, showing that their internal computations can reveal new algorithms with provable guarantees. A softmax-attention transformer trained to solve linear systems $Xw=y$ implicitly improves matrix conditioning. Tracing $XX^\top$ reveals an iterative preconditioning primitive, which we develop into RAPID, a standalone algorithm requiring no learned components. RAPID is an anytime preconditioner for symmetric positive definite (SPD) matrices, using randomized mixing and sparse local corrections to improve conditioning. Proof-of-concept experiments on synthetic and real-world problems consistently improve conditioning and reduce conjugate gradient iterations needed to reach a target accuracy. For a Haar-mixing idealization, we prove convergence to any fixed target condition number regardless of spectral distribution, with only logarithmic dependence on initial conditioning. Together, these results show how a transformer can discover a broadly applicable numerical algorithm.

0.32

A single inference process underlies the behavior and neural circuits of working memory

cs.LG q-bio.NC ·2026-11-17 ·Dongyu Gong, Mario Belledonne, Ilker Yildirim
Abstract · 39 pages

Working memory holds past inputs online for future behavior. Despite decades of research, accounts of working memory remain divided across levels of analysis, with long-standing debates on whether its capacity reflects discrete slots or continuous resources, how attention improves recall, and whether its representations rely on persistent neural firing or activity-silent states. Here we show that MemGP, a theory that is specified at the computational, algorithmic, and circuit levels of analysis simultaneously, can explain diverse behavioral and neural phenomena in working memory. At the computational level, MemGP casts working memory as probabilistic inference over a continuous function on feature space, with a variational objective that balances memory strength against a Gaussian-process prior. At the algorithmic level, writing, storing, and reading of memories correspond to approximating this posterior with a finite set of inducing points, the resource that all items share. In simulations, MemGP reproduces load-dependent precision loss, attentional benefits during maintenance, and inter-item attraction and repulsion, in each case revealing multiple complex computational dependencies. At the circuit level, the algorithm analytically (without training) maps onto a recurrent circuit, including its architecture and dynamics. Beyond reproducing neural phenomena such as delay-dependent drift, this analytical map deduces center--surround connectivity, a property long assumed in attractor models, and a structured combination of persistent activity and synaptic weights to hold memories. Thus, a single inference objective not only jointly predicts the behavioral and neural signatures of working memory, but also addresses, across several instances and different levels of analysis, existing competing explanations.

0.32

Lost in the Funnel: Selection-Induced Miscalibration in Ranking and Decision Systems

cs.LG ·2026-11-17 ·Aditya Somani, Rahul Ragesh, Vikas Joshi, Amit Singh
Abstract · Under Review

Modern ranking and decision systems combine components with divergent objectives, and scores calibrated at the model boundary are used as calibrated probabilities throughout the funnel, across filtering, reranking, business-rule enforcement, and auctions. Yet calibration is typically validated only where the score is produced, so its breakdown where the score is used is routinely overlooked. We ask when calibration survives this selection, how inherited error differs from selection effects, and what can fail for a perfectly calibrated score at a component boundary. We study these questions through a boundary-wise diagnostic framework of residual identities. We target rubric-judged relevance rather than presentation-dependent metrics such as CTR, so calibration is measurable at every funnel layer of interest. We first present an industry-wide case study that applies our diagnosis to settings where calibration deteriorates at decision time, tracing the degradation to selection rather than model error. This diagnosis motivates Funnel-Aware Recalibration (FAR): post-hoc, selection-policy-aware recalibration at the boundary where scores drive decisions. In a live two-region production A/B test, FAR cuts decision-time expected calibration error (ECE) by 60-80% and improves auction efficiency. We further adapt two public datasets to simulate industry settings, providing additional evidence for the mechanism. We plan to release the adapted datasets and their generation pipeline to support further research. Finally, we distill these findings into practical operating principles for calibrating multi-stage scoring systems.

0.32

ConfGuard: a physics-grounded reference monitor for predictive-maintenance large-language-model agents

cs.AI cs.CRcs.LG ·2026-11-17 ·Arunkumar V, Gangadharan G. R., S. Senthilkumar
Abstract · 24 pages, 5 figures

Large-language-model agents now close the industrial predictive-maintenance loop, reading sensor windows, retrieving procedures and dispatching work orders. Two attack channels meet there: injected text in a retrieved document can hijack the decision, and a bounded sensor perturbation can corrupt the remaining-useful-life estimate behind it. Recent agent defenses police data provenance and tool privilege but cannot tell whether a prognostic quantity is physically plausible. ConfGuard is a reference monitor with a physics-grounded policy. Its conformal monitor tests each prediction, anchored to the asset's admitted history, against degradation physics: capped remaining life never rises and falls at most one cycle per cycle. Calibrating its threshold over whole asset trajectories lets the operator set the expected share of a clean asset's windows the running monitor escalates. A physics-policy validator checks the agent's maintenance-urgency decision against the decision a deterministic policy derives from the same interval. On all 10,196 C-MAPSS test windows the monitor cuts a covert attacker's reach on healthy assets from 0.529 to 0.046 and closes suppression of failing assets (0.151 to 0), escalating 0.027 of clean windows against a 0.05 target, while the fixed physical range it replaces leaves both attacks open. Behind a Qwen3-14B planner the composed monitors hold urgency-class injection to 0.00 and the universal sensor attack to 0.12, against 0.29 and 0.54 for a defended prompt. An adaptive attacker defeats a Mahalanobis detector on 0.504 of windows and adversarial training on 0.485. We report an attack class the projection misses.

0.31

Reversing the text is not reversing the path: a fixed-text measure without a zero, and an assembly-order comparison that vanishes identically

cs.LG physics.soc-phcond-mat.stat-mech ·2026-11-17 ·Y. -H. Taguchi
theoryunusual
Abstract · 41 pages. Ancillary files contain the measurement and verification scripts. Prepared with AI assistance (Anthropic Claude for drafting, revi…

For an ensemble of trajectories the fluctuation theorems give the entropy production its meaning and zero. The same expression has recently been evaluated on one fixed text, $Δ= [\log P(x) - \log P(\tilde x)]/L$ with $\tilde x$ the text in reversed order, read as an entropy production separating causally from arbitrarily ordered prose. A single text is neither a distribution nor a dynamics, and the texts at issue were produced by other models. No theorem fixes where the zero of a fixed-text quantity lies: the expression carries over from an ensemble, its zero does not. We supply a zero from an exact antisymmetry. Scoring a text's $n$ sentences in all $n!$ orders, reversal carries $π$ to $π\circ \mathrm{rev}$, a measure-preserving involution, so $Δ(π) = -Δ(π\circ \mathrm{rev})$ identically and the values are centred at zero for any score. The authored arrangement's midrank is then unitless, bounded, and 0.50 for a score blind to order. Across five autoregressive evaluators from 124M to 4B parameters, and taking the prompt-defined labels at face value, 8% to 15% of the median midrank's distance from the null is attributable to causal order and the rest to the arrangement being the order the passage was produced in. The thermodynamic reading needs a comparison whose zero does hold on the text read, and a masked model supplies one: the any-order chain rule gives a genuine log-probability, and its left-to-right and right-to-left orders for one text give $σ$, a comparison of two assembly orders. Both run from the empty sequence to the same text, so it is not a path against its time reverse. It vanishes identically when the evaluator's conditionals come from a joint, and a non-zero reading reports the evaluator's incompatibility, not the text. Neither is a fixed-text entropy production, for different reasons.

0.31

Proximal Targets: Training Looped Refiners on Problems with Many Correct Answers

cs.AI cs.LG ·2026-11-17 ·Syed Affan Daimi
benchmarktheory
Abstract

Looped refiners such as HRM and TRM apply one small network repeatedly to a draft answer, and deep supervision trains every pass toward one stored answer per puzzle. We prove that when a puzzle has many correct answers this target sets a loss floor that no model capacity removes, and that the loss-minimising pass blends the answers into an invalid grid. We propose the proximal target: train each pass toward the correct answer nearest to its own input, found by scanning a list of answers, by a search solver started at the draft, or within a band of quality set by a knob; correct answers become fixed points, and a pass only has to make fewer errors than its input has for the loop to converge, which its training loss bounds. With the same 1.7M-parameter network and 5 seeds, the proximal target solves 98.8% of Sudoku puzzles with 2-64 answers (standard target: 2.5%), 93.8% of sparse puzzles with more than 10,000 answers without any list of answers (0.0%), and 90.9% of 200-node 3-colourings after training on graphs of at most 50 nodes (36.9%), and different random starts end at different answers. Against GRAM, a stochastic-latent reasoner, a single run is correct far more often (98.8% vs. 57.7%) and twenty runs find more of the answers (53.6% vs. 37.4%), while GRAM is slightly ahead on its own 10-node colouring benchmark; in Scrabble the knob becomes a calibrated strength dial.

0.31

A Neural Envelope Method for Rapid Gradient-based Optimum Takeoff Trajectory Design

cs.LG ·2026-11-17 ·Samuel Sisk, Xiaosong Du
Abstract · 16 pages, 12 figures, submitted to AIAA Aviation conference

Urban air mobility arises to mitigate ground traffic congestion. Multidisciplinary analysis and optimization (MDAO) has been introduced to reduce the energy consumption, thereby expanding aircraft performance. Surrogate-based MDAO lowers the cost of optimization, but multimodality disrupts global optimum search for both simulation-based and surrogate-based optimizations. Convex envelopes convert multimodal spaces into unimodal spaces, permitting efficient convex optimization; however, convexifying a design space may produce infeasible "optimal" designs due to complex design constraints. To address these challenges, this work proposes the coupling of a physics-constrained generative adversarial networks (physicsGAN) and neural envelope to transform original, constrained, non-convex optimizations into unconstrained, convex optimizations. Specifically, the physicsGAN discovers a feasible design space where all design candidates inherently satisfy all constraints, by penalizing surrogate-predicted constraint violations during training. Moreover, the neural envelope constructs a neural network-based convex envelope for the transformed objective function within the physicsGAN-discovered feasible space. Note that the neural envelope only needs to deal with objective function since constraints are inherently satisfied in the feasible space. The neural envelope is demonstrated on takeoff trajectory optimizations of the Airbus A3 Vahana eVTOL aircraft. Results show that neural envelopes achieve a 99.91% mean accuracy within a flight-condition space compared with reference convex envelopes. Neural-envelope-enabled optimizations achieve 99.84% accuracy on the global optimum design, costing only 0.04 seconds on average. The neural envelope accelerates the trajectory optimization by 53 and 162 times on average compared to the physicsGAN-enabled gradient-based and global optimizations, respectively.

0.30

Where Should Physics Enter? Information, Identifiability and Calibration in Probabilistic Sea Surface Temperature Forecasting

cs.LG ·2026-11-17 ·Parmitha Vangapandu, Jonnakuti Pavan Kumar, Sathwik Narkedimilli, MSVPJ Sathvik, T. M. Balakrishnan Nair +1 more
benchmark
Abstract

Advection-diffusion priors are a common way to make neural sea surface temperature (SST) forecasts physics-informed. We test where such a prior must enter a heteroscedastic ConvLSTM to help, using daily NOAA OISST v2.1 for the Bay of Bengal (1982-2024) and a synthetic benchmark on the same grid with known currents of controlled strength. (i) On real data, a soft advection-diffusion residual with a learned velocity gives no reliable benefit (lead-15 CRPS -0.007 degrees C, 95% interval [-0.022, +0.008]); a plain temporal-smoothness penalty matches it, and equivalence within 3% is not established either. (ii) Learned velocities do not recover the currents, whether placed in the loss, in an explicit transport step, or in a transport step without neural correction: they correlate with the reference no better than a shuffled-date null, even on synthetic data where knowing the currents would improve day-15 forecasts by 16-23%. (iii) Supplied currents help on synthetic data, but mostly as information: giving them to the network as input channels, without transport, recovers 92% of the transport design's lead-5 gain, and a frozen-current numerical forecast does as well as either. (iv) With GLORYS12 reanalysis currents on real data, the transport design is 6% and 4% worse at leads 1 and 5; reanalysis advection explains essentially none of the daily OISST change (correlation 0.02), and frozen-current transport is worse than persistence. (v) The defect that matters is lead-dependent miscalibration shared by every model. One variance multiplier per lead, fitted on validation data, brings 90% coverage to 0.85-0.91 at all leads and lowers day-1 CRPS by 13-21%; fine-tuning on 15-day rollouts fixes long-lead coverage but costs short-lead skill. This version withdraws an earlier claim of an 81% CRPS reduction that came from an uncontrolled comparison.

0.30

VTT: Reading Vehicle Size and Powertrain Type from Motion Trajectory Across Unseen Car Models

cs.LG cs.RO ·2026-11-17 ·Yuhang Wang, Yuechen Guo, Jingxin Yang, Chuheng Wei, Suyang Chen +5 more
benchmarkdataset
Abstract · 8 pages

Vehicle size and powertrain information can enrich trajectory data for traffic analysis and simulation. We present VTT, a dataset and benchmark for identifying these attributes from motion and controller area network (CAN) signals across unseen car models. The dataset contains 9,062 drives, about 3,950 h, from 1,431 vehicles and 174 car models. It provides three footprint-based size classes, powertrain labels, and evaluation on unseen car models and unseen drivers. Using 60-s windows, we compare four sequence architectures, statistical features and physical features proposed by large language models. Training with auxiliary physical targets, data augmentation and task-specific weighting raises drive-level macro-F1 to 92.6-93.0% for battery-electric versus combustion vehicles and from 62.9% to 67.1% for size classification. On public logs from 4,061 additional drivers, the full-input model identifies powertrain correctly for 86 of 88 car models; size accuracy is 63.2%, close to the internal vehicle-level result of 63.5%. Physical-feature analysis links powertrain recognition to deceleration behaviour and size recognition to steering response. Errors near size-class boundaries and limited gains from longer observations highlight the importance of label resolution and training-fleet coverage for cross-model generalisation.

0.30

Spatially local mini-batching for implicit neural representations

cs.LG ·2026-11-17 ·Minseok Choi, Swarup Dhar, Heejae Kwon, Kookjin Lee, Donsub Rim
Abstract · 72 pages, 36 figures

Neural surrogates based on implicit neural representations (INRs) are typically trained on mini-batches spread over the whole domain, since a generic INR loses accuracy when every point of a batch lies in one small region. We show that low-rank neural representations (LRNRs), INRs whose weight matrices are products of shared bases with coefficients emitted by a small hypernetwork, can be trained with spatially local mini-batches. Each step uses the points of one time snapshot of an adaptive-mesh numerical solution inside one small cell, with snapshots and cells drawn according to the refinement of the mesh. We posit that such training is enabled by the structural property of LRNRs, as a local step reaches the shared bases only through the $r$ coefficients of each layer, and keeps one cell from rewriting the network. By a Kolmogorov-width argument, we show that when restricted to a small cell, the hidden states of LRNRs lie close to a low-dimensional subspace, which suggests the local steps only affect low-rank subspaces of the shared bases. On six hyperbolic-conservation-law problems with solutions computed using adaptive mesh refinement, one local training recipe reaches 1.1-2.3% relative $L^2$ error with a tenth of the parameters of a dense network of the same width. Finally, we provide a comparison by training a hypernetwork-SIREN, a DeepONet, and a Fourier neural operator (FNO) of the same size with the same local batches and training budget. We observe that these baselines reach 16-78% where the LRNR reaches 4.9%. Interestingly, we also observe that local training has a significant effect on the per-cell low-dimensional structure even in a wide generic hypernetwork-SIREN.

0.30

Multistage Contextual Optimization for Vehicle Repositioning

cs.AI cs.LG ·2026-11-17 ·Jinyu Zhang, Di Huang, Zhiyuan Liu, Nikolas Geroliminis
theory
Abstract · 44 pages, 8 figures

Ride-hailing platforms reposition idle vehicles to improve spatial supply-demand balances. This problem is prediction-driven and multistage: at each stage, the platform forecasts future demand, executes current decisions, and uses the resulting state to initialize the next stage. We formulate a multistage linear program (LP) with a macroscopic fundamental diagram (MFD) linking accumulation to trip completion. We then develop a multistage contextual optimization approach that uses decision performance to train the prediction model. The learning problem combines multiple coupled stages, context that depends on earlier decisions, and predictions that affect the objective, the constraint matrix, and the right-hand side of the LP. We derive the exact gradient, which has three parts: parameters change forecasts and hence decisions; earlier decisions change the current decision through the inherited state; and this state also changes forecasts as context. An equivalent adjoint uses reverse vector-Jacobian products, avoiding full solution and state-parameter Jacobians. Finally, forward propagation solves many stage LPs one after another, which makes training slow. We propose a parallel block evaluation algorithm that solves blocks of consecutive stages in parallel and repeatedly feeds each block output to the next block. We prove that the algorithm recovers the serial trajectory within a finite number of passes and that its gradient error remains bounded when it stops early. On New York City ride-hailing data, multistage contextual optimization yields 7.7% to 9.1% more business benefit than the same predictor trained by mean squared error, outperforms all compared baselines from five paradigms, and even outperforms realized demand fed to the same policy. With eight workers, parallel block evaluation shortens training by a factor of 1.86 without loss of decision performance.

0.30

Learning to Undo: Transfer Reinforcement Learning under State Space Transformations

cs.LG cs.AI ·2026-11-17 ·Mridul Mahajan, Aldo Pacchiano, Xuezhou Zhang
published
Abstract · NeurIPS'26 paper

Transfer learning in reinforcement learning (RL) has shown strong empirical success. In this work, we take a more principled perspective by studying when and how transferring knowledge between MDPs (a source and a target) can be provably beneficial. Specifically, we consider the case where there exists an undo map such that applying this map to the target's state space recovers the source exactly. We propose an algorithm that learns this map by matching state feature statistics gathered from both MDPs, and then uses it to transfer the source policy to the target. We theoretically justify the algorithm by analyzing the setting when the undo map is linear and the source is linearly-$Q^\star$ realizable, where our approach has strictly better sample complexity than tabula rasa RL in the target MDP. Empirically, we demonstrate that these benefits extend beyond this regime: on challenging continuous control tasks and Atari games, our method achieves significantly better sample efficiency. Overall, our results highlight how shared structure between tasks can be leveraged for efficient transfer of policies across environments.

0.30

What Does Weight-Space Programming Add to Diffusion Editing? Program Construction versus Score-Guided Refinement

cs.LG cs.CV ·2026-11-17 ·Zitian Gao, Jingyi Xu, Ziqi Han, Chengkai Lei, Ye Liu +1 more
hot direction
Abstract · 25 pages, 7 figures, 13 tables

Weight-space programming (WSP) lets a language model construct executable weight updates for another network. We study diffusion concept editing through an interface in which programs define training contexts, preservation constraints and update directions, while a shared numerical fitter selects their coefficients. We compare generated programs with published editors, a shared-panel hand-written baseline and budget-matched search over conventional edits on 24 TIMED test edits for Stable Diffusion 1.4. The preregistered rewrite pipeline shows no advantage over the strong baseline. In an exploratory comparison, one-shot Astra programs improve fresh-scene VLM joint success, which requires subject presence together with whole-image target matching, by +10.0 points over the shared-panel baseline, while CLIP relative match shows no detectable change. These programs organise adjustable directions around protected text-state subspaces. Probing and rewriting the Astra programs then raises CLIP by +6.2 points while joint success changes by -9.7. On a fixed 144-image sample, five additional evaluators give point estimates in the same directions under whole-image target matching; checks of subject-local change give less decisive differences. Together, these results identify useful behaviour in complete generated programs and show that score-guided revision need not preserve it: program construction and revision should be evaluated as distinct stages of a WSP pipeline.

0.29

Synthetic Crystal Data Generation for Diffusion Pretraining in Material Property Prediction

cs.LG cond-mat.mtrl-sci ·2026-11-17 ·Aamod Varma, Shuyi Jia, Victor Fung
published
Abstract · 14 pages, 2 figures, 6 tables. Accepted at the AI4Mat Workshop at NeurIPS 2026

Diffusion-based pretraining has emerged as a powerful strategy for learning transferable representations in graph neural networks for materials property prediction without requiring quantum-mechanical labels such as energies, forces, and stresses for every pretraining structure. However, the method retains a critical dependency: it requires structures residing near local minima of a potential energy surface, which have traditionally been obtained through costly DFT relaxations of known or physically synthesized compounds. This reliance constrains both the scale and the structural diversity of available pretraining data, leaving large regions of crystal structure space unexplored. In this work, we connect diffusion-based pretraining with synthetic structure generation to reduce this bottleneck. We introduce a synthetic data generation pipeline combining three complementary strategies: perturb-and-relax, generative sampling via MatterGen, and protostructure enumeration. Where structural relaxation is required, we use fast machine-learned interatomic potentials instead of performing a new DFT relaxation for each generated structure. We further compare a coordinate-denoising control with continuous-time variance-preserving (VP) and sub-VP diffusion objectives using a randomly initialized Orb-v2 backbone. The source comparison covers seven MatBench property-prediction tasks, and an eighth task, \texttt{mp\_gap}, is included in the protostructure scaling study. Every synthetic-pretraining condition numerically outperforms random initialization on all seven source-comparison tasks. Scaling the protostructure corpus from 20k to 165k under a fixed 400-epoch schedule numerically matches or exceeds MP-20k pretraining on most tasks; because this protocol increases both data diversity and optimization updates, the result reflects joint data-and-compute scaling.

0.29

CULypso: A Clustered Federated Unlearning paradigm for Vehicular Ad-Hoc Networks

cs.LG cs.AIcs.NI ·2026-11-17 ·Eleftheria Chini, George Margaritis, Dimitrios Katsaros
Abstract · 12 pages, 14 figures

Unlearning has emerged as an essential security and privacy-preserving solution in Machine Learning, enabling networks to support their clients' rights for data erasure, as mandated by data protection regulations. Introducing such a non-trivial task to an edge, highly mobile and time-sensitive vehicular network is an even greater, largely unexplored challenge. This article argues that a hierarchical Federated Unlearning approach could be aptly integrated into vehicular ad-hoc networks (VANETs), leveraging topology and movement to reduce communication, resource and time costs, while effectively forgetting a client's influence on the global model updates. In order to achieve this, we propose CULypso, an approximate, surgical, hierarchical unlearning algorithm tailored for clustered VANETs. Inspired by the nymph in Greek mythology, Calypso, whose name translates to 'she who conceals the knowledge', CULypso, with the capital letters referencing \textbf{C}lustered \textbf{U}n\textbf{L}earning, aims to achieve exactly what its name suggests: conceal, erase or forget in our case, knowledge it contains. Extensive experiments uphold the potency of our paradigm with it performing approximately 65\% better than FedEraser in terms of time, server storage load, and unlearning fidelity while utterly outperforming baseline full retraining.

0.28

Beyond Attribute-Centric Fairness: Modelling and Mitigating Unified Bias Representation in Omni-modal Large Language Models

cs.LG ·2026-11-17 ·Runqi Lin, Suqin Yuan, Zihao Fu, Chris Russell
hot direction
Abstract

Omni-modal large language models (OLLMs) provide joint multimodal understanding and generation capabilities, making them well-suited to complex environments and promising backbones for next-generation AI. Despite these advances, the harmful stereotypes encoded in such models manifest in increasingly diverse forms, leading to broader and more severe societal consequences. Existing debiasing methods typically impose separate fairness objectives for each sensitive attribute and modality to minimise demographic-dependent variations. However, in omni-modal settings, this attribute-centric paradigm struggles to jointly address diverse social biases with distinct fairness objectives and to generalise beyond predefined attributes. To this end, we propose OmniDebias, one of the first methods for promoting fairness in OLLMs, which employs a bias-centric formulation to extract and mitigate a shared bias representation across attributes and modalities while preserving attribute-related knowledge. Specifically, we exploit activation inconsistencies of a given attribute (e.g., gender) across semantic contexts (e.g., occupations) to disentangle bias from legitimate knowledge. We further consolidate multiple extracted biases into a shared representation, formulating a unified objective that captures the underlying structure across social biases and generalises to unseen attributes. Extensive experiments demonstrate that eliminating unified bias with both post-hoc editing and fine-tuning can consistently improve fairness across diverse attributes and modalities in OLLMs while maintaining favourable utility.

0.28

PCGRL for Complex Platformers: Synthesizing Playable and Diverse Levels via Inflated Envelopes and Forward Curricula

cs.LG cs.AI ·2026-11-17 ·Di Wu
Abstract · 9 pages, 6 figures, 1 table

Procedural Content Generation via Reinforcement Learning (PCGRL) has emerged as a promising paradigm for automated game development. However, existing PCGRL frameworks still suffer from severe exploration bottlenecks and scaling limitations when faced with long-sequence generation tasks governed by strict rigid-body rules and non-linear kinematic boundaries (e.g., platformer constraints). Consequently, traditional approaches frequently fail to maintain reliable level solvability on large-scale canvases. To address these challenges, we present a physics-aware, curriculum-guided PCGRL framework. First, we construct an "Inflated Volume Envelope" mechanism, which provides a physics-constrained representation that incorporates continuous kinematic feasibility into the discrete state space of the MDP. Second, to alleviate the sparse reward challenge induced by extremely long decision sequences, we introduce intermediate state-transition feedback and a multi-stage forward extension curriculum learning framework, enabling informative value estimation and stable credit assignment over long-horizon trajectories. Finally, to resolve the issue where deterministic policies after convergence generate highly monotonous map paths with a lack of diversity, we activate high-entropy stochastic inference during the synthesis phase, regularized by a post-hoc parallel rejection filtering mechanism. Experimental results demonstrate that our framework significantly enhances structural diversity while maintaining high physical playability across multi-scale environments. Comparative evaluations across multi-scale canvases demonstrate that our framework maintains superior convergence stability and dimensional scalability, reliably synthesizing solvable levels where traditional PCGRL baselines degrade.

0.28

Efficient Machine Learning Interatomic Potentials

cs.LG cond-mat.mtrl-sci ·2026-11-17 ·Yury Lysogorskiy, Anton Bochkarev, Ralf Drautz
benchmark
Abstract

Machine learning interatomic potentials (MLIPs) reach an accuracy close to quantum mechanical reference calculations at a small fraction of the cost. Efficiency is the purpose of MLIPs, yet strategies for efficient MLIPs have not been reviewed systematically. This review takes efficiency as its guideline. To compare MLIPs on a common footing, we represent them in a complete basis expansion, the Atomic Cluster Expansion (ACE) and its graph generalization (GRACE). The expansion is canonical, built from products of single-particle functions. Classical interatomic potentials, such as the embedded-atom method, and many modern MLIPs, including equivariant transformers, can be represented within this expansion. Much of the apparent diversity of MLIP architectures therefore reflects how non-linearity, layer coupling and optimization are handled rather than which interactions the models can represent. We show how the cost of an MLIP becomes linear in system size and, at fixed rank of the tensor decomposition, in model complexity, and we discuss the role of machine learning, which provides non-linear parameterizations and optimization. Because efficiency is meaningful only at a given accuracy, we also compare accuracy and cost. On the Matbench Discovery leaderboard, models with three to five layers approach the accuracy of deeper models, and one basis, truncated along the Pareto front of error and speed, spans more than two orders of magnitude in cost. Ultimately, training data determine the quality of any MLIP.

0.27

Do LLM Agents Still Need to Be Told How to Self-Improve?

cs.LG cs.AI ·2026-11-17 ·Shobhita Sundaram, Cansu Sancaktar, Phillip Isola
code/data
Abstract · Project page: https://ssundaram21.github.io/self-org/

LLM agents increasingly drive their own learning. Self-improvement methods delegate various components of model optimization to the models themselves. As models become more capable, this motivates a basic question: do these increasingly capable models still need us to specify how they should self-improve? In this paper we study an autonomous paradigm in which, given basic primitives, resources, and a general goal for self-improvement, agents construct their own self-learning algorithm for context-based optimization. Using the setting of gameplay as a testbed, a frozen LLM agent is given a compute budget and tools for collecting experience and updating text documents, and allowed to freely experiment without an externally imposed loop or access to real-world data. We find that sufficiently capable models do self-improve under these conditions, demonstrating strategies that align with classic cognitive frameworks for self-regulated learning and constructs from machine learning literature on self-generated curricula. Furthermore, we study the extent to which the learned text artifacts transfer to external benchmarks, observing improvement across different game horizons, mechanics, and 3D navigation.

0.27

An inexact stochastic variance-reduced gradient method for constrained convex finite-sum optimization

cs.LG math.OC ·2026-11-17 ·Zhaosong Lu, Yifeng Xiao
published
Abstract · Accepted by SIAM Journal on Optimization

In this paper, we study a class of constrained convex finite-sum optimization problems. In many practical applications, all or some of the constraints are hard constraints that represent imperative requirements. Consequently, any desirable approximate solution must (nearly) satisfy these constraints. Motivated by this, we focus on computing an $ε$-$surely\ feasible\ stochastic\ optimal$ ($ε$-SFSO) solution, in which the constraint violation is deterministically bounded by $ε$, and the expected optimality gap is at most $ε$. While a recently proposed stochastic variance-reduced first-order penalty method can compute an $ε$-SFSO solution, there remains a significant gap between its first-order oracle complexity for computing an $ε$-SFSO solution and that for computing an $ε$-$expected\ feasible\ stochastic\ optimal$ solution, where both the expected constraint violation and the expected optimality gap are bounded by $ε$. To narrow this gap, we propose an inexact stochastic variance-reduced gradient (SVRG) method that significantly improves the previously best-known first-order oracle complexity for computing an $ε$-SFSO solution. Furthermore, when the objective consists of a single function, the proposed method reduces to a novel deterministic first-order method for solving standard convex conic optimization. It achieves a significantly lower gradient evaluation complexity with respect to the objective function compared to existing methods, and can outperform them when computing constraint gradients and projections onto the associated cone is cheaper than evaluating the objective gradient. Preliminary numerical results are presented to illustrate the performance of the proposed method.

0.27

How Many Training Labels Are Wrong? Auditing a Fitted Model through Gradient Allocation

cs.LG ·2026-11-17 ·Yuchen Li, Zongqi Fan, Xintong Li, Nguyen H. Tran, Ken-Tye Yong
Abstract · 14 pages, 8 tables, 1 figure. Preliminary version

How many labels in a training set are wrong, and which ones? The usual answers read a held-out set or the model's outputs. A held-out reading depends on how the held-out set was drawn, and an output-based reading fails once the model has fitted its wrong labels, because the loss of a fitted wrong label is small and the prediction agrees with it. We read the answer from the per-example gradients of a single trained model. Each example, together with its label, asks the network to change in a particular way. We normalise the magnitudes of its gradient into a gradient allocation and compare it with the mean allocation of every class. An example is flagged when relabelling it as its nearest other class makes its allocation fit better than under its own label, and the share of flagged examples is the reading of the noise rate. The rule has no threshold and needs no clean data, no held-out set and no further training. After the model has fitted 99-100% of its wrong labels, the allocation still separates wrong from right labels with an AUC of 0.94-0.96 on the noisiest human labels of CIFAR-10N, where a feature-space margin falls to 0.79-0.84 and the loss, the gradient norm and the self-influence of an example to 0.73-0.76. On such a fitted model, readings of the noise rate from predictions or feature margins fall to 0-6% at true rates of 8-58% on CIFAR-10 with ViT-S, while the audit reads 54-96% of the true rate, and at least 93% of the examples it counts carry wrong labels. On a briefly trained model, before the wrong labels are fitted, the reading lies within 1.7 percentage points of the true noise rate on CIFAR-10 from 0% to 58% noise. Confident learning, which trains four further models, separates wrong labels about as well (AUC 0.925-0.998 against 0.900-0.995) and mostly reads the noise rate more closely, at a lower precision. An audit of 5,000 examples takes 14 to 27 seconds on one GPU.

0.26

RSI-Prime: Training Models to Train Models

cs.AI cs.LG ·2026-11-17 ·Dewu Zheng, Yang Ye, Hongyu Zhang, Jun Yu, Yanlin Wang
benchmarkdataset
Abstract

Enabling large language models (LLMs) to autonomously conduct training and improve target models is an important foundation for recursive self-improvement (RSI). As LLMs expand their role from being trained to conducting training, their model-training capability warrants dedicated study and systematic development. To this end, we construct RSI-Prime, a dataset for improving model-training capability, comprising 350 complete experimental trajectories across seven task categories. To help LLMs learn to make high-quality experimental decisions from training practice, we perform fine-grained decision identification and quality annotation, then apply decision-level loss masking to exclude supervision on low-quality behavior. Experiments with GLM-5.3 on PostTrainBench, a benchmark for autonomous post-training, show that training on RSI-Prime yields a 31.95% relative improvement in overall performance over the original GLM-5.3, highlighting a path toward future RSI systems in which models autonomously sustain their own advancement.

0.26

A Quantum Support Vector Regression Framework for Missing Data Imputation in Ship Performance Data

cs.LG quant-ph ·2026-11-17 ·Daniel Maroto Sanchez, Christian Velasco-Gallego, Miriam Bastante Chichon, Jaime Álvarez-Benayas, Tsoulakos Nikolaos +1 more
Abstract

The Internet of Ships enables vessels to generate vast amounts of data, yet missing values remain a persistent issue. To the best of the authors' knowledge, existing imputation techniques in this domain rely exclusively on classical machine learning. This study investigates quantum support vector regression (QSVR) for imputing missing values in ship performance data. Classical support vector regression and QSVR models based on different quantum feature maps are compared for estimating marine fuel consumption under varying levels of data availability. The case study uses real-world data from a Panamax bulk carrier, covering 37 variables across six categories. Three experiments are conducted: exploratory analysis with three input features, scalability analysis with increasing dimensionality (3 to 18 features), and comparative evaluation under a selected representative configuration. Across all scenarios, both classical and quantum approaches captured the relationship between the inputs and the main-engine fuel-flow rate, maintaining strong imputation performance even under severe data scarcity. ZZFeatureMap_1 achieved the highest average R2 (0.991 in the performance comparison conducted with 60% of the dataset, 90% missing data, and 14 input features, for instance) and the lowest aggregate imputation errors, while the radial basis function kernel was the strongest classical baseline, delivering highly competitive results.

0.26

TabDAG: Structured Evolutionary Search for Tabular Feature Engineering

cs.NE cs.LG ·2026-11-17 ·Anastasia Volkova, Andrey Kuznetsov, Ivan Oseledets, Valentin Khrulkov
Abstract · Preprint under review

Feature engineering can improve tabular prediction, but useful transformations are compositional, data-dependent, and subject to leakage constraints. We present TabDAG, a framework that combines evolutionary search with a graph representation of feature-engineering programs. Nodes implement row-wise or data-fitted transformations; edges encode dependencies among raw variables, intermediate values, and final features. The representation supports interactions, statistical encodings, clustering- and neighborhood-based features, and multi-stage compositions in a single executable program. Mutations are structured edits constrained by a parent-specific schema and checked before evaluation. In our experiments, no structurally invalid program reached evaluation, although proposals could be rejected and valid graphs could still fail at runtime. Across ten regression and classification datasets and three tabular model families, evolved features improve predictive performance over raw inputs. TabDAG and its combination with OpenFE achieve the two best mean ranks among the compared methods; pairwise rank differences between feature-engineering methods do not survive multiple-testing correction. The LLM-based searches use the same backbone and candidate-evaluation budget. A controlled ablation compares free-form Python, whole-graph JSON, and constrained graph edits with the search engine, budget, and backbone held fixed. Constrained edits yield fewer failed evaluations than either rewrite baseline. The resulting graphs expose the transformations and dependencies selected by search, enabling feature-level analysis of the observed gains.

0.26

DLS: Dual-Logic Graph Neural Networks with Chaos-Fuzzy Residuals for Robust Link Prediction

cs.LG cs.SI ·2026-11-17 ·Mohammad Mehdi Keikha, Abolfazl Nadi
Abstract · Submitted to World Wide Web (Springer Nature) on January 5, 2026. Preprint version, currently under peer review

Link prediction is a fundamental task in graph mining, critical for applications ranging from biological interaction discovery to recommender systems. While Message Passing Neural Networks (MPNNs) have achieved state-of-the-art results, they are fundamentally limited by the inadvertent suppression of high-frequency signals, as the recursive aggregation that stabilizes structural learning inherently leads to over-smoothing, rendering node representations indistinguishable in deep layers. This limitation is particularly severe in heterophilous graphs, where structural proximity does not imply feature similarity. In this paper, we propose DLS (Dual-Logic SAGE), a novel inductive framework designed to break the over-smoothing barrier via a "controlled instability" mechanism. Unlike conventional methods relying on static structural heuristics, DLS introduces a dynamic residual pipeline. First, a Chaotic Injection Module utilizes the Logistic Map to diversify the feature manifold, exploiting sensitivity to initial conditions to separate topologically similar but semantically distinct nodes. Second, to mitigate the aleatoric uncertainty introduced by chaotic dynamics, a Fuzzy-Logic Gating Mechanism applies soft-thresholding to filter noise and retain only high-confidence signals. Experimental results on diverse benchmarks demonstrate that DLS effectively balances feature discrimination with robustness, significantly outperforming state-of-the-art baselines in heterophilous settings while maintaining competitive efficiency on standard homophilous graphs.

0.26

Residual-Guided Lightweight Reinforcement Learning for Risk-Aware Trajectory Selection with a Frozen World Model

cs.AI cs.ETcs.LGcs.RO+1 ·2026-11-17 ·Jiaqi Wu, Yongqi Dong, Yijie Zhong, Maple Shao, Yide Tao +4 more
Abstract · 6 pages, 2 figures

In automated driving, world models provide future scene predictions for evaluating candidate trajectories, but fixed selection rules may not capture scene-dependent trade-offs. We propose a residual-guided reinforcement learning (RL) framework for lightweight trajectory selection with a frozen world model and a fixed candidate bank. The framework uses proximal policy optimization (PPO) to fine-tune a joint policy over candidate selection and bounded trajectory residuals. A context-conditioned ranking network learns candidate preferences from feedback obtained by evaluating local perturbations of the selected trajectories. At inference, the selector combines learned preference and predicted risk ranks using fixed weights and returns the selected candidate trajectory without modification. Fine-tuning updates 16,106 parameters across the actor and critic, while the deployed selector retains only the 11,725-parameter ranking network, with the residual branch and critic used exclusively during training. In open-loop evaluation on the nuScenes validation set, the method reduces average Euclidean trajectory error (L2) by 3.99% relative to the RiskWorld model baseline without RL-based selection. Under matched budgets for feedback accesses and actor updates, residual exploration reduces L2 by 0.83% relative to the categorical PPO variant trained without residuals. The method also maintains lower L2 than RiskWorld when risk features are removed or shuffled.

0.26

GP--RL--MPC: Uncertainty-Aware Reinforcement Learning-Augmented Model Predictive Control with Gaussian Process Postdecision State-Value Functions

cs.LG eess.SY ·2026-11-17 ·Jon Li, Jay H. Lee
Abstract · 19 pages including references and appendices, 3 figures, 2 tables. Submitted to L4DC 2027

We propose Gaussian Process Reinforcement Learning-Augmented Model Predictive Control (GP--RL--MPC), which uses a learned critic as the MPC terminal penalty to account for costs beyond the prediction horizon. Gaussian process regression (GPR) enables data-efficient critic learning from limited, noisy rollouts within optimistic approximate policy iteration (OAPI), with soft updates mitigating policy chattering. A tunable risk coefficient weights GP epistemic uncertainty in the terminal cost, regulating sensitivity to uncertain critic predictions. The postdecision state formulation shifts disturbance averaging offline to account for aleatoric uncertainty from the terminal transition onward. Under matching objectives and admissible actions, a one-step horizon recovers greedy value-space RL. After fitting GP hyperparameters by exact marginal likelihood, we construct a sparse deterministic training conditional (DTC) posterior using inducing points selected by pivoted Cholesky (PC), reducing online evaluation costs as the training dataset grows. In a continuous stirred-tank reactor (CSTR) subject to exogenous disturbances, GP--RL--MPC with an appropriately selected risk coefficient and policy screening reduces the sum of squared errors (SSE) by an average of 80.52% relative to an otherwise matched baseline MPC across five random seeds (sample standard deviation: 3.48%).

0.25

Amortizing Tempered Sampling with Corrected Self-Training

cs.LG stat.ML ·2026-11-17 ·Hoang-Viet Nguyen, Hai-Dang Dau, Alexandre H. Thiéry
Abstract

Learning a neural sampler can reduce the cost of repeated sampling from an un-normalized distribution, but obtaining good training data remains difficult. We develop a corrected self-training scheme across a sequence of tempered targets, making use of importance sampling and Markov chain Monte Carlo to correct the samples and replay buffers to store them. We introduce Sequential Score Sampler (SSS) as a particular case. In SSS, the proposal is a score-based diffusion trained by regression on energy gradients; and the samples are corrected by importance weights on the path space. After training, samples can be collected directly from the model, from its importance-corrected proposals, or from the final-temperature buffer. The method amortizes repeated sampling by encoding the global structure of the target in the model. SSS achieves competitive sampling accuracy across the benchmarks and, on several targets, uses shorter runtimes with fewer energy evaluations than competing samplers.

0.25

InfuseLoop: Cross-Trajectory Latent Injection for Scalable Parallel Looped Transformers

cs.LG ·2026-11-17 ·Zhaofeng Sun, Lanxiang Hu, Yu-Yang Qian, Jiaqi Leng, Yuanbo Yang +4 more
Abstract

Parallel Looped Transformers (PLTs) use otherwise idle decoding compute by overlapping recurrent iterations. However, vanilla PLTs scale poorly with depth: they exhibit training-loss spikes, and their quality gains shrink as more loops are added. The benefit of additional loops depends on the information available to each iteration, so we examine whether the existing computational dependencies limit effective depth scaling. We find that only the preceding in-trajectory state is passed across loop boundaries, even though completed cross-trajectory representations are already available. Based on this finding, we propose InfuseLoop, a cross-trajectory latent injection mechanism. It aggregates causally available hidden states from later prediction trajectories through lightweight attention residuals. This design preserves token-parallel training and concurrent loop execution, and requires only $O(Rd)$ additional hidden-state storage. Experiments show that InfuseLoop eliminates training-loss spikes and improves validation-loss scaling up to eight loops. At eight loops, InfuseLoop reduces perplexity by 3.4--17.1% relative to vanilla PLTs and by 10.5--37.3% relative to the non-looped baseline. It also improves average downstream accuracy by 0.84 and 1.99 percentage points over these two baselines. InfuseLoop incurs only 0.3% decoding time over vanilla PLTs and 23.3% over the non-looped baseline, while keeping the same logical KV-cache footprint as vanilla PLTs.

0.24

Large-scale Neuron Community Discovery Reveals Neuron-Sparse Behaviors in Language Models

cs.LG cs.CL ·2026-11-17 ·Igor Shilov, Thomas Fel, Tal Haklay, Lucius Bushnaq, Owen Lewis
Abstract

What reusable units of computation can we discover within a language model's neurons? Prior work has shown that individual forward passes can often be sparsified post-hoc to small groups of neurons, but it has been unclear whether these groups form coherent computational entities reused across inputs. We introduce neuron community discovery, an unsupervised method for finding neuron groups that recur across prompts and are causally necessary and sufficient, within the layers they span, for common computations. The method combines gradient-based attribution, a graph connecting neurons that are repeatedly attributed together, and community detection. Applied to Llama-3.1-8B using ten million prompts, it discovers thousands of candidate communities. We evaluate them on held-out prompts through three causal tests: community necessity, sufficiency within a layer subset, and layer necessity. These tests identify groups whose removal disrupts predictions and whose retention preserves predictions despite ablating the remaining MLP neurons in the tested layers. Discovered communities support functions like grammatical agreement, token-probability calibration, fixed phrases, in-context copying, addition, and multiplication. With communities capped at 128 neurons across all layers, 10.3% of held-out prompts have at least one layer's contribution explained on a neuron-sparse basis. Coverage remains 8.7% when requiring at least four layers and 3.3% when requiring at least twelve. These results show that small groups of existing neurons can provide reusable computational components across prompts, and that such groups can be discovered at scale.

0.24

Escaping Latent Traps by Remembering: Training-Free Episodic Kernel Repulsion for World-Model Planning

cs.LG ·2026-11-17 ·Yican Zhao, Bolin Yang, Ruizhe Wang, Haiyang Guo, Fuji Yang +2 more
Abstract · 17 pages, including references and appendix

Latent world models plan towards a goal image by minimising the distance between predicted and goal latents. This distance is not geodesic: it has local minima, latent traps, behind walls, at corners and in dead ends, and memoryless receding-horizon replanning keeps the agent in them, revisiting states far from the goal. We propose Episodic Kernel Repulsion (EKR), a training-free cost term that penalises candidate plans by a top-$k$ Gaussian kernel to a novelty-gated memory of the latents the agent has actually visited, rescaled by the current base cost so that its strength is comparable across models. Across two JEPA world models and four benchmarks, EKR raises average success from 31.8% to 46.7% (LeWM) and from 54.7% to 64.0% (H-JEPA) with the same model, planner and budget. Same-memory counterfactuals show that, in most settings, EKR pushes plans away from visited states, in navigation, often at a higher base cost; stuck episodes and revisits drop accordingly. In seven of eight settings it matches or exceeds $4\times$ samples at $0.8$--$1.5\times$ compute. Without tuning, it transfers to DINO-WM with CEM: $+6.7$ and $+19.3$ points on official and longer-goal Wall tasks, and from 44.7% to 100.0% and from 21.0% to 32.0% on long-horizon PointMaze and Push-T.

0.24

Outcome-Balanced PPO and Policy-Induced Flow Rectification for Flow-Based VLA Policies

cs.AI cs.LG ·2026-11-17 ·Yuhan Wang, Ruiyang Wang, Jiaguan Zhu
Abstract · 15 pages

Reinforcement learning (RL) enables post-training of flow-based vision-language-action (VLA) policies by optimizing expected task return through environment interaction. However, this optimization raises two challenges: policy updates can alter the velocity-field geometry underlying the score approximation used for stochastic transitions, and sparse multi-task rollouts can yield uneven learning signals across tasks and advantage signs. We introduce an RL framework that combines Outcome-Balanced Proximal Policy Optimization (PPO) with policy-induced flow rectification. Outcome-Balanced PPO reweights positive advantages using task success-rate estimates and rescales positive and negative advantages while preserving their signs. Policy-induced rectification periodically fits a linear-bridge flow to ODE endpoint pairs generated by the current policy, refreshing the representation used in subsequent stochastic optimization. We evaluate the framework on simulated manipulation benchmarks and additionally study a component combination in a Real2Sim2Real task, examining task performance, optimization behavior, and flow consistency. The proposed framework improves success rates over the evaluated baselines across the simulated benchmarks.

0.24

Physically Admissible Generative Flow Matching for Sparse Flow Reconstruction

cs.LG ·2026-11-17 ·Yuhao Wang, Yuanxi Peng
Abstract

Reconstructing full fluid velocity fields from sparse, localized observations is a severely ill-posed inverse problem: most of the spatial domain remains unobserved, discrete conservation laws must hold strictly, and predictive uncertainty must be reliable. Existing generative models typically enforce physical constraints only at the terminal output or via soft penalties, leaving intermediate generative trajectories unconstrained. We introduce Feasible Subspace Consistency (FSC), a framework that preserves physical admissibility as an invariant of continuous generative dynamics: tangent-space matching confines the entire ODE trajectory to the discrete conservation manifold, while mass-orthonormal coordinates preserve the kinetic-energy metric. Sparse reconstruction then reduces to amortized posterior transport in the latent space of a frozen FSC prior, ensuring every posterior sample inherits incompressibility and boundary conditions without per-sample projection loops. We validate FSC on a periodic Kolmogorov flow, where exact spectral projection isolates path tangency and metric isometry, and a wall-bounded cylinder wake with immersed obstacles, realized via matrix-free iterative projection. FSC prevents the mode-coverage collapse induced by terminal projection (increasing mode coverage from $0.000$ to $0.790$) while eliminating high-wavenumber spectral distortion. In sparse reconstruction, amortized FSC reduces global relative error by $50\%$ and $46\%$, and sensor residuals by $63\%$ and $55\%$, relative to the unconditioned prior, while delivering calibrated, spatially heterogeneous uncertainty.

0.23

TempQuant: Learning Adaptive Temperatures for Quantized Language Models

cs.LG ·2026-11-17 ·Mukund Athmaram, Liang Ma, Manoj Rao
Abstract

Large language models are increasingly deployed at low precision, yet quantization can degrade their performance. We study how quantization interacts with decoding settings such as the sampling temperature and top-p by examining how it distorts the model's next-token probabilities. Measuring entropy during generation across deployed temperatures, we find that quantization warms the model on average, and that temperature matters much more for the performance of a quantized model than for its full-precision counterpart. Because quantization and temperature both shape the next-token distribution, we investigate how much the sampling settings can undo what quantization changes, through our method TempQuant, which learns the quantized model's temperature and top-p to bring its next-token probabilities closer to the full-precision model's. We learn these settings at two granularities. The first is a single global setting per requested temperature, which finds nearly the same temperatures that tuning on labeled answers would, without any labels. Model cards typically recommend sampling a quantized model with the same settings as its full-precision counterpart, while we find that different settings can substantially improve its accuracy. On 3-bit Qwen3-4B, for example, global TempQuant raises MATH accuracy from 34% under the configuration recommended for Qwen's quantized releases to 51%. The second sets the temperature and top-p for every token, motivated by our finding that tokens whose rankings flip under quantization tend to need warming, while tokens whose rankings are preserved tend to need cooling. When quantization degrades accuracy substantially, TempQuant recovers much of it, and when degradation is lighter, it still brings the quantized model's behavior closer to the full-precision model's.

0.23

AlignedHPO: Efficient Surrogate-Guided Loss Hyperparameter Optimization

cs.LG cs.AI ·2026-11-17 ·Ivan Karpukhin, Andrey Savchenko
Abstract

Modern deep learning losses involve multiple hyperparameters, such as weights, scales, and margins, typically tuned manually or through expensive black-box optimization despite their differentiability. We propose AlignedHPO, an efficient framework for tuning differentiable loss hyperparameters using gradient alignment and surrogate validation objectives. Our method employs Jacobian-vector products to reduce hypergradient computation time and memory, while differentiable metric surrogates guide optimization toward task-specific evaluation criteria. We address gradient magnitude ambiguity and stochastic instability through normalized alignment and gradient smoothing. Experiments with generalized Focal Loss and ArcFace demonstrate effective adaptation across long-tailed classification and open-set face recognition. AlignedHPO achieves competitive or improved performance compared to implicit differentiation, with nearly 12x faster hypergradient computation. These results highlight the potential of efficient surrogate-guided loss hyperparameter optimization.

0.23

MedGuardBench: A Patient-Oriented Benchmark for Personalized Guardrail Development

cs.AI cs.LG ·2026-11-17 ·Sara Khosravi, Xiaofei Wen, Majid Sarrafzadeh, Muhao Chen
benchmark
Abstract

Safety is personal: the same action can be safe for one person but unsafe for another. Yet current guardrails judge content without considering who is affected. Using medication safety, we introduce \textit{patient-conditioned medication guarding}, which predicts a safety verdict and 17 risk categories from a patient profile and prescription. We release \textbf{MedGuardBench}, a drug-disjoint benchmark of 5{,}001 scenarios across 87 medications. We find that standard guardrails rely on surface cues, such as wording and profile presence, rather than patient facts. To address this, we develop \textit{personalized reasoning guardrails} with three components. \textbf{Personalized Reasoning Distillation} links risks to relevant patient fields. \textbf{Context Unrolling (CU)} makes models check each field themselves, reducing false alarms on most backbones; controls show that providing the check does not yield the same benefit. A \textbf{rule-based reward (guided GRPO)} improves unsafe recall and complex-risk detection, including allergies and adverse reactions, on some backbones, but increases false alarms on others. Our models outperform base models and matched no-profile controls in accuracy, while reducing false alarms relative to no-profile controls. Profile removal and swapping confirm that gains come from patient information, not fine-tuning alone.

0.23

RiDEx: Reward-Informed Directional Exploration for Diffusion Reinforcement Learning

cs.LG cs.CV ·2026-11-17 ·Song Yan, Chenfeng Wang, Wei Zhai, Xingchen Wang, Tao Zhang +4 more
Abstract

Reward-based diffusion post-training updates model parameters from task rewards. While existing approaches improve sampling and policy optimization, they still rely largely on isotropic random perturbations in a high-dimensional latent space for exploration. Such variation often spends the perturbation budget on changes with little influence on the target reward, making exploration inefficient. We study where diffusion models should explore and how to learn from that exploration. We introduce Reward-Informed Directional Exploration (RiDEx), which concentrates local exploration in reward-relevant subspaces of the denoising velocity field. Conditional residuals or differentiable reward gradients propose the subspace, and equal-and-opposite perturbations at the same generation state measure its terminal reward response. The reward difference supplies signed local supervision in the explored coordinates, while all generated images contribute to global DiffusionNFT learning. RiDEx thereby couples exploration-space selection with feedback about the changes it produces. On Stable Diffusion 3.5 Medium, RiDEx attains the highest peak rewards among the evaluated methods on four alignment and preference tasks. Its gradient-direction instantiation achieves 15x and 29x target-attainment speedups over DiffusionNFT on HPSv2 and PickScore, measured in H200 GPU hours. Controlled experiments confirm that reward-aligned directions elicit stronger responses at equal perturbation magnitude, and that applying their measured credit improves curve-average reward across all four tasks.

0.23

HADVQ: Hessian Aware Directional Vector Quantization

cs.LG ·2026-11-17 ·Daniil Li, Mars Davletshin, Vadim Klyzhenko, Artem Danilishin, Valery Sandul +2 more
hot direction
Abstract

Ultra-low-bit quantization of large language models is a challenging task because aggressive compression severely distorts model parameters and disproportionately harms a small number of highly sensitive channels. We present a vector quantization method tailored for this regime. Our approach begins with random preconditioning, which regularizes the weight space and makes it more amenable to quantization. We then represent weights in polar coordinates and decouple the quantization of directions and magnitudes. To quantize directions, we leverage the structure induced by random preconditioning and design a structured codebook based on the $E_8$ lattice, yielding a compact yet expressive representation. To further reduce degradation, we detect extremely important channels using a proxy Hessian and preserve them with an additional cost of only 0.16 bits. Together, these components provide an effective quantization pipeline for LLMs in the extremely low-bit regime. Empirical results demonstrate improved accuracy, highlighting the benefit of combining geometric structure, lattice-based codebooks, and selective channel preservation.

0.23

What RL on Expert-Built Tasks Teaches Nemotron 3.5 Super

cs.LG cs.AIcs.CLcs.SE ·2026-11-17 ·Sushant Mehta, Edwin Chen
Abstract · 18 pages, 2 figures, 9 tables

We post-train NVIDIA's Nemotron 3.5 Super (120B total, 12B active parameters) with reinforcement learning (RL) on five expert-built task datasets, one at a time: real-world software engineering, complex constraints, long-horizon tool use, professional chart understanding and enterprise document reasoning. Each run trains a fresh rank-32 LoRA adapter with CISPO on Tinker, with no supervised stage, and each adapter is compared with the base model on the same tasks, harness and settings. On twelve headline benchmarks, eight built by other groups and four by Surge AI, the gains range from +4.5 points on IFBench to +13.8 on GDP_pdf, which rises from 10.4% to 24.2%; SWE-bench Pro V2 rises from 86.9% to 92.1%, Terminal-Bench 2.1 from 63.9% to 70.3%, MultiChallenge from 58.3% to 66.6%, AutomationBench from 14.8% to 27.5% and Chartography from 39.4% to 50.6%. The document adapter, trained without API tools, also gains +8.0 points on AutomationBench. In attempts read side by side with the base model's on tasks each adapter newly solves, the base model's failures recur in a few modes, and each adapter does more often what the base model missed: the tool-use adapter finds the data and the rules before acting, the coding adapter writes its own test or scratch files on SWE-bench Pro V2, the constraints adapter carries earlier turns into its answers, the chart adapter measures the marks in code, and the document adapter reads AutomationBench's spreadsheets.

0.23

Encoding, Decodability, and Use: A Five-Layer Framework for Benchmark Measurement Audits with Evidence from UAV Intent Recognition

cs.LG cs.AI ·2026-11-17 ·Zhirong Su, Tian Zhang
benchmark
Abstract

Benchmark accuracy is routinely read as evidence that a model understands the construct the benchmark measures. We propose and instantiate on AeroIntent a five-layer framework that decomposes that reading into observation, representation, decoding, use, and behaviour. The framework separates encoding, decodability, and use as distinct levels. In the instantiation, the observation layer is unreliable on 7 of the 20 Drone programs and floors the direct visual correlate of radial velocity at a 0.26\% scale change. The representation layer exposes two task-designed RGB features with nominal signals. RAFT flow radial bias reaches $r=+0.257$ ($p=0.030$), and bounding-box optical-flow divergence reaches $r=+0.203$ ($p=0.029$). No multi-feature readout gives a fold-mean test correlation above the cluster minimum detectable effect (MDE), and the largest positive global out-of-sample correlation is $r=+0.080$. The decoding layer sets the MDE at $|r|=0.599$ at the multi-feature Bonferroni threshold ($α=0.00714$, power 0.80), and the observed $r=+0.080$ has power 0.00. Two injection experiments establish the MDE's condition from both sides: aligned injections are detected at or above the simulated power, and orthogonal injections are not detected. The use layer finds accuracy consistent with program memory (program-identity 0.937/0.597/0.210) and multi-region static cues. The behaviour layer is consistent with accuracy not requiring a radial-velocity readout (appearance 0.6552 vs bounding-box trajectory 0.4080, cluster-level paired $Δ=+0.2471$, $p=0.0132$).

0.22

Ambient LLMs: Data-Efficient Supervised Fine-tuning via Temperature Matching

cs.LG cs.AIcs.CL ·2026-11-17 ·Giannis Daras, Marios Papachristou, Adam Klivans, Antonio Torralba, Constantinos Daskalakis
Abstract · Work in progress

Collecting large-scale, diverse, and high-quality post-training datasets for autoregressive language models is challenging. Consequently, Supervised Fine-Tuning (SFT) is typically applied to datasets significantly smaller than those used for pre-training, which leads to overfitting and the catastrophic forgetting of the tails of the data distribution. To prevent this collapse, we introduce Ambient SFT, a principled framework for leveraging imperfect predictions to augment the fine-tuning set. In our method, a "frozen" version of the pre-trained model serves as a noisy teacher that the fine-tuned model must match at a coarse level. In particular, we require that the high-temperature versions of the conditional distributions produced by the post-training model should match the high-temperature versions of the pre-training ones. The temperature hyperparameter controls the level of trust in the pre-trained model, with the limit $T\to \infty$ reverting back to regular SFT and $T=1$ to knowledge distillation. We demonstrate that our approach outperforms standard SFT across domain adaptation, instruction-tuning, and reasoning, leading for up to $21\%$ average relative improvement to standard SFT.

0.22

VRLLM: Turning LLMs into Powerful Tabular Classifiers

cs.LG cs.CL ·2026-11-17 ·Yingtong Dou, Shubham Jain, Junpeng Wang, Hadi Abdullah, Vineeth Rakesh Mohan +3 more
benchmark
Abstract · Technical Report. WIP

Large Language Models (LLMs) have shown strong performance on text, but whether pretrained LLMs are effective for large-scale tabular classification remains unclear, as many previous works demonstrate that LLMs' tabular data classification performance falls behind tree-based models or neural architectures designed for tabular data. At Visa Research, we managed to fine-tune a pretrained Llama 3.2 1B LLM on billions of rows of labeled tabular data and make it outperform a wide array of tabular classifiers and even our competitive production deep learning model. For ease of reference, we refer to the model as the Visa Research LLM (VRLLM). In this paper, we share how we achieved this exciting result, covering data preparation, training recipe, ablation studies of design choices and deployment validation. We also benchmark frontier LLMs on the same task and analyze their prediction behavior across model families and generations. Our interpretability experiments suggest that VRLLM learns feature-specific predictive associations, while component ablations indicate that the earliest two FFN layers contribute to the emergence of linearly decodable task information. Our paper provides an engineering recipe and a focused empirical case study on adapting compact LLMs to industrial tabular prediction tasks.

0.22

Action-outcome coupling: a pre-investment feasibility audit for offline reinforcement learning in supply chain operations

cs.LG cs.CEecon.EMstat.AP ·2026-11-17 ·Tung Dang, The Hung Phung, Son Lam Nguyen, Tu Nguyen
theory
Abstract · 40 pages

Prescriptive analytics initiatives in logistics, such as offline reinforcement learning and next-best-action systems, are often funded before anyone verifies that the available historical logs can support an improved decision policy. We formalise this pre-investment feasibility question via action-outcome coupling: a total-effect variance index measuring how much expected outcomes change with the logged action. Coupling provides an algorithm-independent upper bound on any policy's achievable value, expressed in monetary units per decision for both single-step and discounted sequential objectives. We also analyse the asymptotic false-licensing risk of an invalid workflow: fitting a value model and evaluating the model's greedy policy on the same data. The false-licence rate is driven mainly by the number of state cells over which selection is performed, rising from the nominal test level with one cell to near certainty with a few dozen, while overlap has a smaller effect. We then propose an audit that assigns logs to four tiers of decision support, separately for fixed and contextual policies, and returns an explicit stopping rationale. On a randomised e-commerce log, the audit endorses a fixed recommendation policy, certifying gains equal to 30 per cent of the theoretical ceiling; the remaining gap reflects bound looseness under an 80-action design. Two public logistics logs fail at different checkpoints: one because the target is an accounting identity of the action, and another because, with 27 clusters, estimated coupling stays below the design's detectability threshold even though a conservative offline learner certifies an improvement in 29 of 30 seeds.

0.21

High-Dimensional Black-Box Optimization via Guidance Sampling of Diffusion Models

cs.LG cs.AI ·2026-11-17 ·Xiaoyan Hu, Rohit Sonker, Jalal Butt, Jiayu Chen, Egemen Kolemen +1 more
Abstract

Many real-world optimization problems involve optimizing black-box objectives for which only a limited number of evaluations are available. Bayesian optimization (BO) is a powerful framework for such problems, but its effectiveness often deteriorates in high-dimensional spaces. Recent work leverages diffusion models for black-box optimization, often through pre-training a reward-conditioned score network to steer generation towards high-reward samples. Alternatively, gradient-based guidance methods assume access to a differentiable reward function from which guidance gradients are queried. However, these approaches either require large amounts of labeled data for offline training or access to exact reward gradients, limiting their applicability in black-box settings. We propose BReaD (Black-box Reward-guided Sampling of Diffusion Model), a guidance sampling framework for optimizing black-box rewards without assuming access to a labeled dataset or reward gradients. The proposed method uses a Gaussian process (GP) as a differentiable surrogate for the unknown objective and actively queries labels for the generated samples. We provide a theoretical analysis of the total-variation distance between the generated distribution and the target distribution. Using Upper Confidence Bound (UCB) as guidance signals, our method balances exploration and exploitation during generation and enables the discovery of high-reward regions in the data space. Numerical experiments on scientific discovery and image generation tasks demonstrate that our framework can effectively optimize black-box rewards in low-data scenarios.

0.20

Eyes in the Update? Learning Where Rewards Land in Multimodal Reasoning

cs.LG cs.AI ·2026-11-17 ·Lixuan He, Tianzhe Chu, Shulin Tian, Shikang Zheng, Feng Chen +1 more
Abstract

Multimodal post-training has moved from broad instruction tuning toward reinforcement learning for image and video reasoning, where outcome rewards can improve final-answer accuracy. Yet these rewards are assigned to complete generated trajectories, while visual evidence is used only at sparse and uneven positions inside the response. Recent work has begun to make multimodal RL more token-sensitive through grounding signals or external token-level weighting. We ask a more specific question: can the allocation of the same trajectory-level reward be inferred from the model's own multimodal computation, rather than from an auxiliary annotator, perception module, or process label? We propose Gaze-RL, a representation-driven policy optimization method that redistributes completion-level advantage through a latent reward landing field. The field is computed from internal cross-modal geometry and local predictive uncertainty, assigning larger update weight to positions that are both visually anchored and decision-bearing while remaining mean-one, so the total credit of each rollout is preserved and only its positionwise allocation changes. Gaze-RL requires no process supervision, auxiliary perception module, or architectural modification, and improves both Qwen2.5-VL and InternVL3.5 across a broad range of image and video benchmarks.

0.19

Robust Actor-Critic Learning under Transport and Divergence Uncertainty

cs.LG stat.ML ·2026-11-17 ·Kunqi Huang, Lican Kang, Hongrui Shu, Jincheng Ying
Abstract

Learning reliable control policies requires accounting for uncertainty in transition dynamics and errors introduced by function estimation and action generation. We develop an actor--critic method for discounted robust Markov decision processes with continuous states and actions and an unknown nominal transition law. We consider transition uncertainty sets defined by Wasserstein distance, total variation, Kullback--Leibler divergence, or Pearson chi-square divergence. A conditional diffusion model is fitted once to offline transitions. At each subsequent iteration, we estimate the robust action-value function from current-policy observations using a shared-target minimax criterion, then update a generative actor through exponentially weighted score matching. Regularization preserves exploration and provides the density regularity needed for conditional distribution estimation. Under coverage and regularity conditions, we establish finite-sample bounds for critic estimation, the learned conditional action law, and the robust return of the implemented policy. The bounds distinguish critic estimation error, uncertainty in the nominal transition model, and policy regularization, and characterize how the four robustness radii affect the resulting guarantees.

0.19

Two Counterexamples for e-means

cs.LG stat.ML ·2026-11-17 ·Zhenisbek Assylbekov
Abstract

e-means (Bessa et al., 2026) combines exponential averaging and cluster revival. We give two counterexamples to its published recipe with higher final loss than standard k-means (Lloyd, 1982) from identical initial centers. Mass-based revival can destroy a global optimum and make the ratio of expected losses unbounded even with k-means++ seeding.

0.19

SkillBuffBench: Measuring What Agent Skills Contribute across Harnesses, Models, Skill Access Modes and Memory

cs.AI cs.CLcs.LG ·2026-11-17 ·Gengyue Dong, Ye Wei, Zheng Zhu, Yidong Han, Jiale Miao +2 more
benchmark
Abstract · 18 pages, 2 figures, 6 tables. Work in progress: benchmark design and protocol; results to be added in a later version

Agent Skills, file-based packages of procedural knowledge that a coding agent can discover and load on demand, are now shipped by every major agent harness, yet the evidence for how much they help comes almost entirely from settings in which the right Skill is hand-picked for the task and the agent starts from a blank slate. Neither holds in deployment, where an agent chooses from whatever library happens to be installed and frequently carries memory from earlier work. We present SkillBuffBench, a benchmark and paired evaluation protocol that treats Skill access as a three-level factor (no Skill, a curated Skill, and automatic invocation from an installed library), crosses it with memory (off and on), with four agent harnesses and with multiple models (the state-of-the-art proprietary models from Anthropic, Google and OpenAI, and the state-of-the-art open-weight models from Chinese providers), every model being run under every harness, and evaluates the result on research tasks contributed by domain experts together with the Skills they use, currently in materials science and biology, with the suite extended on a rolling basis. Every run is marked by a deterministic verifier; uncertainty is reported with a paired, task-level bootstrap; and cost, agent rounds and wall-clock time are reported alongside the score rather than folded into it. This version of the paper describes the design of the benchmark, its evaluation protocol and its current tasks and Skills. The evaluation is in progress: results will be added once the runs are complete, and the task suite, results and analysis will continue to be updated on a rolling basis.

0.19

Scaling Laws for Vision-Language Fusion

cs.LG cs.CV ·2026-11-17 ·Seng Pei Liew, Teppei Suzuki, Seiya Tokui
Abstract

Vision-language models (VLMs) have emerged as central foundation models combining multimodal perception with text comprehension. While VLMs are traditionally built via late fusion by continued pre-training an off-the-shelf language model on multimodal (image-text) data, mixing modalities from the start of pre-training (early fusion) has recently gained traction. However, the fundamental trade-offs between these paradigms remain poorly understood, resulting in expensive and heuristic-driven training recipes. In this paper, we systematically analyze the learning dynamics of these multimodal pre-training processes, showing that validation losses during vision-language fusion can be modeled by separating the base learning trajectory, where different data types contribute at distinct rates, from a distribution shift driven by the competition between modalities. We derive a scaling law that unifiedly describes these dynamics, enabling accurate loss predictions across mixture ratios, training steps, and model sizes. Crucially, our law characterizes the trade-off between early and late fusion: under a fixed data budget, early fusion preserves text performance while late fusion favors vision performance, yielding a trade-off that increasingly tilts toward early fusion as model size scales. Extensive experiments confirm that our scaling law reliably captures multimodal learning dynamics and governs fusion decisions.

0.18

SOTA or Luck? The Winner's Curse in LLM Leaderboards

cs.LG cs.CLstat.MEstat.ML ·2026-11-17 ·Sikun Xu, Raphael Thomadsen, Dennis J. Zhang, Heng Zhang
benchmarktheory
Abstract

Public LLM leaderboards guide model evaluation and adoption, yet the reported performance of top-ranked models can be systematically overstated. This leaderboard winner's curse arises because ranking noisy benchmark scores rewards sampling luck alongside genuine ability. The same selection effect can distort scores throughout the leaderboard, a phenomenon we call the rank-induced bias. We develop an $m$-out-of-$N$ bootstrap to correct for this bias. It assumes a fixed model pool with independent, identically distributed bounded tasks while allowing arbitrary within-task correlation among models. Under exact population tiers separated by positive gaps, we prove that the bootstrap reduces $Θ(N^{-1/2})$ ranking bias to expected residual bias $o(N^{-1/2})$. Across 14 benchmarks from 5 benchmark families, the rank-1 adjustments range from approximately 0.09 to 3.86 percentage points. In our SWE-bench Verified snapshot, the tied leaders receive an adjustment of 0.90 percentage points. Such overstatement is large enough to make the apparent winner statistically fragile. Thus, we use the same bootstrap to further quantify the intervals of plausible rank positions for each model and group models into tiers. We establish sufficient conditions for consistent recovery of population tiers. On SWE-bench Verified, this procedure places 13 of 173 models in the leading tier. We combine the score adjustment and rank intervals in the Rank Uncertainty and Leaderboard Evaluation Report (RULER). We validate its finite-sample performance using sample splitting and calibrated simulations, and advocate publishing RULER alongside raw rankings so readers can assess both reported performance and the precision of the ordering. RULER also provides guidance for benchmark design that accounts for selection effects and ranking uncertainty.

0.18

Testing World Models on Actions They Have Never Seen: Held-Out Plans, Forecast Accuracy and Price Rules in Retail

cs.LG cs.AI ·2026-11-17 ·Karthik Iyer
Abstract · 12 pages, 3 figures, 5 tables. Companion to arXiv:2611.20811

A world model is trusted to answer "what if": how its world responds to an action it has never seen. We give a test of that trust for any action-conditioned world model: actions held out of training by design and scored against a known answer; whether forecast accuracy predicts accuracy on those actions; and the same structure given to simpler models. We run it in retail, where the action is a price and promotion plan, across 2,853 fitted models. First, forecast accuracy is a fair guide between kinds of model (rank correlation with error on new plans 0.32 to 0.66) and a poor guide between versions of one world model: where merchants plan into high-demand weeks the link reverses, down to -0.69. Second, training the world model to learn the effect of actions (JEPA pretraining, a latent consistency check, adaptation at the plan date) leaves its error on new plans unchanged. Third, in generated worlds the recorded regular price and three price rules improve it. Given the same rules, a log-linear model is best when its curve can match the truth; in a world whose curve lies outside the rules, the world model with the rules is ahead of the log-linear model with the rules in 36 of 40 worlds, and does as well without its recurrent cell. On real data (Breakfast at the Frat, The Complete Journey and M5) the rules lower the world model's error on pure price cuts deeper than any it saw (0.63 to 0.56 on the Frat; 0.45 to 0.44, within chance, on The Complete Journey), but not on lasting price moves. Fourth, a small share of random test plans helps every model a little. In a generated dosing world, a stress test in a second domain, the first finding repeats, and the built-up drug level lowers the world model's error (2.86 to 1.91) though it still predicts only 0.12 to 0.32 of the true gain. The test carries to any domain where a world model is asked about actions outside its history.

0.18

ReMaRL: Learning to Unmask and Remask in Frozen Diffusion Language Models

cs.LG ·2026-11-17 ·Egor Shibaev, Danila Biktimirov, Ivan Klimov, Artem Makoian, Viacheslav Meshchaninov +1 more
Abstract

Masked diffusion language models generate tokens in parallel, and the order in which tokens are unmasked affects final-answer accuracy. Several methods learn unmasking policies from task rewards, but these policies cannot remask earlier predictions. Remasking lets the model correct those predictions using newly generated context. Heuristic rules guide both unmasking and remasking, while supervised remaskers learn to detect token errors. Neither approach directly optimizes these decisions for final-answer correctness. To address this gap, we introduce ReMaRL (Remasking with Reinforcement Learning), a lightweight policy that learns which tokens to unmask, retain, or remask. We train this roughly 0.3M-parameter policy with Group Relative Policy Optimization (GRPO) from final-answer rewards while keeping the diffusion model frozen. With LLaDA-8B-Instruct, ReMaRL achieves the highest scores among the compared unmasking and remasking policies within the tested inference budgets of 4-32 model calls on math and 8-16 calls on code.

0.17

DSADiff: Data Structure-Aware Diffusion for Crystal Structure Prediction via Mitigating Complexity- and Imbalance-Induced Biases

cs.LG cond-mat.mtrl-sci ·2026-11-17 ·Jiang Xiaobo, Dinghong Lai
benchmark
Abstract

Crystal Structure Prediction (CSP) is an important step in accelerating the generation and discovery of new materials. In recent years, diffusion-based CSP methods have made significant progress. However, existing models still suffer from generation bias when facing diverse materials spaces, while systematic approaches for tracing the sources of such performance variations remain limited. To address this issue, we propose LAS (Layered Analysis Strategy), a mechanism analysis framework that identifies crystal complexity and elemental composition imbalance as two key factors affecting generation difficulty and further reveals their relationships with model generation performance. Building on these findings, we further propose DSADiff (Data Structure-Aware Diffusion), a data structure-aware diffusion model that transforms the key factors identified by LAS into a unified generation difficulty representation and uses it to jointly drive feature modeling and training strategy optimization. This enables the model to adapt its representation and generation capability to samples with different data structures and difficulty levels, thereby mitigating learning bias across samples. Experiments on three CSP benchmark datasets, including Carbon-24, demonstrate that DSADiff consistently improves both Match Rate and RMSE, with a maximum Match Rate improvement of 9%, surpassing existing state-of-the-art methods and achieving new state-of-the-art performance. This work deepens the understanding of diffusion-based CSP models from the perspective of generation bias and provides a new direction for designing less biased generative models for complex materials spaces.

0.17

Projected Penalty-Gradient Tracking for Communication-Efficient Byzantine-Resilient Federated Learning

cs.LG cs.AI ·2026-11-17 ·Yanjie Dong
benchmark
Abstract

We investigate communication-efficient federated learning over heterogeneous datasets in the presence of Byzantine adversaries capable of uploading arbitrary information. Penalty-based methods bound the influence of each worker through the bounded gradient of a penalty function; however, the bound is lost once the uploads are compressed, since a Byzantine worker can accumulate an unbounded state at the server. To retain the Byzantine resilience under compressed uploads, we propose a projected penalty-gradient tracking framework, where each honest worker uploads a compressed residual of its penalty gradient and the server projects each reconstructed state onto the known penalty-gradient ball. The framework accommodates any contractive compressor and yields two algorithms: fault-resilient proximal gradient with consistent sparsification (FRPG-CS), whose shared mask is protected by freshly drawn reserve rows, and lazily aggregated FRPG (LAG-FRPG), whose uploads are event-triggered. For smooth and strongly convex losses, we establish a finite-time bound on the expected objective gap, where the Byzantine error floor is independent of the compressor and the compression error decays faster than the optimization error. We also bound the expected number of lazy uploads and derive certificates that separate the finite-penalty model bias from the optimization error. Experiments on MNIST and Fashion-MNIST show that the proposed algorithms attain the highest accuracy among twelve methods under the Min-Max attack, whereas a compressed-momentum benchmark and a local-update benchmark remain preferable in attack-free accuracy and total traffic, respectively.

0.16

Which Past Matters? Training History and Multi-Forecast Horizon in Equity Transformers

cs.LG q-fin.STq-fin.CP ·2026-11-17 ·Chariton Chalvatzis
Abstract

More historical data can improve estimation while introducing relationships that transfer poorly to current market conditions. We therefore examine how training quantity and historical selection interact with forecast horizon in intraday equity prediction. We deploy twelve configurations which combine four sampling and validation policies with 6, 12 and 24 month training quantities under a fixed cross-sectional Transformer and feature set. A monthly walk-forward experiment produces 264 refits and eight cumulative return forecasts from 5 to 120 minutes. Evaluation covers January 2025 through 20 October 2026 and uses cross-sectional Spearman information coefficients against future five-factor residual returns. Six-month recency sampling leads individual models at 5-15 minutes, while multi recency configurations lead at 20-30 minutes and uniform configurations at 60-120 minutes. Increasing training quantity does not uniformly improve accuracy while an equal-weight ensemble of all twelve models exceeds every individual mean at 15, 20 and 25 minutes, but not at the remaining horizons. These findings support evaluating historical selection jointly with forecast horizon without identifying a universal policy or a causal benefit from recency.

0.16

Similar Is Not Useful: Selecting In-Context Demonstrations by Closed-Loop Utility

cs.LG ·2026-11-17 ·Jiachen Li, Zhigen Zhao, Shihao Li, Dongmei Chen, Long Chen +1 more
Abstract

Retrieval-augmented vision-language-action (VLA) policies fill their context with the demonstrations whose frames look most like the current observation under frozen DINOv2 embeddings. We ask whether visual similarity selects the demonstrations that make the policy succeed. For a frozen retrieval-augmented VLA (RICL on $π_0$-FAST, LIBERO), we make the demonstration pool swappable at run time and score each demonstration's utility by regressing the closed-loop success of 60 fixed-size pools on their membership. On initial states disjoint from those used for scoring, pools of the five highest-utility demonstrations succeed on 66.4% of 760 paired episodes across 19 tasks, pools of the five most similar on 55.7%, and random pools on 55.5%. Utility beats similarity by 10.8 points ($p=3.2\times10^{-6}$; positive on 15 of 18 untied tasks), including on 12 tasks the policy was never primed on, and similarity-selected pools do no better than random ($+0.1$ points). Within a task, a pool's visual similarity to the test scenes tracks its success at Spearman 0.10; across tasks, similarity separates tasks whose scenes differ and falls to chance when two tasks share a scene. The gain comes from on-task closed-loop evidence: best-of-ten direct search over pools recovers two-thirds of it with 100 rollouts. Against the shipped 20-demonstration pool, utility-curated pools of five lead by 7.0 points over 560 paired episodes from two evaluation runs ($p=0.006$), and utility-ordered retrieval inside the full pool leads by 4.5 points ($p=0.055$).

0.16

Communicate the Change: Efficient Expert Parallelism via Activation Compression

cs.LG cs.AIcs.DC ·2026-11-17 ·Fengyuan Liu, Jay Gala, Siva Reddy, Dzmitry Bahdanau
Abstract

Scaling has driven the success of large language models (LLM), with sparse mixture-of-experts (MoE) architectures enabling greater parameter capacity without a proportional increase in computation. At large scales, MoE models rely on expert parallelism to distribute experts across devices. Each MoE layer requires two synchronous collective communication steps during inference: dispatching hidden states to the activated experts and combining the resulting expert outputs. As compute throughput has outpaced interconnect bandwidth, these synchronous collectives have become a major bottleneck to MoE scaling and efficiency. We address this challenge with EchoMoE, a lightweight architecture modification that retrofits existing MoE models, requiring approximately 8,500x fewer training tokens than full pretraining. Motivated by the observation that transformer blocks update the residual stream incrementally, we use a hidden state from an earlier layer as a reference and communicate only the compressed difference between it and the current state, rather than communicating each hidden state in full. EchoMoE periodically refreshes these reference states and transmits only the compressed differences synchronously, reducing the traffic on latency-critical paths without significant computational overhead. EchoMoE reduces inter-node communication volume by 58% and improves throughput by 22.6%, while preserving average performance within 3.8 percentage points across 18 benchmarks. EchoMoE further improves inference throughput by up to 2.38x over bandwidth-constrained Ethernet, paving the way for wide-area distributed inference. EchoMoE also tolerates quantization of the communicated representations without measurable quality loss. We release an efficient implementation for applying EchoMoE to pretrained MoE models, together with an inference integration for vLLM.

0.15

Learning Stratified Policies for Inference-Time Objectives

cs.LG cs.AI ·2026-11-17 ·Mridul Mahajan, Xuezhou Zhang, Aldo Pacchiano
Abstract

Inference-time scaling improves the performance of large language models by sampling multiple responses and evaluating their collective utility. However, many existing approaches obtain these samples by repeatedly sampling from a single policy. In this work, we study if the sample budget at inference time can instead be allocated across a population of policies. We introduce stratified policies, where each of $k$ learned policies contributes one response to a $k$-sample inference-time objective. We realize this policy population using separate LoRA adapters on a shared language model backbone and optimize it for pass@$k$. We theoretically motivate the approach by analyzing the linear setting; a population of LoRAs can achieve zero best-of-$k$ approximation error in settings where every single LoRA has strictly positive prediction error. Empirically, we demonstrate that stratified policies consistently improve pass@$k$ over a single-policy baseline across multiple reasoning tasks. Overall, our results highlight how distributing samples across complementary policies can be a more effective use of sample budget than repeated i.i.d. sampling from a single policy at inference time.

0.15

Learning Optimal Transport Plans Via Autoregressive Token Regression

cs.LG cs.AI ·2026-11-17 ·Ivan Jaen Marquez, Sotirios Panagiotis Chytas, Vikas Singh
Abstract

Optimal Transport (OT) provides a principled framework for finding the minimum cost coupling between probability distributions. Classical solvers remain the gold standard for this task. At the same time, OT is an ideal testbed for studying whether learning-based amortized models can internalize the structure of OT type problems. Here, ground-truth solutions are available and marginal constraints are explicitly defined. Also, exact discrete solutions have well-characterized sparsity. We describe ToR-OT, a transformer-based model that learns to predict discrete OT plans through autoregressive token regression. We reframe OT into a structured sequence generation task. ToR-OT decodes sparse transport plans as sequences instead of solving an optimization task at test time. We focus on three properties: (a) Zero-shot generalization: pretraining on a large synthetic corpus of OT instances allows a single model to generalize to unseen distribution pairs and heterogeneous cost metrics without retraining, (b) Sparse primal decoding: by exploiting the inherent sparsity in exact discrete OT solutions, the decoder predicts only nonzero entries (avoids dense coupling outputs), (c) Amortized inference: on low-dimensional discrete problems, ToR-OT gives accurate approximate plans and compares favorably with per-instance neural OT baselines. Overall, ToR-OT is not yet a replacement for classical solvers, but a step toward understanding how autoregressive transformers can learn structure of constrained optimization problems like OT.

0.15

An Optimization Framework for Denoising Generation

cs.LG ·2026-11-17 ·Xinyu Zhou, Jiawei Zhang, Stephen J. Wright
Abstract

Noise-unconditioned denoising learns a single score, so generation can be viewed as gradient ascent on a fixed log-density landscape. This optimization interpretation leaves two questions: do trajectories approach meaningful data, and do different initializations retain diverse outcomes? We establish both phenomena for empirical and population multi-scale Gaussian-mixture objectives. Under explicit posterior and landscape conditions, ascent converges almost surely within $O(mσ_{\min}^2/τ)$ of an $m$-dimensional reference manifold with reach at least $τ$. We derive a source--sink identity for the same fixed field and use it to prove finite-resolution regional coverage. In an affine construction, we further identify a positive-bandwidth regime where even empirical-score optimization can avoid training-sample memorization while preserving diversity at the training resolution. The same fixed objective naturally supports penalty-guided controllable generation. Experiments support our predicted mechanisms, show improved robustness under large-step guidance, and demonstrate downstream gains consistent with finite-data generative generalization.

0.14

Correcting Flows with Marginal Matching

cs.LG ·2026-11-17 ·Meital Bojan, Nadav Bojan Sellam, Ev Zisselman, Aviv Tamar
Abstract

Flow matching models, ODE-based generative models, generate samples by gradually morphing a simple source distribution into a target distribution. However, these models still face challenges in optimally balancing computational efficiency with generation fidelity. While predictor-corrector schemes have been proven to help address this issue in diffusion models effectively, their application to flow matching models remains under-explored. In this work we present a novel training-based predictor-corrector framework for flow matching, which exploits the fact that the ODE can be easily reversed. Our key idea, termed Marginal Matching, is to correct samples at a particular time step in the forward ODE towards a matching marginal distribution of the backward ODE. This scheme proves to reduce a bound on the distance between the data distribution and the learned distribution, motivating two different implementations for the correction function. We show that our proposed method improves sample quality on CIFAR-10 and ImageNet-64, with minimal overhead in computation time.

0.14

PACT: Plan-Action Consistency Training by Preserving Correct Plans and Repairing Actions in GUI Agents

cs.AI cs.CVcs.LG ·2026-11-17 ·Zhuohan Cai, Zhangxuan Gu, Runze Li, Beitong Zhou, Ming Wen +3 more
benchmark
Abstract · 21 pages, 11 figures, including supplementary material

GUI agents can formulate a correct immediate plan in their reasoning yet emit an action that fails to execute it, creating an execution gap. This recurring mismatch undermines reliable task completion, as a single execution error can derail a multi-step workflow. In outcome-supervised reinforcement learning, broadcasting a negative response-level advantage across all generated tokens risks penalizing valid planning together with erroneous execution. To address this credit-assignment conflict, we introduce PACT (Plan-Action Consistency Training), a training framework that combines reference-aligned plan protection and plan-aligned action repair without altering the original scalar action reward. Using annotated reference actions to identify execution gaps, PACT protects the verified plan by masking negative policy-gradient contributions across its token span. Keeping the generated reasoning unchanged as context, it then favors the reference action over the emitted erroneous action through a preference surrogate with stop-gradient pair weights. Compared with the outcome-only baseline, PACT raises AndroidWorld task success from 60.34% to 68.97% (+8.63 pp), while improving rule-based plan-action consistency on AndroidControl High from 83.71% to 89.74% (+6.03 pp). Adapting PACT to desktop agents also improves task success from 40.92% to 44.17% (+3.25 pp) on the full 369-task OSWorld benchmark. Together, these results establish PACT as an effective credit-assignment framework for bridging execution gaps and improving task completion across mobile and desktop GUI environments.

0.14

PACD: Provenance-Aware Consensus Decoding for Mitigating Cross-Modal Hallucinations in Omnimodal Large Language Models

cs.AI cs.LG ·2026-11-17 ·Chang Sun, Yijun Tian, Zhenwei Tang, Yuchen Deng, Yifan Simon Liu +2 more
Abstract

Omnimodal large language models (Omni-LLMs) integrate audio, video, and text for multimodal reasoning, yet remain vulnerable to cross-modal hallucinations, where one modality improperly influences predictions about another. Existing training-free decoding methods mitigate these errors through modality perturbation or relevance-based weighting. However, due to interference from irrelevant modality cues during fusion and difficulty detecting shared errors, these methods may suppress correct unimodal predictions and allow erroneous consensus to persist. To address these limitations, we propose Provenance-Aware Consensus Decoding (PACD), a training-free framework for mitigating cross-modal hallucinations. PACD comprises two complementary paths based on whether the branches reach consensus. For non-consensus cases, Relevance-Guided Candidate Arbitration identifies task-relevant unimodal evidence and selectively arbitrates between predictions using alternative candidates. For consensus cases, Provenance-Based Consensus Recovery leverages the outranking principle to evaluate the consistency and strength of layer-wise modality evidence for alternative answers, selectively resolving erroneous consensus under confidence constraints. Extensive experiments on four Omni-LLMs and five benchmarks across two tasks demonstrate the superiority of PACD. Compared to existing methods, PACD achieves up to 27.6% relative improvement over standard decoding across CMM, AVHBench, and SVHalluc for cross-modal hallucination mitigation, and an average relative improvement of 7.7% over best baseline across OmniVideoBench and FutureOmni for general audio-visual reasoning.

0.14

Variance Restoration over Covariance Rank: Memory-Budgeted Analytic Continual Learning

cs.LG ·2026-11-17 ·Elena Voss, Pixel Nomand, Marcus Hale, Sofia Reyes
Abstract

Analytic continual learning replaces iterative gradient updates with closed-form ridge regression over frozen pre-trained representations, summarizing all past sessions in two running sufficient statistics: the feature autocorrelation matrix $Σ_t$ and the class cross-correlation matrix $Γ_t$. Lifting backbone features into a wide random code of dimension $M \gg D$ substantially improves linear separability across tasks, but forces $Σ_t$ to grow quadratically as $\mathcal{O}(M^2)$. Under a fixed persistent memory envelope $\Bcal$, a learner faces a fundamental structural trade-off: store exact pairwise co-activations inside a narrower feature code, or expand to a wide code while compressing cross-coordinate interactions. We observe that the coordinate self-variances $v_t = \diag(Σ_t)$ and class statistics $Γ_t$ grow only linearly with $M$, whereas off-diagonal interactions dominate memory. We introduce \method (\textbf{Co}ordinate-\textbf{Re}stored \textbf{A}nalytic \textbf{C}ontinual \textbf{L}earning), which retains $(v_t, Γ_t)$ exactly, maintains off-diagonal co-activations through a streaming rank-$q$ factor $F_t$, restores exact diagonal variances via $\Sighat_t = F_t^\top F_t + \Diag(v_t - \diag(F_t^\top F_t))$, and solves the full $M\times C$ normal equations via the Woodbury identity. Across five class-incremental benchmarks on frozen ViT-B/16, \method matches dense full-covariance readouts within $0.09$ points while cutting persistent state by $90.7\%$ ($35.9$ vs.\ $385.3$ MiB). Under equal memory envelopes of $16$--$80$ MiB, prioritizing expansion width $M$ over covariance rank $q$ outperforms exact narrow readouts in $17$ of $18$ configurations by up to $+2.19$ points.

0.13

DAQ: Decision-Aligned Post-Training Quantization for Spiking Transformers

cs.AI cs.LG ·2026-11-17 ·Jialin Chen, Xiaochen Zhao, Aili Wang
Abstract

Spiking transformers combine attention-based representation learning with sparse spike communication, yet their weights and membrane variables still incur substantial storage and computation overhead. Post-training quantization (PTQ) offers a practical way to reduce these overheads, but preserving neuronal firing under weight and membrane quantization remains challenging. In this paper, we propose Decision-Aligned Post-Training Quantization (DAQ), a method that calibrates the quantizers to align neuronal firing decisions with those of a floating-point reference. We analyze how errors from these quantizers affect firing at decision boundaries and how they propagate over time steps in recurrent neurons. Based on this analysis, we formulate an objective that encourages quantized membrane variables to remain within the reference firing regions along the quantized model's own state trajectory. Experiments across multiple spiking transformer architectures on image classification, semantic segmentation, and event-based tracking demonstrate the effectiveness of DAQ with 4-bit weights and membrane variables. On FE108, DAQ achieves 58.8\% AUC and 90.9\% PR with SDTrack, while reducing energy by 66.2\%.

0.13

ResoLUT: Distilling Residual Networks into Lookup-Table Networks

cs.LG ·2026-11-17 ·Katarzyna Fojcik, Aidas Mikalauskas, Matthias Kampa, Renaldas Zioma, Jogundas Armaitis
Abstract

Logic networks built from Boolean gates or lookup tables (LUTs) promise efficient hardware inference, but training them at depth and learning their connectivity remain difficult. We do not train logic as logic. Instead, ResoLUT distills a trained residual network (ResNet) layer by layer into LUT trees, preserving its multi-bit residual stream and BatchNorm while extracting connectivity from the teacher rather than learning it. To our knowledge, ResoLUT is the first convolutional logic network in computer vision to retain a ResNet-style multi-bit residual stream. Hard LUT decisions are used throughout training and inference, avoiding a discretization gap in the logic. ResoLUT-20 reaches 95.66% on CIFAR-10 with twice the trees, and ResoLUT-56 reaches 78.87% on CIFAR-100 and 53.69% on ImageNet at 64px, the best results of logic and lookup-table networks on CIFAR-10 and CIFAR-100, showing that LUT networks can scale in depth. We explore and quantify the role of the residuals and the tree budget and shape. The most optimized model, a ResoLUT-8 with one eighth of the trees and further optimizations, fits on an FPGA and reaches 90.04% at 189k images per second, the highest throughput among the compared models above 85% on CIFAR-10.

0.12

Storing the Gemma 4 K=V Cache Once: Exact Global-KV Memory Savings in a Paged Serving Engine with Unchanged Attention Kernels

cs.LG cs.CLcs.DC ·2026-11-17 ·Kushagra Bainsla
Abstract · 21 pages, 4 figures, 3 tables

In the global attention layers of Gemma 4 31B, the value is the normalized key projection, and the key is the value scaled by a fixed weight, with rotary position encoding on one quarter of its dimensions. The Gemma 4 report states that this design reduces the global KV cache by up to 37.5%, but the open serving engines that we checked still store a full key and a full value: 1,024 values per KV head and token. We store one 640-value row instead, exact in real arithmetic, and an unchanged paged attention kernel reads it through two overlapping strided views as its key cache and its value cache. In a vLLM prototype with the 16-bit 31B model, the outputs are bit-identical to an emulation of the same arithmetic on the stock cache, 2 of 432 final answers differ from the stock engine (within its run-to-run noise), and the steady decode KV memory per request falls by a factor of 1.34 to 1.40 at 32K to 64K tokens. On a prefill/decode pair of two A100 GPUs with NIXL transfer and CPU offload, a decode engine with a full 15 GiB pool ran 1.38 times the stock decode tokens per step for 16K-token prompts (repeated to 0.06%) and 1.24 times for 8K-token prompts; a model of the vLLM block pool predicted both ratios to within 1% (the engine was 0.7% to 0.9% higher) and matched the change between them. The eager decode throughput rose by 1.30 and 1.11 times, and on this path the compact engine gave the same output tokens as the emulation for 8 prompts at batch 1. The run-time prototype makes an eager step 4% to 9% slower at equal batch, and the method does not yet stack with an 8-bit KV cache at the default scale.

0.12

OAK-Mind: Zero-Training Mental Health Change Inference from Passive Sensing via Knowledge-Grounded Multi-Agent LLMs

cs.MA cs.AIcs.HCcs.LG ·2026-11-17 ·Yishu Ji, Zhihan Jiang, Will Ke Wang, Yuang Fan, Anind K. Dey +1 more
Abstract · Under review at ACM IMWUT. 30 pages, 4 figures and 6 tables

Wearable and smartphone sensors capture behavioral signals associated with mental health, yet inferring mental health outcomes from these signals has traditionally relied on supervised models requiring labeled data and offering limited insight into their predictions. Large language models (LLMs) offer new opportunities to reason over behavioral evidence, but existing LLM-based approaches to passive sensing have largely continued to rely on task-specific fine-tuning or sensor-specific pretraining. In this paper, we introduce \textbf{OAK-Mind}, a zer\textbf{O}-training, multi-\textbf{A}gent, \textbf{K}nowledge-grounded LLM pipeline for inferring mental health changes from passive sensing data. The pipeline organizes sensing data into behavioral dimensions including sleep, physical activity, location mobility, and smartphone use, and orchestrates specialized agents grounded in mental health literature to interpret longitudinal behavioral changes. It synthesizes their assessments and applies questionnaire-based symptom rating to infer changes relative to an individual's baseline assessment with human-readable rationale grounded in observed behavioral patterns and domain knowledge. Across datasets and prediction tasks, while without task-specific training, OAK-Mind achieves comparable or better performance than the most predictive supervised baselines trained on additional labeled samples. It also demonstrate consistent improvement upon direct prompting across four tested LLM backbones. Ablation analyses support the benefits of combining behavioral interpretation with symptom rating, distributing analysis across domain-specific agents, and incorporating domain knowledge. Our findings highlight the potential of knowledge-grounded, multi-agent LLMs for training-free inference of mental health changes from passive sensing data.

0.12

An AI-Based OSSE Framework to Evaluate Learned Physical Consistency: A Multi-Model Study Using FourCastNet, Aurora, Prithvi WxC, and GraphCast

cs.LG ·2026-11-17 ·Yash Diggikar, Milton Halem, Aishwarya Raman
Abstract

Observing System Simulation Experiments (OSSEs) are a central methodology for evaluating future atmospheric observing systems. We present the first AI-based OSSE framework to test whether different AI weather models can reproduce comparable physical impact from new observations that traditional physics based models generate. We evaluate four distinct architectural AI models on the Charney conjecture that incomplete historical observations can infer the present state of the atmospheric. We considner- FourCastNet V2, GraphCast, Aurora and Prithvi WxC and integrate each autoregressively over a 60-day forecast horizon. After a 30-day spin-up, we insert ERA5 temperature fields to infer winds and related atmospheric fields. Similar OSSE experiments are conducted for wind insertion. We measure the performance of each model's response to observational information on a 30-day data assimilation cycle forecast with and without data insertion. Distinct architectural AI weather models respond differently to the same OSSE observations. Foundation models (i.e. Prithvi WxC and Aurora), produce the strongest impact. For example, while GraphCast is the weakest, all three of the four models capture the temperature-to-geopotential height response. The two foundational models retain a positive, non-trivial, tropical zonal wind at 500 hPa response to temperature insertion. Various data insertion experiments can vary significantly even within a given architecture. We identify an architecture-specific numerical instability in GraphCast's full 37-level configuration that produces a runaway feedback concentrated at the uppermost levels with ERA5 temperature insertion correctable by excluding levels at or below 100 hPa.

0.11

The Price of Hard Coverage Constraints in Online Conformal Prediction

cs.LG math.STstat.ML ·2026-11-17 ·Mingrui Liu
Abstract

Conformal prediction provides distribution-free uncertainty quantification, but in sequential deployment, standard coverage guarantees can still allow substantial fluctuations in the number of miscoverages over time. We study an online setting in which prediction sets are produced one round at a time and the cumulative number of miscoverages must stay within a prescribed budget throughout deployment.Such a hard coverage guarantee can be enforced by enlarging prediction sets, but how much additional prediction-set size is fundamentally necessary remains unclear. We characterize this unavoidable efficiency cost over $T$ prediction rounds. Remarkably, the cost persists even with i.i.d.\ data, a known score distribution, and perfect knowledge of the population-optimal prediction set. Under suitable regularity conditions relating coverage and prediction-set size, we determine the minimum cumulative excess prediction-set size required by any online method satisfying the hard coverage constraint. This minimum cost is $Θ(\log T)$ for expected cost at a fixed time known in advance, $Θ(T^{1/3})$ for the largest expected cost over all fixed times up to $T$, and $Θ(\sqrt{T})$ for the expected worst realized cost when the evaluation time is chosen in hindsight. We further show that these rates can be attained using past scores, without knowing the score distribution. Thus, even perfect statistical knowledge does not make sequential reliability free.

0.10

OICycle: Observe, Infer, and Intervene Cycle for Belief-State Learning in Longitudinal Clinical Decision Making

cs.AI cs.LG ·2026-11-17 ·Jitindra Fartiyal, Pedro Freire, Sergei K. Turitsyn, Sergei G. Solovski
benchmark
Abstract

Clinical care is a closed loop: clinicians observe partial evidence, infer the patient's state, and intervene, and each intervention determines what is observed next. This makes clinical decision making a partially observable Markov decision process (POMDP), in which a belief over a latent patient state must be updated from irregular, multimodal, and incomplete records. Yet most deep clinical models predict from a single snapshot or from coarsely aggregated history, while POMDP-based clinical methods need hand-specified discrete state spaces and transition models that real electronic health records cannot supply. We introduce OICycle (Observe, Infer, and Intervene Cycle), which learns a stochastic belief state directly from retrospective trajectories. At every episode, a learned recurrent operator fuses a sample of the previous belief with the new multimodal observations and the intervention that produced them. The resulting Gaussian belief is trained to predict clinicians' next interventions, the time to the next episode, and next-episode mortality. To enable reproducible evaluation, we release a benchmark derived from MIMIC-IV, MIMIC-IV-Note, and MIMIC-CXR that organizes 72,075 hospital admissions (5.4M laboratory, microbiology, radiology, prescription, and note records) into temporally ordered therapeutic--diagnostic episodes. Compared with an otherwise identical history-free model, OICycle improves next-intervention adaptive recall from 0.253 to 0.347, time-to-next-episode macro-F1 from 0.293 to 0.351, and next-episode mortality macro-F1 from 0.585 to 0.649. Analysis shows that the learned belief encodes intervention type and trajectory position and evolves smoothly across episodes, and that the benefit of history is concentrated in interventions whose ordering depends on prior findings.

0.10

Bandit Online Submodular Minimisation Using Zeroth-Order Oracles

cs.LG stat.ML ·2026-11-17 ·Amir Ali Farzin, Philipp Braun, Iman Shames
benchmark
Abstract

We consider online submodular minimisation under bandit feedback, where at each round a decision maker plays a subset and observes only its cost, while the cost functions, and hence their minimiser, change over time. Existing bandit methods only guarantee static regret, which compares against a single fixed subset and is a weak benchmark when the minimiser moves. This paper provides, to the best of our knowledge, the first method with guarantees for tracking a moving minimiser in this setting. We propose a zeroth-order method built on a two-point oracle for the Lovász extension whose two evaluations are both subsets that can be played. The perturbed and base points are rounded by a common random threshold, and drawing the smoothing direction from the unit sphere makes the base evaluation a zero-mean control variate, so the oracle is an unbiased estimator of the smoothed gradient. We establish dynamic regret bounds of order $O\big((1+\sqrt{P^\ast_N})N^{3/4}\big)$ in general and $O\big((1+\sqrt{P^\ast_N})N^{2/3}\big)$ under a sufficiently small drift, where $P^\ast_N$ is the path length of the minimiser sequence, together with the corresponding $O(N^{3/4})$ and $O(N^{2/3})$ static regret bounds. Numerically, the method attains lower regret than the algorithm of (Hazan and Kale (2012)), with the advantage growing with the path length.

0.08

Get Hint-Augmented GRPO Done Right Through Activation Steering

cs.LG ·2026-11-17 ·Khiem Le, Ting Hua, Nitesh V. Chawla
Abstract

GRPO has become the dominant reinforcement learning method for eliciting reasoning capability in LLMs using verifiable rewards. However, it suffers from a fundamental failure mode: when a training question exceeds the base model's current capability, all generated responses fail and receive zero reward, yielding no gradient signal and stalling further capability expansion. An emerging class of GRPO extensions, termed Hint-Augmented GRPO (HA-GRPO), adds a "hint" to the context to reduce the difficulty of the question, thereby producing a usable gradient signal. Nevertheless, existing HA-GRPO methods share a key issue: no hint is available at inference time to append to the question, creating a training-inference context mismatch that leads to inconsistent performance gains. In this work, we reinterpret the function of hinting -- not as reducing question difficulty, but as steering model behavior -- and propose ASH (Activation Steering for Hinting), a novel HA-GRPO method built on this view. During training, the model is presented with both hinted and plain questions, while a trainable steering vector is optimized to widen the margins between the model's generation probabilities for responses to the hinted version relative to the plain version, thereby capturing the question-agnostic component of the behavioral shift induced by the hint. During inference, the trained steering vector is directly injected into the model's activation, reproducing the hint-conditioned behavior without requiring any hint in the context. Experiments with four LLMs (Qwen3-1.7B/4B-Base, Llama-3.2-1B/3B-Instruct) show that ASH consistently improves pass@$k$ on competition-level mathematics benchmarks (AMC, OlympiadBench, AIME24, AIME25) and out-of-distribution benchmarks (Minerva, GPQA-Diamond). Notably, it improves average pass@8 of Qwen3-1.7B-Base and Qwen3-4B-Base by \textbf{2.60}\% and \textbf{2.21}\%.

0.08

LoCaR: Localized Canvas Retrieval With Frozen Vision-Language Model

cs.AI cs.CVcs.LG ·2026-11-17 ·Rahim Hossain, Dipanwita Saha, Ashirbad Mishra, Xinxin Shu, Svetlana Strunjas +1 more
benchmark
Abstract · Zero-shot retrieval , Local Retrieval

Whole image embeddings can under represent small objects and local attributes that distinguish relevant images for localized queries. We present LoCaR, a training free framework for localized text to image retrieval with frozen vision language models. LoCaR constructs and utilizes independent proposals (SAM2) focusing on the object so as to get a more localized representation that independently encodes each region using the frozen visual encoder. We further study paired-support adaptation as an optional residual that uses matched text region exemplars without updating the underlying encoders. We also introduce LOCFG, a Visual Genome derived benchmark designed to evaluate localized retrieval for smaller and more challenging referents. Across several vision language backbones, our results show that improving the region representation itself is the main source of gain.

0.06

Sample-efficient Personalization of Large Language Model Alignment with Multidimensional User Preference Data

cs.LG ·2026-11-17 ·Zhen-Yu Zhang, Jerry Huang, Gang Niu, Takashi Ishida, Masashi Sugiyama
Abstract

Preference alignment has become a central tool to align large language models (LLMs) with human values and desired behavior, and its online variant can further advance their capabilities beyond what static data alone offer by repeatedly generating and labeling new responses with human feedback. But costly, time-consuming annotations limit its scalability, making sample efficiency essential, and existing methods typically seek to improve it by expanding coverage of the response space or selecting pairs with the highest uncertainty for comparison. Although they are effective, these methods quantify uncertainty from the perspective of the LLM, which may differ from the uncertainty perceived by the user: a pair can appear highly uncertain to the LLM even when past comparisons from the user already imply a clear ordering. In this paper, we propose multidimensional preference-guided online direct preference optimization (MPG-DPO), which leverages low-cost multidimensional evaluations from automated verifiers on the user historical comparisons to construct a personalized preference cone guiding response selection for online preference alignment. Specifically, it first discards responses that are confidently dominated under the estimated user preference, then selects two responses that are competitive under the retained preference cone yet exhibit the strongest unresolved disagreement. Theoretical analysis shows that, with high probability, the fraction of queries spent on cone-dominated responses vanishes as learning progresses. Experiments across multiple LLM families show that MPG-DPO outperforms strong contenders under matched query budgets.

0.04

Language, Vision & Media 26

Do Diffusion Models Learn to Generalize Basic Visual Skills?

cs.CV cs.LG ·2026-11-17 ·Amish Sethi, Boya Zeng, Wenhao Chai, Zhuang Liu
publishedcode/data
Abstract · NeurIPS 2026. Project page: https://amishsethi.github.io/visual-generative-lab/

Diffusion models generate high-quality images, but how well they generalize is hard to measure. Current evaluations rely on human preference and prompt-following benchmarks for text-to-image models and on distribution-level metrics such as Fréchet Inception Distance (FID) and Inception Score (IS) for class-conditional models, but these cannot tell whether models learn basic visual skills or largely match training patterns. To address this gap, we introduce the Visual Generative Lab (VGL), a controlled experimental framework for studying how diffusion models learn and generalize basic visual skills, including size, position, rotation, count, color, shape, and their compositions. By generating synthetic training data for each skill from explicit rules, VGL allows precise measurement of generalization with task-specific, rule-based metrics. Training models from scratch on these data reveals three patterns in this controlled setting: (1) with our default design, extrapolation is limited: size and position degrade within a few pixels of the training range, count reaches only about 50% accuracy on out-of-range counts, and rotation extrapolation accuracy varies widely across seeds; (2) for compositional generalization, the number of distinct basic visual skill combinations matters more than dataset size; (3) skills mostly fail independently, except position. Code and data-generation scripts: https://github.com/AmishSethi/visual-generative-lab.

0.49

Runtime Verification of Multiple Natural Language Criteria for Agent Governance

cs.CL cs.AIcs.LG ·2026-11-17 ·Silviu Pitis, Parand A. Alamdari, Jessica Tang, Toryn Q. Klassen, Sheila A. McIlraith
top venuebenchmark
Abstract · Accepted at NeurIPS 2026

AI agents are governed by hundreds of natural language criteria, yet existing verifiers cannot efficiently evaluate an agent's output against a large set of such criteria and report which are satisfied, violated, or inapplicable. We introduce the VFM, a tree-attention verifier that encodes the shared context and target text, and scores N natural language criteria in parallel, returning independent per-criterion ternary verdicts. The VFM is pretrained on CriteriaBank, a corpus of 394K (context, target, criterion, verdict) tuples spanning over 321K distinct criterion strings. A finetuned VFM-4B outperforms GPT-5.4 on a synthetic governance benchmark. On the DynaBench multi-criterion benchmark, it outperforms DynaGuard-4B on trace-level failure detection, and its per-criterion verdicts localize the violated rule and support focused escalation to a larger chain-of-thought judge. We show that CriteriaBank pretraining improves cross-domain transfer, particularly in the low-data regime, while matching a strong direct-finetuning control at full data. On a single H100, VFM-4B evaluates 1,000 criteria against a 4K-token context in under 2s at <11 GB VRAM, making it suitable for runtime verification.

0.45

Efficient Self-Supervised Adaptation of 3D Abdominal Vision-Language Model for Institution-Specific HCC Classification via Full Fine-Tuning and PEFT

cs.CV cs.LG ·2026-11-17 ·Febryan Putra Kartika, Cheng-Yu Ma, Ying-Jia Lin, Chi-Tung Cheng, Kuan-Fu Chen
published
Abstract · 25 pages, 12 figures, 7 tables, including appendices. Corrected author revision of the MIDL 2026 paper. Revised preprocessing yields lower H…

Medical vision-language models (VLMs) capture cross-modal relationships between image and text, yet their adaptation to institution-specific clinical tasks remains underexplored. In this study, we fine-tuned a pretrained 3D medical VLM for hepatocellular carcinoma (HCC) classification using paired abdominal CT scans and radiology reports from a different institution with acquisition characteristics that differ from the model's original pretraining corpus. We compared two adaptation strategies: full fine-tuning and parameter-efficient fine-tuning (PEFT), motivated by the common use of PEFT to reduce computational cost and enable adaptation under limited-data constraints. Both approaches achieve strong downstream HCC classification performance despite the cross-institutional domain shift, with PEFT reaching an AUC of 0.79 and F1 of 0.70, and full fine-tuning achieving an AUC of 0.82 and F1 of 0.71. Although these results remain below the performance reported by supervised HCC classification approaches, our method achieves improvements over the pretrained model through CT-report contrastive learning without requiring lesion-level annotations or HCC classification labels as training targets. Full fine-tuning converges rapidly but overfits within a few epochs, whereas PEFT (ConvLoRA for the image encoder and LoRA for the text encoder) attains comparable performance while updating approximately 1% of the model parameters, although requiring more training steps. We also examine the role of contrastive temperature, observing that temperature initialization significantly affects classification performance. This study demonstrates that a 3D medical VLM can be efficiently adapted to institution-specific HCC classification using self-supervised CT-report contrastive learning, while highlighting the practical trade-offs between full fine-tuning and parameter-efficient fine-tuning.

0.42

Multimodal Masked Diffusion Autoencoder for Modeling Impaired Driving

cs.CV cs.LGcs.RO ·2026-11-17 ·John Gideon, Anshul Gupta, Megan Applegate-Kenton, Kayli Battel, Todd J Rowell +5 more
dataset
Abstract · Equal contribution: John Gideon and Anshul Gupta. 8 pages, 6 figures. Submitted to CVPR 2027

Human cognitive state (e.g., distraction, stress), gaze, and behavior interact with the driving environment in subtle ways that AI could use to assist drivers, but ground-truth labels for human cognitive state levels are costly to obtain at scale. We propose a multimodal masked-diffusion autoencoder, ID-MAE, that adds a self-supervised objective to impairment prediction: alongside predicting the impairment label as a masked token, it reconstructs masked gaze and trajectory. To probe the effects of different modalities and types of impairments on model inference, we introduce DIVERT, a driving-simulator dataset containing 103 hours of driving data (n = 109) with ten impairment inductions spanning cognitive, visual, and emotional states, recorded with gaze, vehicle state, and scene video. We demonstrate how ID-MAE affords significant improvements in impairment inference results on both the DIVERT dataset and existing datasets, and enables strong results in additional tasks, such as trajectory and gaze prediction. For location-matched pairs, our model achieves a paired AUC of 0.706 for impairment detection on DIVERT, compared to 0.590 without reconstruction loss, improving over existing baselines. We probe the design space in terms of input modalities, masking scheme, positional encoding, attention mechanism, and decoding. We also investigate the interaction between various impairments in training and inference, demonstrating that visually demanding distractions are separable from cognitive distractions. Code and data will be available upon acceptance.

0.42

Sub-Meter Canopy Height Synthesis from Sentinel-2 via Anchored Residual Flow

cs.CV cs.LG ·2026-11-17 ·Kiarie Ndegwa, Tony Chang, Andreas Gros, Vincent A. Landau, David D. Diaz +3 more
Abstract · 30 pages, including appendices

Canopy height models (CHMs) - per-pixel maps of vegetation top height - are essential for forest carbon accounting, wildfire fuel mapping, and tree-level ecology, yet sub-meter CHMs require airborne LiDAR refreshed on multi-year national cycles. We produce 0.5 m CHMs from free 10 m Sentinel-2 imagery, a 20$\times$ cross-modal super-resolution task where optical reflectance must predict geometric height it does not directly observe. Motivated by the perception-distortion trade-off, we derive two operating points with shared conditioning. A deterministic regressor (Anchor) targets low distortion; a conditional flow matching model, trained on the Anchor's residual in a learned CHM latent space (AR-flow), targets perceptual fidelity. Trained on 197k LiDAR-paired tiles spanning 69 ecoregions across the contiguous United States and evaluated on 2,530 held-out tiles against eight baselines, the Anchor achieves 2.83 m MAE using only free satellite input. The AR-flow reduces sliced Wasserstein distance from the Anchor's 1.13 to 0.208, at a higher MAE of 3.68 m. Crown-scale statistics improve, but strict individual-tree detection remains weak: generated sub-pixel detail represents plausible structure rather than measured individual trees. Quarterly continental refresh becomes tractable, letting satellite monitoring inform timing and extent for more expensive data acquisitions.

0.38

Déjà View: A Shared Working-Memory Model for Visual Tasks

cs.CV cs.LGq-bio.NC ·2026-11-17 ·Mranmay Shetty, Gabriel Kreiman
hot direction
Abstract · 6 figures; includes appendices

A working-memory model must remember stimuli and use its memory flexibly to perform different tasks. Here we introduce a model of working memory for images that combines dynamic activity with temporary associations between image features. We train the model on five tasks, including remembering a sample while other images are presented, categorizing distractor images while remembering the sample, and determining whether an image is new or has been presented before. The same trained model performs all tasks, guided by explicit cues about which response is required. We use this flexible memory model to ask how delay, distraction, and image history affect performance. Delays have little effect on sample-memory performance for the delays we tested, while seeing more images leads to more false alarms even when the target image appeared equally recently in all cases. On the sample-memory-plus-categorization task, the probability of correct responses on all questions decreases as more questions are asked, despite robust sample-memory performance. Finally, the model maintains stable performance across 40 brief trials of each task under the tested conditions. Together, these results demonstrate how a single working-memory system can be flexibly used to perform different tasks, and how its errors depend on both previous images and task demands.

0.35

OfficePrism: Benchmarking Agents on Single-Domain Operations Across Enterprise Business Functions

cs.CL cs.AIcs.LGcs.SE ·2026-11-17 ·Hanfeng Liu, Zhenyu Luo, Dongchao Zhang, Zexi Liu, Zulong Chen +3 more
benchmark
Abstract

Broad workplace scores can mix competence within one business function with the extra cost of coordinating across functions and systems, making failures hard to localize. OfficePrism evaluates this prerequisite through 420 single-domain enterprise tasks spanning seven business functions under shared, benchmark-defined L1--L3 task-structure levels. Each task runs in a fresh environment combining a shared synthetic organization with one function-specific domain and yped state access. Our pipeline uses grounded records, business rules, and tool schemas to jointly construct a Chinese-language ticket, a hidden executable reference witness, and an expected response, followed by automated filtering and human curation. The same task set is evaluated through full-schema and search-and-dispatch packages. Across 124 model--harness--package configurations, the per-case joint natural-language-and-database pass rate ranges from 22.6% to 76.4%, with a median of 67.9%. L1 is highest in all 124 configurations, and L2 exceeds L3 in 122. Among 50 configurations sharing the same harness and model labels across both packages, search scores higher in 30, ties in one, and scores lower in 19.

0.34

LoopZip: Visual Token Compression for Efficient Looped Vision-Language Models

cs.CV cs.LG ·2026-11-17 ·Weizhong Huang, Zhicheng Liu, Yan Zhang, Fei Chao, Xiawu Zheng +1 more
code/data
Abstract · 11 pages, 10 figures

Looped vision-language models (VLMs) deepen computation by reusing shared layers, approaching larger models' capabilities with fewer parameters and easing the storage and deployment burden of stacking independent layers. However, repeatedly processing long visual sequences still incurs substantial compute and GPU memory costs. Visual token compression reduces these costs by shortening sequences. Existing methods mainly select original tokens directly or combine selection with aggregation. In looped models, the former may overlook changing token importance across calls and discard information needed later; the latter lack joint consideration of output changes across calls, limiting information preservation. We propose LoopZip, a training-free visual token compression algorithm addressing both issues. Visual aggregation builds a compact candidate pool before the looped backbone, and cross-invocation selection averages instruction attention from the first two existing calls of the same shared attention layer to retain tokens needed by both states. Weight-tied value correction then jointly fits shared visual-value adjustments to both pre-selection layer outputs, reducing output deviations across states. To validate LoopZip's effectiveness, we conduct extensive experiments covering two open-source looped VLMs, eleven benchmarks, and three pruning ratios. On LoopVL-1B, LoopZip retains 97.16% of the original model's performance at 50% pruning ratio and exceeds the strongest baseline by 5.36% at 87.5% pruning ratio. Across Hugging Face Transformers, vLLM, and SGLang, LoopVL-1B with LoopZip reduces peak allocated GPU memory by 58.36%-58.98% and achieves 2.01-2.47x end-to-end speedups over the original model.

0.33

Social Signal Bench: Evidence-Weighted Evaluation of Co-occurring Social Signals in Video and Audio

cs.CV cs.LG ·2026-11-17 ·Siddharth Ravi, Anders R. Bargum, Hugo Anselme, Patrizio Cugia di Sant'Orsola, Alberto Cardaci +3 more
benchmark
Abstract · 25 pages, including appendix

Recognizing social signals such as interest and hesitation requires interpreting context-dependent behaviors that may co-occur in conversation. Existing affective computing benchmarks rarely evaluate such co-occurring signals, and conventional multi-label F1 overlooks semantic relationships and annotator disagreement. We introduce Social Signal Bench, a 47-hour densely annotated benchmark for the evaluation of social signal recognition and engagement, comprising two datasets: Social Signal Bench Video (25 hours), covering ten social signals in audiovisual recordings, and Social Signal Bench Audio (22 hours), covering six signals in audio-only recordings. Each dataset pairs a public training set of online conversational video, of which 10 hours are expert-annotated and the remainder pseudo-labeled, with a private expert-labeled test set spanning four conversational domains. We also propose Evidence-Weighted F1 (EW-F1), which weights reference labels by annotator evidence and assigns graded credit to semantically related predictions through optimal per-clip matching. Controlled perturbations demonstrate that EW-F1 distinguishes semantically related errors from unrelated ones, unlike conventional F1. Under a shared evaluation protocol, 7B open models post-trained on Social Signal Bench outperform general-purpose frontier models on video and match them on audio.

0.33

Spool: Low-Latency Streaming Inference for Hybrid Vision-Language Models

cs.CV cs.LG ·2026-11-17 ·Rui Ning, Om Chabra, Hari Balakrishnan
benchmarkcode/data
Abstract

Applications such as visual accessibility agents and robotics require answering questions about recent video content promptly, over a bounded window of past frames (e.g., the last two minutes). Hybrid vision-language models (VLMs), which interleave attention with linear recurrent layers such as Gated DeltaNet (GDN), are attractive for this setting, but maintaining a sliding window is expensive. The reason is that recurrent layers combine all past frames into a shared state, so expiring old frames normally requires re-running the entire window through the model. We observe that the recurrent update is affine in the state with input-dependent coefficients, so the effect of each frame can be captured as a transformation that composes associatively. We present Spool, a training-free framework that caches per-frame state transformations and composes them efficiently as the window advances, removing expired frames from each recurrent layer's state without replaying frames through the backbone. Removal is exact within each layer; as with key-value eviction in attention, expired frames can retain residual influence through the cached inputs computed by earlier layers. We design a dedicated data structure and GPU kernels to achieve low caching and composition overhead. Across five open-source hybrid VLMs and four video reasoning benchmarks, Spool achieves the lowest p95 time-to-first-token in all 20 model-benchmark tests, reducing it by a geometric mean of $6.6\times$ over the faster of two recent vLLM baselines, while improving accuracy by 2.3 percentage points on average.

0.32

What Does a Bias Head Measure? Non-Identifiability in Head-Level Bias Attribution

cs.CL cs.LG ·2026-11-17 ·S. M Seefat Alam
Abstract · 18 pages, 13 tables. Under review at ACL Rolling Review (October 2026 cycle)

Methods that localise social bias in large language models to individual attention heads and remove it by ablation are validated by their effect on a fairness metric, never by evidence that the statistic selecting the heads measures bias. We audit this on Qwen-2.5-7B-Instruct and Llama-3.1-8B-Instruct. The head map correlates at $ρ= -0.950$ and $-0.768$ with a label-blind response-length map, apparently identifying a surface confound. An intervention that breaks the length-label coupling rejects this reading: with the prompt held fixed, response length explains 1% of the statistic's variance. What does explain it is prompt identity, at 64% of variance, in an ordering that is not monotonic in length. Within any fixed prompt the fairness label, response length and the presence of a protected-group token are collinear at $φ= 0.95$, partly by definition of the labelling rule, so no observational analysis can attribute the map to any of them, and no method in this family performs an experiment that could. The substantive claims fail independently: bias selectivity is not late-layer-concentrated on either model under either reference reading, and no head set we tested reduces the unfairness rate when ablated. Its complement, scored on the opposite label, reliably destroys refusal behaviour ($p = 2.7\times10^{-5}$), demonstrating that the pipeline detects effects at this scale.

0.31

Sample-Aligned Decoders for Latent Generative Models

cs.CV cs.LG ·2026-11-17 ·Yu Lei, Yi Li, Hongcheng Gao, Yipeng Wei, Liang Li +2 more
code/data
Abstract · Project page: https://torpedo2003.github.io/SAD/

Latent generative models synthesize images by decoding generated latents, yet their decoder is trained only to reconstruct encoder latents and never sees the latents it decodes at test time, so the residual errors of the generator are rendered directly into the final image. Existing remedies retrain the generator, add cost at every sampling step, or make the decoder robust to perturbed real latents, assuming that generator errors resemble such perturbations. We find that they do not. Generated latents lie on the encoder latent manifold as closely as real ones, yet follow a different distribution. The decoder should therefore be adapted to the distribution of generated latents, which raises two challenges: (1) unpaired supervision, since a generated latent has no ground-truth image, and (2) latent-space drift, since a decoder freed from paired targets forgets the latent space it must decode. We propose Sample-Aligned Decoders (SAD), a post-training framework that fine-tunes only the decoder of a frozen generator at no inference cost. Generated-Sample Distribution Alignment matches decoded generated latents to unpaired real images in a frozen foundation-model feature space, and Encoder Anchor Preservation keeps the decoder close to the original one on encoder latents. Our analysis shows that the alignment is effective only on generated latents, where it corrects each sample in proportion to how much more often the generator produces such latents than the encoder; on encoder latents it can only work against the anchor. SAD improves four state-of-the-art ImageNet generators across three tokenizers, e.g., FID-50K from 1.15 to 0.98 for REPA-E, complements discriminator guidance, and extends to text-to-image generation. Our code and models are available at https://torpedo2003.github.io/SAD/.

0.30

When Anchors Shift the Target: Advantage-Gated Adaptation for Diffusion Reinforcement Learning

cs.CV cs.LG ·2026-11-17 ·Pu Cao, Min Zhou, Ying Shu, Tiezheng Ge, Qing Song +1 more
Abstract · 12 pages, 4 figures, including supplementary material

Forward-process reinforcement learning adapts diffusion and flow-matching models by re-noising rewarded generations into regression targets. Because negative advantages induce concave updates with unbounded repulsive gradients, existing methods stabilize training using cached predictions from a moving policy copy, yet these anchors also alter what positive feedback teaches the model to fit. We formalize this conflict as anchor coupling: after separating frozen-reference regularizers, affine moving-anchor rules cannot eliminate negative curvature while preserving every positive regression target. We propose Advantage-Gated Adaptation (AGA) to decouple these roles. Positive samples retain standard target fitting, whose self-limiting update vanishes at the observed target independently of the moving anchor. Negative samples instead use the first-order tangent of the concave objective at the moving-policy prediction, preserving the local repulsion direction with a prediction-independent gradient. Within the quadratic-anchor family, these requirements uniquely determine the parameter-free gate $ω(a)=\max(-a,0)$. Because the moving copy provides local linearization rather than retention, AGA assigns long-horizon behavior preservation to a frozen pretrained reference regularized at its operating classifier-free guidance scale, interpreted through an implicit-reward lens. Across compositional generation, text rendering, and preference alignment, our experiments show that displacing positive targets degrades visual quality and heightens sensitivity to policy-copy schedules, whereas AGA consistently improves the reward--quality trade-off.

0.29

Propose with the Gradient, Accept with the Loss: One Hedged Attack for Pixel-Sparse Robustness

cs.CV cs.CRcs.LG ·2026-11-17 ·Hon Ming Yan, Chun Pong Lau
Abstract · 22 pages, 13 tables, 4 figures

A pixel-sparse adversary may change a few pixels of an image to any value. Robustness to such attacks is currently measured with an ensemble, sparse-AutoAttack (sAA), because its members fail in different places: sparse projected gradient descent (sPGD) trusts a first-order model that a full pixel change breaks, and sparse random search (Sparse-RS) ignores the gradient even where it is informative. We show that both attacks are local searches, each a proposal distribution plus an acceptance rule, and ask what such a search can learn from the gradient. A Taylor bound says what to measure, and measurements at the states the search visits give five rules. The sign of the gradient is its most reliable use at a single pixel and sets a new pixel's value. Its ranking of candidate pixels should be followed loosely, and its ranking of the pixels to remove, which is often more reliable, can be followed more closely. A pixel should not be proposed twice before a move is accepted, and acceptance should be decided by the loss alone. These rules define Gradient-guided Random Search (GRS), which runs once and costs one forward and one backward pass per step. On 13 CIFAR-10 configurations and 5 ImageNet models, GRS matches or beats the best single member of sAA and is within one point of sAA itself or stronger on 17 of 18 settings, and it is significantly stronger than sAA on 8. It uses 3 instead of up to 7 forward-pass units, and on CIFAR-10 it runs 3.7 to 5.0 times faster than the members of sAA run in full.

0.29

MoE--LoRA Methods for Multilingual Adaptation: A Systematic Empirical Study

cs.CL cs.LG ·2026-11-17 ·Nikit Srivastava, Joel Dag, René Speck, Axel-Cyrille Ngonga Ngomo
published
Abstract · To Appear - EACL 2027 (Findings)

Routed low-rank adapters offer a parameter-efficient means to improve the performance of large language models (LLMs) across tasks. Hence, it is assumed that they can also be used to improve the support of low-resource languages in LLMs, which currently serve English far better than the thousands of other languages spoken worldwide. However, this assumption has never been verified empirically under a shared multilingual evaluation protocol. We address exactly this gap by comparing eight low-rank adaptation (LoRA) methods on a frozen base model using 200 unique languages in more than 9{,}000 evaluation settings. Our results suggest that a reduced-scale search can identify viable configurations of LoRA methods but does not reliably predict their relative performance, as the top-ranking configuration is never best at scale. LoRA reliably improves language-model fit on translation data, consistently degrades reading comprehension, and benefits low-resource languages most. These findings suggest that LoRA methods should be regarded as targeted tools rather than general-purpose means for improving the multilingual capabilities of LLMs.

0.27

Is model self-assessment enough for diverse LVLM ensembles?

cs.CV cs.LG ·2026-11-17 ·Michael Smith, Frank P. Ferrie
Abstract

Ensembles of diverse Large Vision Language Models have demonstrated performance beyond that of single models across multiple datasets and tasks. However, the necessary aggregation process usually relies on training, parameter fitting, or complex task-specific pipelines. Furthermore, evaluation is limited, making a fair comparison difficult. We create a unified pipeline where we perform candidate response selection: we build an ensemble by selecting the best answer from a set provided by different models. We comprehensively evaluate across 34 different means of confidence estimation, 24 ensemble configurations, 20 datasets, and two prompt specifications. We show that simple training and parameter-free approaches can be sufficient, even for ensembles of different architectures. These simple self-assessment approaches rely solely on each constituent model's confidence, without explicit calibration, yet are able to perform better than said constituent models. They are also competitive with more involved alternatives on standard metrics such as accuracy, including on open-ended problems. However, we also discover that dataset, ensemble composition, and prompt wording are key confounding variables insufficiently explored by prior work that significantly affect ensemble performance, such that the ensemble can perform worse than its best constituent model. These must be optimized for to get the most out of ensembles, at the cost of limiting their generalizability. We recommend that practitioners consider training-free ensembles only if deployment conditions can be precisely characterized, with Best-of-N configurations of strong models being a good starting point.

0.27

Zero-Dimensional Attention Topology for Linguistic Acceptability

cs.CL cs.LG ·2026-11-17 ·Saeed AlMarri
Abstract · 10 pages, 4 figures, 4 tables. Code and data are provided as ancillary files

Topological summaries of Transformer attention provide sentence representations for linguistic acceptability classification. We examine a zero-dimensional persistent-homology ($H_0$) representation that summarizes minimum-spanning-tree merge values from 144 symmetrized BERT attention graphs into 1,440 features. The retrospective evaluation uses 8,517 downstream training sentences and 521 in-domain and 497 out-of-domain CoLA validation sentences after data-integrity filtering. The recorded out-of-domain Matthews correlation coefficients are 0.438 for the topological representation and 0.508 for attention statistics, compared with 0.545 for BERT. Attention-feature results are retained aggregate measurements; BERT predictions are rescored and length controls are refitted from the accompanying records. Analytical results show that the representation retains token-count information and discards a class of directional attention differences. Source resampling of BERT predictions also demonstrates the sensitivity of uncertainty estimates to the sampling unit. The reported results do not demonstrate a predictive advantage for this fixed topological representation over simpler attention summaries. They identify representation properties and evaluation controls relevant to subsequent studies of attention topology.

0.27

Centroid Spectral Coupling: Implicit Alignment and Representation Diversity in Self-Supervised Learning

cs.CV cs.LGstat.ML ·2026-11-17 ·M. Hadi Sepanj, Benyamin Ghojogh, Saed Moradi, Paul Fieguth
Abstract

Self-Supervised Learning (SSL) aims to align augmented views while preserving representation diversity and preventing collapse. Existing non-contrastive methods often address these goals using multiple weighted loss terms. We introduce Centroid Spectral Coupling (CSC), a single-term SSL objective that implicitly couples view alignment, representation diversity, and collapse resistance. CSC constructs centroids from independently normalized augmented-view pairs and applies a ridge-based spectral objective to their uncentered second-moment matrix. Preserving centroid magnitudes encodes view agreement while encouraging balanced representation energy across spectral directions. We establish an exact decomposition into nonnegative alignment and spectral-imbalance components, characterize the geometry of global optima, and derive finite-loss guarantees for alignment and diversity. We also show that complete constant-output collapse is a strict saddle in normalized representation space. Experiments on CIFAR-10, CIFAR-100, and Imagenette show that CSC achieves the highest linear Top-1 accuracy among the evaluated non-contrastive methods on all three benchmarks, and the highest kNN Top-1 accuracy on both CIFAR datasets. A ridge-parameter sensitivity study reveals an empirical alignment--diversity trade-off: increasing $κ$ strengthens view alignment but reduces representation effective rank. These results show that a unified spectral objective yields competitive visual representations without explicit contrastive negatives or separately weighted alignment and diversity penalties.

0.26

What's in a Tree ? Hierarchical $n$-gram models as controlled approximators of language

cs.CL cond-mat.dis-nncs.AIcs.LG ·2026-11-17 ·Aritra Das, Francesco Cagnetta, Darshil Doshi, Victor V. Albert, Maissam Barkeshli +1 more
Abstract · 7+13 figures, 10+30 pages

Hierarchical models have been proposed as synthetic settings for studying language. However, their relevance to realistic language corpora remains underexplored. In this work, we introduce tree-grams, autoregressive generative $n$-gram models that build the conditional probability tensor $p(x_n \mid x_{1:n-1})$ through a hierarchical composition of order-2 and order-3 tensors. $n$ controls the maximal context window, whereas a rank hyperparameter $d$ controls model capacity, yielding an interpretable and systematically scalable family of approximators of language. We show how tree-grams outperform existing $n$-gram baselines, while remaining competitive with deep transformer-based language models. We demonstrate the utility of such controlled approximations by using tree-grams as generative models. In particular, we show that hyperparameter transfer under $μ$P can be made more robust when narrower models are tuned on simpler data, generated by tree-grams with moderate $n$ and $d$. In addition, we show that entropy and effective range of correlations control performance gap after fine-tuning. This is because tree-grams allow us to explicitly control horizon and capacity. When allowed to route across candidate topologies, we observe emergent sparsity in tree-grams. Such concentrated routing permits pruning upto 95\% of topologies without significant change in performance.

0.21

Event-Based Dynamic Lip Region Estimation During Speech

cs.CV cs.LG ·2026-11-17 ·Hamze Hammami, Omar Eldardeer, Francesco Rea
Abstract

Event cameras offer high temporal resolution and low power consumption, making them promising for speech-related applications. These include automatic lip reading, audiovisual speech analysis, and human-robot interaction, particularly in challenging environments characterized by low illumination or high dynamic range. Event-based lip reading, speaker recognition, and voice activity detection, however, mostly operate on a cropped region of interest (ROI), by hand or by the dataset, and no event-based method supplies that ROI as a spatial region; existing event-based face analysis gives the mouth as landmarks, which presents a mismatch between application and localization. The present work predicts a dense heatmap over the lip region from the full event "frame". Events are encoded as a polarity-separated voxel grid and processed at every location by a two-pathway continuous-time recurrent network (CT-RNN) with learned, log-space bounded time constants acting on the real elapsed time of each bin: a stateless fast pathway and a stateful slow pathway, utilised for refining and anchoring the prediction. The model reaches 69.5% on lips on four unseen speakers and switches target only 2.1 times per second, against 8.9 and 11.0 for convolutional baselines with more parameters. Adding a spatial encoder improves precision but fails on a new face scale. On a real event sensor, trained only on simulated events, the proposed model reaches 58.0% on lips against 43.2% for the CNN, with fewer than half its target switches.

0.19

What MoE Routing Controls: Causal Interventions on Expert Trajectories

cs.CL cs.LG ·2026-11-17 ·Mukund Athmaram, Manoj Rao, Liang Ma
Abstract

In mixture-of-experts (MoE) language models, a token is routed to a sequence of experts across layers. We investigate how forcing tokens onto chosen expert paths influences model behavior. We build a target path from the experts that a concept's or role's tokens most often use, and we add a fixed bias to each target expert's router logit during generation to steer the token toward this expert path. We study two kinds of targets. Format targets concern the form of the output, and concept targets concern its topic. In Qwen3-30B-A3B, we show that steering a format path changes the text the model writes. We also show that steering a concept path shifts the model's next-token probabilities toward the concept, even when the path is built in another language, yet this shift has little effect on the generated text. We investigate why formats reach the text but concepts do not by steering only part of the generation, and find evidence that a format, once started, reinforces itself. We show that the concepts religion and money steer most strongly, and that the effect is concentrated in the middle layers. We attempt to explain why the strength of the effect varies by concept through an analysis of embedding distinctiveness and word frequency.

0.18

Hierarchical Diffusion Transformer for End-to-End Image Generation

cs.CV cs.AIcs.LG ·2026-11-17 ·Xiaoyu Wu, Yifei Wang, Weihang Guo, Chen Wei, Lydia E Kavraki +2 more
Abstract · Under Review

Pixel-space diffusion learns visual representations and image generation jointly in a single end-to-end training stage, without a separately pretrained visual encoder. However, plain pixel-space Transformers still trail leading two-stage latent diffusion models in generation quality and training efficiency. We hypothesize that part of this gap arises from retaining one token per patch throughout the backbone, which does not explicitly encourage a visual hierarchy that summarizes related content at progressively coarser levels. To address this, we introduce Hierarchical Diffusion Transformers (HiT), which progressively group related tokens and represent each group with a single token. The main Transformer bottleneck processes these compact group tokens, and ungrouping restores features to their original image positions. A compression budget regulates the average compression ratio across the batch, allowing group memberships and counts to vary across images and diffusion timesteps during both training and generation. The model exhibits emergent semantic structure, learning to group coherent objects and object parts without segmentation labels. In text-to-image generation, HiT further learns to group caption words with image patches depicting the same concepts, without region annotations. HiT achieves 1.48 FID on ImageNet $256\times256$ and reaches comparable generation quality to JiT using approximately 40\% of its cumulative training FLOPs. For text-to-image generation, HiT exceeds UniLiP's average score across five benchmarks with roughly $6\times$ fewer denoiser parameters and $3\times$ fewer forward FLOPs. These results suggest that learning a visual hierarchy within the generator breaks down the barrier to stronger image generation without external visual pretraining.

0.18

RISA: Recoverable Sparse Attention for Block-Diffusion Language Models

cs.CL cs.LG ·2026-11-17 ·Xinrui Zhong, Fanjiang Ye, Yuke Wang
Abstract

Block-wise diffusion language models (DLMs) refine multiple tokens in parallel, offering an alternative to token-by-token autoregressive decoding. Their long-context inference, however, remains bottlenecked by memory-bound attention: every denoising step repeatedly reads a growing historical key--value (KV) cache. Directly applying sparse attention does not fully solve this problem, because different queries can select different pages, inflating the block's collective read set, while repeated support construction adds overhead and permanent eviction can discard information needed later. We present RISA, a recoverable sparse-attention method with shared read plans for block-wise DLMs. RISA exploits the fact that historical KV remains fixed within a block while queries evolve across denoising steps. Newly generated history is organized into exact priority and recovery tiers; seed queries then build a protected, bounded read plan whose indices are reused across the block while attention is recomputed with fresh queries and current-block KV. This separates history preservation from repeated reading and amortizes sparse-support construction without changing the backbone or decoder. On DreamReasoner-8B, RISA retains comparable task performance. At 32K, the default configuration speeds up attention by $2.67\times$ and fixed-trajectory requests by $1.36\times$ over dense, and attention speedup reaches $5.53\times$ at 128K. Across eight settings on three models, RISA decodes faster than the tested MAGE, QUEST, and LoSA budgets at matched KL.

0.18

RLDE: A Reinforcement Learning Framework for Interpretable Iterative Document Editing

cs.CL cs.AIcs.LG ·2026-11-17 ·Nikhil Parkar, M. Marimuthu, Riya Joshi, Karthikeyan N
Abstract

Improving the quality, clarity, and flow of long documents remains a challenging problem in NLP, as human editing typically occurs through iterative refinement rather than a single-pass rewrite. Most existing automated systems act as black-box generators or rely on supervised learning, limiting interpretability and failing to capture the sequential nature of revisions. In this paper, we propose RLDE, an automated end-to-end metric-driven RL framework that formulates document revision as a reference-free MDP problem (unsupervised with respect to paired document-revision targets). Working with a limited action space (NOOP, INSERT, DELETE, REPLACE), RLDE leverages Pointer Networks for dynamic span selection at the token level and sends generative actions to the frozen BART-base model to obtain context-aware suggestions. PPO agents learn to optimize a multi-objective reward function based on readability, perplexity, and grammar metrics with the help of a novel SBERT-based Semantic Preservation Penalty constraint to avoid deletion degeneracy. Evaluations on the BEA-2019 dataset showed statistically significant results (p < 0.01) across all metrics, where Flesch Reading Ease increased from 49.8 to 59.4, perplexity decreased from 63.7 to 39.8, while preserving 93% of semantic information. When compared to Chain-of-Thought (CoT) prompt-engineered models like GPT-4 and LLaMA-2, RLDE shows high performance at a very low computational inference cost. Moreover, the quality of edits generated by RLDE has been verified via human evaluations, scoring 4.9/5 for meaning preservation and 4.4/5 for edit quality.

0.17

When AUC Meets Generative Reconstruction: A Unified Principle for Balanced Visual Representation Learning

cs.CV cs.AIcs.LG ·2026-11-17 ·Boyu Han, Qianqian Xu, Shilong Bao, Zhiyong Yang, Qingming Huang
Abstract

Multimodal large language models (MLLMs) require visual encoders that extract fine-grained visual features while constructing semantically disentangled and discriminative representations. Conventional methodologies encounter a fundamental trade-off: contrastive learning enhances inter-class discrimination by suppressing intra-class variability, whereas generative reconstruction emphasizes perceptual fidelity but lacks explicit comparative supervision. The naive aggregation of these objectives via composite loss functions often induces optimization conflicts and instability due to gradient interference within the shared latent space. To resolve this, we introduce a unified representation-learning paradigm that reformulates generative reconstruction as a discriminative ranking problem within the generative output manifold. Our approach maintains perceptual reconstruction as a supervisory signal and integrates a ranking-based discriminative constraint by systematically ordering matched and mismatched reconstruction outputs. This formulation embeds discriminative information seamlessly into the generative process, alleviating the risk of loss competition. Specifically, we propose AUC Generative Reconstruction (AUCGR), which optimizes the area under the ROC curve (AUC) by ensuring higher scores for matched pairs than for mismatched pairs, and Partial AUC Generative Reconstruction (pAUCGR), which addresses supervision bias by focusing on mismatched comparisons within reliable feature subspaces. The proposed objectives are designed to be plug-and-play, requiring no architectural modifications to existing generative reconstruction frameworks. Theoretical analysis establishes rigorous connections to AUC/pAUC optimization and provides generalization guarantees. Extensive experiments validate that our method advances both discriminative accuracy and fine-grained perceptual representation.

0.16

Uncovering the Invisible in Routine Radiology: Cross-Modal Detection of MRI-Verified Soft-Tissue Pathologies from CBCT via Physics-Guided 3D Self-Supervised Learning

eess.IV cs.CVcs.LG ·2026-11-17 ·Evgeny Alves Limarenko
Abstract

Routine X-ray modalities like cone-beam computed tomography (CBCT) are optimized for high-contrast mineralized tissues, leaving soft-tissue abnormalities buried beneath photon noise and below human visual perception. Thus, detecting non-calcified pathologies requires secondary magnetic resonance imaging (MRI). We demonstrate that sub-visual radiographic signatures of soft-tissue pathology are physically preserved within routine CBCT scans and can be extracted for opportunistic screening using a targeted, physics-guided deep learning pipeline, even under extreme paired-data scarcity. Using the temporomandibular joint (TMJ) as a model system (N = 212 paired CBCT-MRI joints), we address the severe overfitting of high-capacity 3D networks through a two-stage framework. First, we employ a frozen full-grid 3D masked autoencoder (BertsWin) pre-trained on 6,550 unlabeled CBCT volumes via the GradientConductor (GCond) optimizer and an anatomy-aware objective (PhysLoss) redirecting gradient flow into the [-300, 300] HU soft-tissue window. Second, automated localization of the condylar zenith drives a deterministic 3D cropping that eliminates 81% of background osseous variance. Evaluated via a patient-isolated 3x10-fold stratified cross-validation protocol, a statistically robust sub-visual signal for MRI-verified synovial effusion (ROC AUC = 0.793) emerges exclusively from representations pre-trained with the PhysLoss and GCond regime. Crucially, only this setup captures true 3D anatomical chirality (p < 0.001) and high intra-patient structural consistency (0.791) over inter-patient noise; conversely, identical architectures pre-trained with standard L2 loss or unmasked variance decomposition (MVC) collapse to random guessing. These findings establish a generalizable paradigm for repurposing routine, low-contrast radiographic scans into opportunistic screening tools for occult soft-tissue disease.

0.14

Systems & Networking 3

PVTSizing 2.0: A Fast and Reliable Analog Circuit Sizing Operator for Agentic EDA Workflows

cs.AR cs.LG ·2026-11-17 ·Zichen Kong, Minghuan Song, Xiyuan Tang, Ziming Wei, Yi Zhong +1 more
Abstract · 10 pages, 4 figures, 5 tables

Analog circuit design automation requires both coordinated tool execution and efficient numerical search. Large language models (LLMs) offer new ways to organize tool execution, but finding one sizing that satisfies specifications across process, voltage, and temperature (PVT) corners remains costly. PVTSizing 2.0 addresses these needs through a human-agent workflow built around an independently usable numerical optimizer. An agent carries the designer's approved task through execution and result review. The optimizer reduces unnecessary training through loss-guided updates and uses critic-guided local refinement to explore additional candidates while accounting for cross-corner conflicts. An optional LLM adviser supplies parameter relationships to focus numerical search on circuit-informed directions, while the optimizer selects adjustment amplitudes and verifies their effects through simulation. Across four circuit benchmarks, the standalone optimizer completes all 40 runs and achieves a 7.3x speedup in overall mean runtime over the original PVTSizing, with a comparable mean sample count. In a separate study, the best tested knowledge-injection configuration further reduces mean runtime by 39.9% and simulation samples by 35.4% relative to the standalone optimizer.

0.24

XM-Sparse: High-Throughput Sparse-Attention LLM Serving over Scale-up Memory Pools

cs.DC cs.LG ·2026-11-17 ·Yijun Sun, Xudong Liao, Zhuobin Huang, Kai Chen, Nikita Lazarev +1 more
Abstract · 13 pages, 11 figures

Sparse attention has been increasingly adopted by production LLMs for long-context inference, yet serving systems must still retain the full KV history, making GPU HBM capacity a key limit on serving concurrency. Existing systems offload the KV cache to host memory, but suffer from the limited bandwidth of PCIe and the limited capacity of host memory. Emerging scale-up memory pools (e.g., X-Mem) relieve both limits and open new opportunities for sparse-attention offloading. In this paper, we first conduct an in-depth analysis and reveal that, even with scale-up memory pools, decode throughput remains bounded by HBM capacity, as hot KV buffers and the fully resident indexer key (K-index) cache occupy substantial HBM. Moreover, the two caches exhibit distinct access patterns. Based on these insights, we present XM-Sparse, the first sparse-attention offloading system over scale-up memory expansion. The core idea of XM-Sparse is to offload both caches aggressively while matching each with an I/O mechanism tailored to its access pattern. Specifically, for dynamically selected KV entries, XM-Sparse fuses TMA-based retrieval with attention in a tile-level pipeline to hide on-demand retrieval latency. For predictable K-index accesses, it prefetches pages via SM-efficient GPU-initiated DMA and coordinates prefetching with latency-critical operations to avoid interfering with decoding. We have implemented XM-Sparse atop SGLang, and our experiments on DeepSeek-V3.2 and GLM-5.3 with an emulated scale-up memory pool show that XM-Sparse delivers up to 4.98$\times$ and 4.52$\times$ higher decode throughput than PCIe-based ECHO and HiSparse, respectively, and up to 1.34$\times$ over a naive integration on the same platform, across 16K-1M contexts.

0.21

Manufacturing Differential Sparsity: Planner-Hardware Co-Design for Efficient World Model Planning

cs.AR cs.AIcs.LGcs.RO+1 ·2026-11-17 ·Wenbo Zhu, Paul Yi-Chia Chen, Yuanhan Li, Zishen Wan, Christopher Torng
Abstract · 14 pages, 11 figures

JEPA-style latent world models let embodied agents plan toward new goals with model predictive control, but they are still too slow for real-life embodied control: sampling-based MPC repeats predictor inference tens of thousands of times per plan, and even on high-end GPUs one planning step can take tens of seconds. We observe substantial latent locality in this repeated-inference workload: the predictor maps close actions to close latents, and inside a quantized predictor this closeness takes a form that hardware can exploit, differential bit-level sparsity. Unlike an input stream, whose temporal similarity is given, a planner itself decides which candidates it evaluates together, so the planner can choose its own references and, with reference forest sampling, manufacture more sparsity: it couples the sampled action sequences of each group to a shared root, with competitive planning quality. We build the accelerator on the insights of the rich line of bit-level sparsity accelerators and design it and the planner tightly coupled as a single co-design: the planner increases sparsity and exposes structure for hardware scheduling, and the bit-exact hardware effectively uses the sparsity that the planner creates. On three JEPA world models, our co-design is 1.28-1.80x faster than a dense accelerator and uses 38-52% less energy, and it is 1.22-1.58x faster than prior bit-sparse design and uses 16-36% less energy, with the same area. To our knowledge, our planner is the first planner that manufactures differential bit-level sparsity across its candidates and exposes its structure for hardware scheduling, and our co-design is the first algorithm-hardware co-design for JEPA-style world-model planning.

0.17

Software & Programming 2

From Constraints to Causality: Petri net Encodings for Explaining Binary Neural Network Inference

cs.LO cs.LG ·2026-11-17 ·Alex Chan, Mohamed Tarraf, Alex Yakovlev, Rishad Shafik
publishedtheory
Abstract · In Proceedings FMAS 2026, arXiv:2611.21375

Binary Neural Networks (BNNs) are attractive for resource-constrained hardware and on-device components in autonomous systems. However, their discrete, non-linear behaviour makes them difficult to analyse and trust in safety-relevant settings. Existing approaches based on SAT and Satisfiability Modulo Theories can verify a BNN's input-output properties, but treat it as a monolithic constraint system, which reveal only whether a property holds and not how the computation reaches its result. In this paper, we present an early work that provides a complementary, discrete-event view. We give a compositional translation from the inference semantics of a fixed, trained BNN to a 1-safe Petri net (PN), where each computation step is an explicit transition, so that a firing sequence is a causal trace of the BNN's decision. On a worked example, we establish faithfulness via exhaustive input-output equivalence against a reference implementation, and show how contrasting firing sequences localises where a single-feature change alters the prediction. Finally, we outline two directions this opens: a correctness-of-translation theorem, and a causal-trace analysis of where two trained BNNs diverge in their decisions, a structural question that constraint-based verifiers cannot express.

0.65

Higher-Scoring Coding Models Did Not Get More Work Done

cs.SE cs.AIcs.LG ·2026-11-17 ·Yining Hua
unusual
Abstract · 20 pages, 4 figures

New coding models advertise themselves on score improvements of coding benchmarks. Yet no study has tested if a model that scores higher on these benchmarks gets more real work done. On June 12, 2026, a U.S. export control directive banned access to Claude Fable 5 for 19 days, which forced users to switch to other models. During this time, most former Fable 5 users switched to Opus 4.8 (11 points lower on SWE-Bench Pro compared to Fable 5). We treat this suspension as a natural experiment, where the people and projects stayed the same and only the model changed, so we could isolate the effect of Fable 5 on user productivity. We tracked GitHub pull requests of 297 Fable 5 users before, during, and after the suspension. We used 8,908 developers who stayed on Opus 4.8 all the time as the control group. We also repeated the comparison across five model upgrades before and after the suspension in 2026 (1,208 developers, 15,442 pull requests). Fitting a difference-in-differences model with developer and date fixed effects, we found that having access to Fable 5 increased the 7-day merge rate of pull requests by only 1.4 percentage points [-1.4; +4.3], which was not statistically significant. Similarly, no significant increases of merge rates were found with the upgrades (weighted mean: -0.9 percentage points [-1.9; +0.2]); all five estimates were negative (95% intervals included 0, did not differ from each other). We also did not observe a significant increase in the number of pull requests merged per week (+2.4% [-4.9%; +10.2%]).

0.44

Theory & Algorithms 1

Computed Permutation Trellis Coding: Near-Optimal Quantization from Two Multiplications and a 64-Byte Table

cs.IT cs.AIcs.ARcs.LG ·2026-11-17 ·Muhammed Fatih Balın
Abstract · 38 pages, 6 figures, 27 tables

We propose Computed Permutation Trellis Coding (CPTC), a low-bit quantization format for tensors written once and read many times. Formats that round each value on its own, such as INT4 and FP4, waste a large part of their bits against the information-theoretic ideal. Trellis codes recover this loss by encoding values jointly, but prior decoders either need a large lookup table for their best quality or compute values with a hash that restricts what they can decode. CPTC's decoder multiplies its state by two coefficients and adds the top bits of the products to index a 64-byte table. Because that sum is not uniformly distributed, the rarely used ends of a small table can hold values far out in the tails. One decoder serves every bit width from 1 to 6, reproduces the standard integer and floating-point formats exactly, and comes within 0.07 bits of the ideal at 4 bits. It matches a random-permutation trellis code whose table is thousands of times larger, and exceeds every QTIP codebook. In LLM weight quantization without fine-tuning, CPTC has the lowest KL divergence to the unquantized model at every tested bit width on Llama 3.1 8B and 70B Instruct against QTIP's codebooks (by up to 27%). The decoder has been built into the AWS Trainium 4 AI accelerator, where it runs at the throughput of the native FP4 and FP8 formats for a negligible fraction of the systolic array's area. Fused encoding kernels quantize an 8B model's linear layers in under two minutes on one GPU and, in cycle-level simulation, on one Trainium 4 chip.

0.40

Robotics, Control & Signals 8

Magflow: Differentiable Simulation of Industrial-Grade 6D Magnetic Levitation

cs.RO cs.LGeess.SY ·2026-11-17 ·Philipp Hartmann, Jannick Stranghöner, Klaus Neumann
top venuecode/data
Abstract · 8 pages, 5 figures, 1 table. Accepted for publication in IEEE Robotics and Automation Letters (RA-L). Project page: https://sites.google.com…

Magnetic levitation with six-degrees-of-freedom enables frictionless transport and micrometer-scale positioning in principle, but low-level control is inherently challenging due to its complex and unstable dynamics. Industrial systems typically rely on deep hardware-software co-design, where mover and stator geometries are jointly designed to enable a compact, hand-crafted approximation of the electromagnetic force computation, thereby keeping control computationally tractable at kilohertz rates. However, this approach imposes overly conservative constraints on the operating workspace and the range of supported mover topologies. Instead of reducing the underlying physics to a hand-crafted approximation, we compute the physics directly. We introduce Magflow, a fully differentiable, GPU-parallel simulator for 6D magnetic levitation. It is used to train the first neural controller for an industrial-grade 6D magnetic levitation system that is learned purely from simulated physics. We show that the learned neural controller transfers zero-shot from simulation to real hardware while satisfying hard real-time constraints at 4 kHz. The learned controller achieves qualitatively new behaviors and even levitates a mover topology for which no hand-crafted solution exists. These results demonstrate that co-design is not a physical necessity for 6D magnetic levitation control. More broadly, control learning from simulated physics can find solutions beyond the reach of control engineering in demanding industrial domains. Source code and demonstration videos are publicly available at https://sites.google.com/view/magflow.

0.59

Auditing a Learned Simulator for Legged Robots: The Robotic World Model on ANYmal D

cs.RO cs.LG ·2026-11-17 ·Joyjeet Singh
code/data
Abstract · 39 pages, 13 figures, 20 tables. Code and data: https://github.com/joyjeet-singh/rwm and https://github.com/joyjeet-singh/rwm-robot-level. E…

A legged robot can learn to walk inside a learned simulator: a neural network that predicts the robot's next state and is run forward on its own predictions. They are usually trained and checked one step ahead, although a policy uses them for a hundred steps or more. We audit one published system end to end, without a GPU: the dynamics model of the Robotic World Model (RWM) for the ANYmal D quadruped, and the uncertainty estimate of its follow-up, RWM-U. We rebuilt the model to match the released code's outputs, losses and gradients exactly, measured 101 trained simulators, 200 compressed copies and 38 simulators of three other architectures, and trained policies inside simulators that we tested in a hand-built MuJoCo model of the robot. Training on the model's own rollouts beat one-step training at 368 steps on all ten held-out episodes (registered), although one-step training was better one step ahead. One-step checks therefore misjudge design choices: pruning 30% of the weights kept 90% of the model's skill one step ahead and 39% at 2 s. Ensemble disagreement ranked rollout error well (Spearman +0.90 at 100 steps), yet covered 4.6% of outcomes at one sigma, against 68% for a calibrated estimate, as each network's own variance collapses under the implemented loss. Two checks that need no recorded future ranked some pools of simulators well, but neither was shown better than 8-step error in a primary test. At the robot level, accuracy has not predicted walking: all 18 fine-tuned copies we tested predicted better than the released model one and eight steps ahead, yet 13 gave failing policies, and registered tests of whether any check ranks simulators by their policies were not settled. In a secondary test, running the authors' policy inside each simulator ranked them well. We also report nine of our own claims that we withdrew. Scope: one robot, one gait, simulated data, no hardware.

0.47

From Temperature-Limited to Battery-Limited Legged Locomotion

cs.RO cs.LGeess.SY ·2026-11-17 ·Raphaël Ségas, Balázs Opra, Léo Ghafari, Thomas Stewart
hot direction
Abstract · 9 pages, 7 figures, 2 tables. Supplementary video in the ancillary files. Dataset: https://doi.org/10.5281/zenodo.24274994

On legged robots, sustained locomotion often ends in thermal shutdown rather than battery depletion: the heat of walking concentrates in a few joint motors, which overheat long before the battery is empty. We train thermally aware locomotion policies by reinforcement learning without ever observing temperature, and demonstrate this on a Unitree Go2 quadruped. We first identify a coupled 13-node lumped RC thermal network, fitted on two runs of a single robot and calibrated against the measured electrical power, that predicts all twelve motor temperatures with a mean RMSE below 1 $^\circ$C across different robot instances, gaits, payloads, and speeds. Under a sustained gait, the temperatures the motors will settle at follow directly from the gait-averaged squared torques and joint speeds. Our core contribution is a reward that drives down the hottest predicted asymptotes: thermal time constants exceed the credit-assignment horizon of reinforcement learning by orders of magnitude, but predicted asymptotes respond to torque and joint speed within a single gait cycle, so the policy learns to lower and spread heat with no temperature input. On the robot, the policy turns sustained locomotion from temperature-limited to battery-limited: with a 5 kg payload at 1 m/s, a learned baseline reaches thermal shutdown after 19 minutes (1.1 km), whereas our policy runs the battery down after 174 minutes (10.0 km) with its hottest motor 14 $^\circ$C below the shutdown threshold in a 26 $^\circ$C room; on an unloaded factory circuit with stairs and ramps, the baseline overheats after 25 laps (6.2 km), while our policy completes 47 (11.2 km) on one battery with 22 $^\circ$C of margin.

0.45

Heave-Ho: Decentralized Long-Horizon Multi-Humanoid Collaboration

cs.RO cs.CVcs.LG ·2026-11-17 ·Ting-Hsuan Chen, Siheng Zhao, Han Jiang, Qi Pan, Louison Lu +5 more
benchmarkcode/data
Abstract · 18 pages, 11 figures, 7 tables. Project page: https://heaveho-humanoid.github.io/

Collaboration promises to extend humanoid capabilities beyond what individual robots can achieve. However, realizing this potential requires coordinated whole-body interactions, coherent long-horizon planning, and precise synchronization across robots. We introduce Heave-Ho, a decentralized framework for long-horizon collaborative multi-humanoid loco-manipulation. At the low level, we develop a multi-agent post-training framework that adapts a humanoid foundation tracking model to physically coupled interactions between humanoids, enabling coordinated decentralized control from local observations alone. At the high level, we train a vision-language planner and design a speech-based leader-follower coordination mechanism to synchronize decentralized execution without network communication. Trained with two identical humanoids, our system scales to teams of up to ten robots and generalizes to heterogeneous teams without retraining. To benchmark these capabilities, we introduce HH-Bench, comprising 12 tasks involving teams of 2 to 10 humanoids, including heterogeneous teams. Systematic evaluations show that Heave-Ho consistently outperforms baseline methods, while real-world experiments demonstrate robust decentralized collaboration across challenging tasks.

0.38

Capability Composition in Heterogeneous Teams via Barrier-Guided Diffusion Sampling

cs.RO cs.LGcs.MAeess.SY ·2026-11-17 ·Diana Laura Morales, Ruoyu Lin, Magnus Egerstedt, Fabio Pasqualetti
theory
Abstract

The agents in a heterogeneous team can cooperate to accomplish tasks that exceed their individual capabilities. The team trajectories for a cooperative task can be generated by a generative model, provided that the model has been trained on demonstrations of the agents cooperating. Yet, demonstrations of cooperation are costly and may not exist for tasks that the team has never performed. The conditions under which the agents can combine their capabilities, however, are physical properties of the team and are known before any demonstration is collected. We encode the conditions as constraints on the team trajectories, and we enforce the constraints while sampling from generative models trained on demonstrations of each agent alone. We prove that the proposed method can generate every team trajectory that satisfies the constraints, although the demonstrations contain no cooperation. Hence the constraints determine which team trajectories can be generated, and the demonstrations determine how likely each team trajectory is. Numerical results on two cooperative tasks show the effectiveness of the method, which generates cooperative team trajectories that satisfy the constraints and requires no retraining when agents join or leave the team. Finally, we implement the method on a multi-robot testbed.

0.33

Temporal Diffusion Models for Joint Signal-Channel Estimation under Gaussian and $α$-Stable MIMO Channels

eess.SP cs.LG ·2026-11-17 ·Rawad Kansoh, Bahaa Ammoury, Charbel Hannoun, Jihad Fahs, Ibrahim Abou-Faycal
Abstract · 14 pages, 10 figures

Diffusion models have emerged as powerful tools for learning complex data distributions and solving inverse problems. However, most diffusion models rely on the Gaussian distribution, which does not accurately represent heavy-tailed data. In this work, we study diffusion-based reconstruction under Gaussian and symmetric $α$-stable fading channels, while also exploring the benefits of using possible temporal dependence in the channel. We consider two variants: pilot-based channel-only estimation and joint image-channel recovery for MIMO systems. Our approach is based on several standard and score-based diffusion models for both Gaussian and $α$-stable distributions. We refine these models by guiding the reverse diffusion process using the observation likelihood and show how a single trained diffusion model can be reused for the location-scale family of distributions without retraining. In blind MIMO recovery, separate diffusion models are used for the image and channel, and their estimates are combined through likelihood guidance. For each setting, we compare independent and temporal reconstruction, where information from previous observations is used to improve the estimation. We show that, under a time-dependent channel model, such a temporal reconstruction improves channel estimation and the quality of the reconstructed image. Finally, we show that, for the same heavy-tailed channel data, using an $α$-stable diffusion prior provides better channel estimation and image reconstruction than using a Gaussian diffusion prior.

0.33

Stochastic Diversity Windows: Diagnosing Noise Influence in Robotic Diffusion Sampling

cs.RO cs.LG ·2026-11-17 ·Tina Lakhani, Jing Wu, Yu-Kun Lai, Ze Ji
Abstract · 26 pages, 3 figures, 6 tables

Diffusion models are increasingly adopted in robotics and control due to their capability to express multiple feasible solutions to a given task. However, the role of stochasticity in the diffusion model-sampler across denoising remains poorly understood. We introduce a per-timestep counterfactual diagnostic that tests how changing the sampled noise affects the final output. Across five model-sampler configurations spanning three robotics benchmarks, diverse architectures, domains, and conditioning mechanisms, we reveal highly varied temporal sensitivity profiles, where noise-induced deviations can amplify, partially contract, or undergo near-total attenuation downstream. From these profiles, we define stochastic diversity windows (SDWs): compact intervals that identify where sampler stochasticity has the strongest measured influence on the final output. Guided by these windows, we apply two heuristic interventions: targeted stochastic amplification, which increases sample diversity exploration of the solution space, and an SDW-hybrid sampler, which replaces DDPM steps with DDIM skips in low-impact regions, resulting in network evaluation reductions of up to 40-60% and wall-clock time by up to 85%, with no consistent changes detected in feasibility, cost, or diversity across held-out tasks.

0.28

Task Convergent Flow Matching for Generative Robotics

cs.RO cs.LG ·2026-11-17 ·Lorenzo Paiola, Maximilian Stölzle, Giorgio Grioli, Daniela Rus, Cosimo Della Santina
Abstract · Submitted to L4DC 2027

Successful manipulation requires the manipulated objects to progress toward a desired task outcome. Generative policies can produce plausible robot motions without ensuring this progress, allowing successive action chunks to stall or undo earlier changes to the scene. We introduce a task-centric inference-time refinement framework for flow matching policies that evaluates candidate action chunks through parallel physics-based rollouts and searches for motions satisfying sufficient conditions for task convergence. Task-level performance funnels bound temporary regressions while requiring net progress over each chunk. We establish sufficient conditions under which repeated funnel satisfaction yields convergence throughout execution, and use interval-wise Lipschitz bounds to verify continuous-time funnel satisfaction from sampled rollouts. To accommodate contact-rich hybrid dynamics, we refine the flow's source seeds using derivative-free Covariance Matrix Adaptation Evolution Strategy (CMA-ES), constrained to radial quantile bands of the source distribution to help retain the learned motion prior. This formulation supports both Euclidean and pose-valued Riemannian Flow Matching (RMFM) policies. On PushT and three-cube stacking, refinement raises success from 31.3% to 88.5% and from 77.0% to 86.0%. On PushT, whose coverage score responds continuously to the motion of the block, a learned funnel lowers the average maximum loss of achieved progress by 19% relative to constraining the endpoint of each chunk alone, at a similar success rate. With the radial band, 11.3% of the executed motion windows are out of distribution, a rate within the range of held-out demonstrations, compared with 34.1% for an unconstrained search.

0.24