In briefParsing logs drops key details: a benchmark for detecting failures in distributed GPU training
SummaryThe paper introduces GF-Bench, a PyTorch-based framework and dataset for generating failures in distributed GPU workloads, with 220 labeled two-GPU runs covering normal execution, CUDA out-of-memory failures, and persistent straggler failures. The authors evaluate three sequence-based methods and the LLM-based LogPrompt on parsed logs, finding that none reliably detects both failure classes, with a best overall F1 of 0.667. Using raw logs raises LogPrompt's overall F1 to 0.983, but at about 9.0 times the input tokens and 8.7 times the per-run cost.
In briefInvariant Atoms: Sparse Directions for Local Meaning in Language Models
SummaryThis paper proposes the Invariant Atom Hypothesis, which holds that local semantic changes in language model representations follow sparse directions that stay stable under meaning-preserving rewording. The authors learn a shared semantic frame and sparse coordinates that reconstruct semantic shifts while suppressing nuisance variation. The reported experiments show good separation between semantic and nuisance factors, causal effects on model predictions, and generalization to unseen semantic neighborhoods and nuisance families.
In briefTheory on how attention behaves when transformers face out-of-distribution in-context learning inputs
SummaryThe paper theoretically analyzes why transformers lose accuracy on out-of-distribution inputs during in-context learning, using the dynamics of attention weights that track confidence in correct and incorrect features. It shows the error depends on interactions between training and OOD features, and can reduce to random guessing in some cases. It also studies finetuning on OOD data, finding source-domain performance may not always degrade and depends on the nature of the feature shift, with experiments on synthetic and real data.
In briefOnline Versatile Incremental Learning: Adapting to Class and Domain Shifts at Any Time
SummaryThis paper introduces Online VIL, a scenario where class concepts and visual domains change simultaneously and online without explicit boundaries. To handle such dynamic conditions, the authors propose TopFlow, which combines Domain-agnostic Flow Matching (DFM) and Global Topology Preservation (GTP). Experiments reported in the paper show that TopFlow achieves state-of-the-art performance in the Online VIL setting.
In briefDARE: A Dual-Path Autoregressive-Aware Method for Reducing Hallucination in Vision-Language Models
SummaryLarge vision-language models often describe objects that do not match the image, a problem called object hallucination. This paper proposes DARE, a training-free representation editing framework that reduces such hallucinations. It builds editing directions from textual contrasts, representation changes during autoregressive decoding, and controlled visual differences between paired images. Experiments on multiple benchmarks show that DARE consistently reduces object hallucinations while preserving multimodal perception and inference efficiency.
WhyAdds autoregressive-aware signals and image-contrast paths to training-free representation editing; consistent gains across multiple LVLM hallucination benchmarks.
In briefBit-width choices for zero-knowledge LLM inference: which quantization settings suit verifiable private proofs
SummaryThis paper is the first systematic study of quantization methods that suit zero-knowledge proofs for large language models, and it formalizes what such quantization requires. The authors test nine language models across different weight, activation, and lookup-table precisions. They find that activation precision matters much more than weight precision, and that nonlinear lookup approximations often cause most of the accuracy loss. The paper also finds that reducing bit-width does not always cut proving cost proportionally.(ZK-friendly quantization)
In briefMerlin Plus: A Large-Scale Multi-Cancer CT Image-Mask-Report Dataset
SummaryThe paper introduces Merlin Plus, the first large-scale CT dataset with radiologist-created tumor masks across 9 organs. It extends the Merlin dataset with 1,153 per-voxel tumor masks and longitudinal metadata. The masks were produced with a report-based active-learning framework in which a segmentation model generates initial masks that radiologists then review and correct.
Pedro R. A. S. Bassi, Wenxuan Li, Szymon Plotka et al.
In briefMemFold: Compact learned memory that helps AI assistants personalize over long user histories
SummaryMemFold compresses a user's long-term history into a fixed set of continuous vectors that serve as a memory interface for a language model. The model is trained on its own generated outputs using task-outcome rewards plus a teacher-based distillation signal. The authors report the highest accuracy they measured on PersonaMem-32K and PersonaMem-128K, and transfer to PrefEval and LongMemEval without training on those domains. Ablations attribute most of the gain to the reward term, with a smaller additional gain from the teacher signal.
In briefSafety Operator: Tuning How Strongly Safety Instructions Shape Model Outputs via Spectral Optimization
SummaryThis paper studies how context tokens in transformer language models can be absorbed into model weights as a multiplicative operator, focusing on safety instructions. The authors propose a Contrastive Safety Loss with a suppression weight that emphasizes safety instructions on harmful queries and suppresses them on harmless ones. Varying this weight reveals a relationship between attack success rate and over-refusal rate, suggesting the operator's dominant eigenvalue works as a continuous control for the instruction's influence.
Benoit Dherin, Michael Munn, Xavier Gonzalvo et al.
In briefRA-CFGCache: Risk-aligned caching for faster diffusion sampling under classifier-free guidance
SummaryThis paper addresses the high computational cost of diffusion models under classifier-free guidance, which requires both conditional and unconditional evaluations. It proposes RA-CFGCache, a training-free caching framework that accounts for how cache errors combine across branches and how local errors affect the final output at different timesteps. An online threshold controller decides when to refresh or reuse both branches. Experiments on FLUX.1-dev, Wan2.1-T2V-1.3B, and CogVideoX-2B report improved efficiency-fidelity trade-offs over evaluated training-free caching baselines.
WhyProposes risk-aligned cache control under CFG, addressing error composition and timestep propagation; validated on three models against baselines.
In briefA paper on a temporal-aware fusion method for more robust outdoor LiDAR localization
SummaryThis paper proposes TempLoc, a framework that estimates a sensor's global 6-DoF pose from LiDAR scans by modeling consistency across consecutive frames. It predicts per-point global coordinates with uncertainties, estimates inter-frame point correspondences via attention, and fuses both through an uncertainty-guided module. The authors report that TempLoc outperforms state-of-the-art methods by a large margin on the NCLT and Oxford RobotCar benchmarks.
In briefPredicting single-cell gene expression from H&E histology images with context
SummaryThis paper introduces CELLO, an end-to-end framework that predicts gene expression for individual cells from H&E-stained histology images, rather than only at the aggregated spot level. It runs one pathology foundation model pass per image and uses grid sampling to extract features for all cells at once, with a distance-decay cross-attention module that adds spatially weighted local morphological context. Evaluated on 52 public Xenium-H&E pairs from HEST-1k covering 12 organs and about 10 million cells, the authors report higher average predictive accuracy than the baselines and a 14.0x average speed-up over DeepSpot2Cell in mean whole-slide inference time, excluding upstream cell segmentation.
In briefBeyond Pixel Reconstruction: Retrieval-Guided Glyph-Aware Restoration for Manchu Documents
SummaryHistorical Manchu documents are hard to digitize because of severe degradation and few paired training examples. The paper proposes a framework that retrieves relevant glyph exemplars to guide restoration with glyph-level structural knowledge. Experiments on Manchu historical documents show improved restoration quality and glyph-level fidelity over existing methods.
In briefPredictive Credit: How Much Scientific Explanations Improve Experimental Forecasts
SummaryThis paper proposes a protocol that measures how much explanations from research agents about planned experiments add to forecasts, using paired forecasts that differ only in the explanation provided. Tested on 336 states across controlled learning, Tox21, and OpenML tasks, the results were inconclusive. Predictive gains from natural explanations remained unconfirmed at the tested donor resolutions.
In briefRules to Tools: Executable Checks That Help AI Coding Agents in Scientific Computing
SummaryThis paper introduces R2T, which turns public scientific requirements such as equations, boundary conditions, and output rules into callable executable checks for LLM coding agents that repair programs. Compared with a group given the same checks only as written text, complete repair was 26/30 for text and 29/30 for tools across two task-ID cohorts. The authors report that outcomes depend on the task, with several tasks tying and some favoring text.
In briefLosing Uncommon Objects After Narrow Fine-Tuning, and a Training-Free Fix
SummaryThe paper studies how fine-tuning a broadly pretrained detector on a narrow domain reduces its ability to propose boxes for objects outside the deployment vocabulary, even as in-domain accuracy rises. It introduces a longitudinal protocol tracking top-K proposal coverage across four architectures and three domains. It also proposes a training-free repair that mixes a quarter of the pretrained state back in, raising coverage for at most 2.47 points of in-domain accuracy.
In briefAgent-based evolutionary model of how individual vehicles reach equilibrium on mixed-autonomy weaving ramps
SummaryThe paper proposes an agent-based model in which individual vehicles adjust lane choices with an evolutionary game-theoretic rule, linking to a macroscopic weaving-ramp theory. The authors prove that these decentralized dynamics converge to the unique equilibrium predicted by the macroscopic model. Simulations show close agreement with macroscopic predictions and examine how convergence rates, adaptation to traffic changes, heterogeneous altruism among CAVs, and imperfect information affect performance and how altruistic burden is distributed.
In briefSample Complexity of Equivariant RL: Using Group Symmetry to Cut Robot Learning Data Needs
SummaryThis paper studies how exploiting group symmetries can reduce the sample complexity of reinforcement learning, meaning the number of environment interactions needed to reach optimal return. For finite-horizon Markov decision processes, the authors show that homomorphisms induced by group symmetries significantly tighten theoretical upper and lower bounds, and they extend these results to continuous state and action spaces. Experiments on high-dimensional continuous robotic simulations indicate that symmetry-aware policy learning improves sample efficiency and performance.
In briefTarget leakage in support sets for relational in-context learning models
SummaryThis paper studies a failure mode in relational in-context learning where target-derived features appear in the support set but not for queries. The authors build 20 controlled leaker features and evaluate them on 13 RelBench tasks across five relational ICL configurations. They find that 0-hop and full leakage cause the largest deviations from clean evaluation, and that effects depend strongly on task and model, sometimes reversing conclusions between model variants.
Roshan Reddy Upendra, Alexandre Dorais, Joe Meyer et al.
In briefFineART: A densely annotated dual-arm robot dataset and vision-language-action model
SummaryThis paper presents FineART, a dual-arm manipulation dataset with 40,543 episodes (about 1,718 hours) and 533,913 subtasks across 151 tasks. It also introduces FineART-VLA, a vision-language-action policy that predicts its own next subtask to guide its actions. The authors report that mid-training on the subtask annotations raised success on spatial instructions from 32.0% to 100.0%, and that step-by-step subtask guidance raised success on unseen long-horizon tasks from 16.0% to 76.0%. The dataset, model weights, and training code are stated to be open-sourced.
Jade Choghari, Pepijn Kooijmans, Mansi Agarwal et al.
In briefEternal Sunshine of the Spotless Mind: Systematically Erasing LLM Memories (Whether persistent chatbots can truly forget what users asked them to delete)
SummaryThe paper studies persistent large language models that accumulate memories of user interactions over time, and asks whether they can forget information a user asks them to delete. The authors find that current models fail to truly delete such information, even when they claim to have forgotten it. Simply removing messages that match a deletion request is insufficient, because message dependencies within conversations cause information to persist. The authors propose the DeLLM framework, which dynamically builds relevant context for each query and maintains a provenance graph to decide which messages must be removed, and report a high deletion rate while maintaining utility.
In briefLonger Records, Broader Invariance: Scaling Pitfalls in Longitudinal Contrastive Learning
SummaryThis paper studies how contrastive learning on longitudinal data, which treats observations from the same person as positives, can erase meaningful behavioral change as records grow longer. The authors separate record span from supervision span and find, using in-home sensing records up to 2.7 years, that broader supervision suppresses recoverable changing-state information. They also report a prospective replication in 199 GLOBEM participants.
WhySeparates record span from supervision span, showing longer histories broaden invariance, with reversal via same-person contrast and prospective replication in 199 participants.
Rameen Mahmood, Xuhai "Orson" Xu, Zachary Beattie et al.
In briefWhy High-Magnification Knowledge Transfers: Cross-Resolution Distillation in Whole-Slide Imaging
SummaryThis paper studies cross-resolution knowledge distillation, which transfers high-magnification representations to improve low-magnification whole-slide pathology analysis. Controlled experiments across ten pathology cohorts covering classification, grading, and survival prediction show that adding teacher regional means to native low-magnification features improves downstream performance in all ten. The authors find that lower reconstruction error does not consistently yield better downstream results, so reconstruction error alone is not a sufficient measure of cross-resolution transfer.
In briefFrom Retrieval to Reasoning: Reliability-aware agents predicting drug mechanisms from cell imaging data
SummaryThis paper addresses predicting a compound's mechanism of action from Cell Painting, a morphological imaging assay. Existing methods match nearby reference perturbations in feature space, but these neighbors are often noisy due to batch effects and cytotoxicity. The authors propose PhenoAIR, a multi-agent framework that treats retrieved neighbors as uncertain evidence, weighted by offline calibration. Evaluated on a benchmark built from JUMP Cell Painting profiles, it reportedly outperforms representation-matching and LLM-based baselines.
In briefCAMEO: Background-Replacing Augmentation Guided by Class Activation Maps for Skin Lesion Diagnosis
SummaryThis paper introduces CAMEO, a framework that uses stable explanations from class activation maps to separate skin lesions from their backgrounds in dermoscopic images. It then replaces the background with realistic synthetic skin while leaving the lesion unchanged. On HAM10000 and dark-skin ISIC images, CAMEO kept accuracy while reducing background-driven errors nearly fourfold and making model attention more consistent when backgrounds changed. The authors conclude that this improves robustness and fairness without costing accuracy.
In briefCalibrated to Whom? Persona and Language Effects on Cultural Values in JEV
SummaryThe paper audits the cultural values of JEV, a decision-only language model that outputs option probabilities rather than text, using the Values Survey Module 2013. It tests Saudi and American personas, with and without persona, in English and Arabic. With Saudi personas, answers shifted toward the human Saudi-US difference, reproducing about 87% in English and 62% in Arabic.
In briefWhere physics should enter a molecular crystal generator.
SummaryThe paper introduces CrystAF, an all-atom crystal flow-map generative model, and uses the UMA interatomic potential to compare where physics can be added during training, post-training, or inference. Post-training on physical preferences improves molecular validity and crystal packing without changing sampling, while UMA relaxation at inference repairs local clashes but slows generation by 6 to 26 times. The two approaches are complementary, and the same post-training strategy also improves the Clari-M and MolCrystalFlow generators.
In briefLOCO-AdaMP: built-in LOCO inference for adaptive minipatch ensembles with better prediction
SummaryThe paper proposes LOCO-AdaMP, a minipatch ensemble framework that adaptively samples features guided by leave-one-covariate-out (LOCO) importance. It enables feature importance inference without data-splitting or model refitting, with asymptotically valid results. Experiments on synthetic and real datasets report better predictive performance, inferential power, and stability than existing methods.
In briefReward-Rate Policy Gradient for Time-Efficient Machine Learning Engineering Agents
SummaryStandard reinforcement learning assumes every action takes the same amount of time, which fails for agents whose steps, such as data loading and model training, vary in duration. The paper proposes Reward-rate Policy Gradient (RPG), which optimizes long-term reward per unit of time and charges each action for its elapsed time. Theoretical analysis in a bandit setting supports that RPG approximates the optimal reward rate. Applied to a small language model (Qwen3.5-4B), RPG earns higher rewards within a fixed time budget than vanilla RL on MLE-Bench and NanoGPT, with margins of 19.2% and 85.7%.
In briefARCagent: An Adaptive Retrieval Calibration Agent for Clinical Question Answering
SummaryThis paper presents ARCagent, a clinical question-answering agent for ME/CFS, a disease where diagnostic frameworks coexist and guidelines contradict each other on treatment. It combines a 1,706-chunk knowledge base from 10 sources with a structured inter-guideline conflict registry, and a retrieval pipeline that re-ranks evidence using query-specific focus and conflict signals. The authors evaluate it with an LLM-as-Judge benchmark and report 95.3% accuracy, outperforming all base LLMs.
In briefConstrained stimulus generation to reveal feature selectivity in early human visual cortex
SummaryThis paper introduces two methods, RF-DiVE and RF-GO, for generating most-exciting inputs (MEIs) that reveal feature selectivity of single voxels in human visual areas V1 through hV4, using fMRI data from the Natural Scenes Dataset. Both methods use a population receptive field-constrained encoding model, with RF-DiVE pairing it with a pretrained latent diffusion model and RF-GO using regularized gradient ascent. The generated MEIs show consistent structure within the receptive field, suggesting selectivity for local features such as contour, color, and texture, and elicit higher model-predicted responses than the most activating natural images.
In briefFinite-sample theory for fitted Q-iteration when decisions are whole functions
SummaryThis paper studies offline reinforcement learning where each action is a function, such as a radiation fluence map or a smooth robot trajectory. It analyzes finite-sample guarantees for fitted Q-iteration with functional actions, using smoothness-regularized policy search under a critic-relative coverage condition that does not require an action density. The main theorem bounds the learned policy's regret relative to the best policy in a fixed smooth class. Numerical experiments show learned functional-action policies outperform constant-action policies.
In briefPersona Dosing: Calibrated Steering to Set How Strongly a Language Model Expresses a Trait
SummaryThis paper studies persona dosing, which controls a language model's persona using a trait description and a requested mean intensity rather than a raw steering coefficient. The authors introduce PersonaDose, which specializes a shared, description-conditioned FLAS controller on persona responses and calibrates its flow time against measured trait expression. Across three models, it raises core-trait expression above contrastive activation addition at a coherence floor of 75. The authors note that the behavioral range a controller learns is separate from how accurately requests land within that range.
In briefStealth depends on the observer: how event representations hide timing attacks in event cameras
SummaryThis paper studies how timing perturbations in event-camera streams can be hidden from a downstream observer depending on how that observer accumulates time. The authors propose Null, a gradient-guided attack that retimes events within a protected accumulation window so the accumulated tensor stays unchanged. Experiments on DVS Gesture and DailyDVS-200 report high attack success, and perturbations hidden from one observer become visible under other temporal views.
Shoaib Ahmed Dipu, Md. Shaown Miah, Kamrul Hasan et al.
In briefTarget leakage from labeled support sets in relational in-context learning, with detection and mitigation
SummaryThis paper studies a failure mode in relational in-context learning, where labeled support examples contain target-derived features that are unavailable for queries. The authors build 14 synthetic leaker types covering 20 columns and evaluate a frozen relational encoder with an ICL head on held-out RelBench databases. Integrated Gradients is used to rank and remove suspicious columns. Target-table leakers cause the clearest performance degradation, while one- and two-hop leakers are not consistently used by the model, and removing top-ranked columns partially recovers performance where leakage matters most.
Roshan Reddy Upendra, Alexandre Dorais, Joe Meyer et al.
In briefLEGO-Anything: Coding Agents Rebuild 3D Scenes from Images
SummaryThis paper presents LEGO-Anything, a framework where a coding agent iteratively writes and runs Blender code to turn a single image into an editable 3D scene program. The authors also introduce LEGO-Bench, a simulator-grounded benchmark of 208 images from 104 indoor and outdoor scenes that scores artifact validity, geometry, and appearance separately. The best evaluated agent still shows large gaps between producing valid scenes and faithfully recovering geometry and appearance, so the authors propose LEGO-Plugin, a training-free harness that improves all six models. Reconstructed scenes are also tested on object detection, instance masks, and relative depth, where they perform non-trivially but below specialized vision models.</summary>
In briefNeural Succession: An Ecology-Inspired Theory of Forgetting in Continual Learning
SummaryThe paper proposes Successional Learning Theory (SLT), which frames continual learning as an ecological process in which a current representation is a resident community and a new task is an invader that can displace it. Its key measure, directional pre-invasion compatibility, is computed on the resident model before the new task is learned and is strongly negatively correlated with later forgetting (r=-0.789 across 20 Split-CIFAR-10 transitions). The authors position SLT as a pre-adaptation diagnostic that complements replay, regularization, and projection methods.
WhyProposes a mesoscopic ecological-succession theory of continual learning with pre-invasion compatibility predicting forgetting, validated across eight experiments and multiple datasets.
Shoaib Ahmed Dipu, Md Salman Shamil, Sayeed Shafayet Chowdhury
In briefOTT3R: Fast 3D Reconstruction and Dataset Generation Using About 1% of the Compute
SummaryThis paper presents OTT3R, a knowledge distillation framework that compresses a 959M-parameter 3D reconstruction model into a 102M-parameter student, achieving up to 7x faster inference while training at 1.6% of VGGT's compute. It also introduces an RGB-only pseudo-label pipeline that generates dense point maps and camera poses for large image collections as an alternative to COLMAP. The authors report that a domain-specialized student is more accurate than COLMAP on 7-Scenes while running at far higher throughput.
In briefLatentSift: Token-Efficient Filtering of Software Engineering Agent Candidates Using Policy Hidden States
SummaryLatentSift is a filter that selects the best candidate trajectories from software engineering agents without an extra LLM pass or test execution. It uses hidden states the policy already produces during generation, comparing them with positive and negative state banks from policy training and combining the distances with a learned linear score. On SWE-bench Verified, it cuts execution-free verifier tokens by 66.6–81.0% and total verification tokens by 49.1–62.1% at K=16, while hybrid Best@16 matches or improves on each reference workflow, rising from 59.26% to 60.06% on DeepSWE-Preview.
Yuning Han, Yangchenchen Jin, Tyler Jandreau et al.
In briefAdaKerNet: A Learnable Kernel Decoder for Predicting from Frozen Multimodal Model Representations
SummaryAdaKerNet is a task-adaptive neural kernel decoder that works on frozen representations from multimodal large language models without needing access to their parameters. It combines learnable multimodal features, a reference kernel prior, and a lightweight neural predictor, optimized jointly. Experiments across four MLLMs and text, audio, image, and tabular inputs report consistent gains over baseline decoders when labels are scarce, with average error reductions of up to 41%.
Konstantinos D. Polyzos, Eleni Oikonomou, Tara Javidi
In briefCross-attention models reveal how the brain dynamically routes video information across visual areas
SummaryThe paper extends cross-attention encoding models to natural video, using per-parcel attention to route features from the self-supervised video model V-JEPA-2 across space and time to predict fMRI responses to short clips. Joint spatiotemporal attention improved predictions of brain responses to held-out videos, most consistently in lateral and dorsal visual areas. The method yields interpretable attention maps that follow moving objects, and maps from face-, body-, and scene-selective networks weight content consistent with their expected selectivity.
In briefEngineering Simplicity: Simple Interfaces Help LLM Agents Make Better Auction Choices
SummaryThis paper studies whether interaction formats and textual scaffolds help large language model agents make better decisions in auctions and matching markets with known optimal strategies. It compares interfaces that ask for a complete bid or ranking against sequential interfaces that make safe choices easier to identify. An ascending auction interface substantially reduced bid deviations across four model families, while some prompts improved choices and others worsened play. The authors also find that better verbal explanations do not always accompany better realized choices.
In briefCompressing Video History into Compact Memory for Longer Generation
SummaryStandard video generators struggle to keep information from earlier frames as generation grows longer, and frame-selection methods can discard needed content. The paper proposes Prediction-Aligned Context Compaction (PACC), which trains a learned compressor via on-policy distillation, using a frozen generator as both student (conditioned on compressed memory) and teacher (conditioned on full history). On MBench, PACC outperforms the strongest baseline by 6.63 points on Causal-rCM and 3.19 points on Causal Forcing. On VBench-Long with MovieGen prompts, it produces minute-long videos with competitive generation quality.
In briefImproving Nearest Neighbor Graph Indexes with LLM-Guided Pruning
SummaryGraph-based approximate nearest neighbor search is widely used for semantic search, but its indexes are built from geometric relationships while results are judged by semantic relevance, creating a mismatch. This paper proposes LLM-Guided Graph Pruning (LGP), which uses LLM reasoning to refine an existing graph index by replacing structurally low-value neighbors with semantically useful alternatives while preserving sparsity and navigability. Experiments on semantic retrieval benchmarks report that LGP improves end-to-end retrieval over vanilla greedy graph search and LLM-based reranking on indexes such as DiskANN and HNSW.
In briefHyperZip: Faster lossless data compression using diffusion LLMs personalized by hypernetworks
SummaryThis paper presents HyperZip, a lossless data compression framework built on large language models. It uses diffusion-based LLMs with multi-token prediction to speed up compression. The authors note that increasing decoding throughput in diffusion-based compression lowers the compression rate, so they use a hypernetwork to generate data-specific updates that adapt the model without costly fine-tuning. Experiments reportedly show a better balance of compression rate and speed than state-of-the-art baselines.
In briefExplaining Models Through Their Training, Applied to TCR-Epitope Prediction
SummaryThe paper introduces explainability from training (EFT), a model-agnostic method that traces how model interpretations change during training to show which features models rely on and how they organize them as predictive evidence. EFT is applied to four TCR-epitope prediction models, and a new benchmark, TCR-XAI2, is built with 388 experimentally resolved TCR-epitope structures plus structures predicted by several tools. The authors report that CNN and transformer models follow distinct learning trajectories, that TCR alpha and beta evidence can conflict during learning, and that real and predicted structures show different feature preferences.
In briefStructRL: Online Reinforcement Learning with Structured Rewards for Long-Task Robot Control
SummaryThis paper presents StructRL, an online reinforcement learning framework for improving vision-language-action models on long-horizon tasks. It splits each task into verifiable subtasks and gives intermediate rewards only after prerequisite subtasks are completed, scaled by completion pace. Experiments on RoboCasa365 and LIBERO-Long with GR00T-N1.5 and pi 0.5 show it consistently outperforms the evaluated online RL baselines.
In briefDesigning Representations Alongside Generation: Cross-View Class-Token Alignment in Diffusion Transformers
SummaryThis paper proposes training diffusion models to learn stronger semantic representations directly, not just as a by-product of generation. Each image is observed twice with independent noise, and a student class token is aligned with a stop-gradient EMA teacher target from the other view, optimized jointly with flow-matching and patch objectives. Reported results include ImageNet linear-probing gains of 9.4% (class token) and 10.1% (mean-pooled patches), a 3.6 mIoU improvement on frozen VOC2012 segmentation, and comparable generation FID. In text-to-image training, FID drops from 2.52 to 2.37.
In briefDeepRewind: Predicting and Repairing Premature Commitments in Deep Research Agents
SummaryDeep-research agents run long investigations of searching, evaluating evidence, and revising beliefs, but they can commit to claims before enough evidence exists. DeepRewind is an additive control layer that tracks the agent's evolving knowledge as a typed graph and uses a prompt-based world model to predict the impact and reversibility of intermediate conclusions. A controller blocks risky commitments, and a consistency monitor performs dependency-aware rollback when later evidence invalidates them. On DRBench and LiveDRBench, it improves insight recall by 3.6 points and reduces premature commitments by 59.1% relative to Open Deep Research.
Amirhossein Abaskohi, Amirhossein Dabiriaghdam, Lele Wang et al.
In briefThuRunel: Dynamic Decoupling for Structured Advisory Dialogue
SummaryThis paper presents ThuRunel, an AI advisory agent for high-stakes fields such as medical aesthetics, legal consultation, and educational planning. It combines a finite-state belief management framework, a chain-of-thought teacher synthesis protocol, and learned generation adapters to decide what to ask, when to stop, and what to forward to a specialist. Compared with eleven baselines, it consistently improves elicitation completeness and the quality of specialist briefs.
In briefLinear-time architectures adapted for tabular in-context learning
SummaryTabular foundation models perform well by conditioning on labelled examples in context, but softmax attention limits their use on large datasets. The paper compares linear sequence mixers for this task and finds that the causal DeltaNet performs best, though it degrades beyond 2-4 times the pretraining context length due to instability in its recurrent state rather than capacity limits. The authors stabilise length generalisation with a time-dependent decay schedule and reintroduce non-causality by reading out from the final state, closely matching a softmax attention baseline on OpenML-CC18 and TabArena.
David Schnurr, Felix Sarnthein, Thomas Hofmann et al.
In briefATLAS: Keeping Latent Structure Aligned for Reliable World-Model Planning
SummaryLatent world models plan using the geometry of their representations, but regularizing only the overall latent distribution does not preserve the relationships between states that guide action choices. The paper proposes ATLAS, a training objective that transfers pairwise structure from an informative encoder to the planning latent while calibrating its marginal distribution via one-dimensional Wasserstein-2 transport. Applied to LeWM, ATLAS improves mean goal-reaching success on PushT, TwoRoom, and OGBench-Cube, with the largest gain on higher-novelty TwoRoom episodes, and yields lower multi-step prediction error.
In briefReducing the Adaptation Gap via Reachable Fisher Geometry
SummaryThe paper studies the reachable Fisher, the full-model Fisher pulled back through an adapter's Jacobian, to explain what its trace reveals about parameter-efficient fine-tuning. The trace can be computed from score-gradient norms and serves as a scalable diagnostic, but it cannot certify local gap behavior or matrix alignment. In experiments on 306 single-seed models, trace matching reduced the best-worst subgroup gap in all 30 combinations, while held-out audits showed operator discrepancy and squared-Frobenius results improved in different numbers of cases.
In briefPILLAR: A private keyword-index lookup system for retrieval-augmented generation
SummaryRetrieval-augmented generation normally reveals a user's query to the party hosting the document corpus. PILLAR is a privacy-preserving RAG system built on private information retrieval, so the server learns neither the query terms nor the access pattern. It runs two-stage hybrid retrieval: a sparse stage uses a small fixed number of PIR queries over precomputed BM25 scores to find candidates, and a dense stage fetches only those candidates' embeddings for local re-ranking. Two protocols are presented, PILLAR-Bin for lower latency and PILLAR-Tree for the highest retrieval quality.
Truong Son Nguyen, Daniel Blackley, Ni Trieu et al.
In briefFew-step text-to-speech via local flow-map distillation and scheduling
SummaryFlow-matching text-to-speech models produce high-quality audio but need many neural function evaluations to generate speech. This paper proposes Local Flow-Map Distillation (LFMD), which builds training targets without integrating teacher trajectories, and derives a sampling schedule (TD-DP) from teacher dynamics and map consistency. The method supports multiple evaluation budgets with one cost graph and adds alignment-aware self-distillation for the single-step regime. On Seed-TTS and LibriSpeech-PC, LFMD improves low-NFE synthesis over a matched integral-distillation baseline.
In briefToward an AI software factory for data systems that speeds up the whole development lifecycle, not just coding
SummaryThe paper argues that AI coding tools speed up only the coding stage, so their effect on the whole software development lifecycle is limited. The authors describe progress toward an AI software factory that aims to accelerate targeting, coding, reviewing, and operations, and that generates metadata used to fine-tune models and update a world model. The work focuses on data systems and evolutionary coding tasks with measurable objectives. It reports scaled deployments across tens of Microsoft repositories, with 3x engineering efficiency over agentic coding and up to 22x token efficiency, along with several open challenges.
Anna Pavlenko, Bogdan Crivat, Brandon Haynes et al.
In briefPeriodic Weak Spots: Chunked Cache Compression Phase Effects
SummaryThis paper studies how chunked KV-cache compression, which merges consecutive tokens into fewer cache entries, makes retrieval accuracy depend on a token's phase relative to compression-window boundaries. The authors call this phase sensitivity and report differences of up to 40 percentage points across phases in some large open-weight models. They pretrain transformers across several compression designs to reproduce the effect and use causal interventions to show that attention components specialize by source phase. The paper argues that high average benchmark accuracy can hide systematic positional failures.
In briefStateTape: Keeping Long-Running Coding Agents' Context Accurate as Code Changes
SummaryCoding agents built on large language models struggle over long tasks because their context keeps growing with every observation. This paper proposes StateTape, which rewrites the agent's context based on changes to the code repository rather than context length, using a symbol-level code graph and a tape that marks which symbols each write changed. The authors also introduce TraceBench, a benchmark for checking what an agent holds against what it needs. Experiments across six coding agents and three edit-heavy benchmarks report higher resolve rates with little computational overhead.
In briefTraining LLMs to Say When They Know They're Being Evaluated
SummaryLanguage models may act differently during audits when they are aware they are being evaluated, which makes such behavior hard to measure. This paper introduces verbalization training (VT), which uses spontaneous statements of evaluation awareness as evidence and trains models with reinforcement learning to verbalize it more readily. Across three models, VT raised verbalized evaluation awareness 2.4 to 2.9 times and transferred to held-out agentic settings, while latent awareness and behavior stayed largely stable.
WhyProposes a new training method to make models verbalize evaluation awareness, with a causal synthetic-document experiment; novel and fairly well supported, but impact is mainly within AI safety evalua
In briefPyroStack: A Multi-Band Spatio-Temporal Sub-Daily Wildfire Dataset for the US
SummaryThis paper presents PyroStack, a dataset combining satellite fire observations with atmospheric reanalysis, vegetation, fuel, and topographic data. It covers 6994 wildfires in the contiguous United States and Alaska from 2012 to 2024. The data offers spatial resolutions from 30 m to 9 km, hourly temporal resolution, and fire progression at 12-hour intervals. It supports both physics-based and machine learning approaches for benchmarking fire spread models.
Arya Kondur, Giosue Migliorini, Cameron Schmitt et al.
In briefFRAC: fractional-dynamics state space model for long sequences
SummaryThis paper introduces FRAC, a selective state space model (SSM) based on fractional dynamics. It replaces the exponential forgetting of common SSMs with power-law long memory, approximated by a log-spaced sum of finitely many exponential modes. The design supports parallel training and bounded-state autoregressive decoding. The authors report that FRAC improves long-context performance over state-of-the-art SSM baselines, including in 1.3B-parameter language modeling, while remaining competitive on short contexts.
Ivan Kobyzev, Abbas Ghaddar, Ali Nasiri-Sarvi et al.
In briefLearning Sample Importance by Parameterizing Dual Variables in Everywhere Learning
SummaryThis paper studies Everywhere Learning, a framework for training AI models under constraints that must hold across the whole data distribution. The authors learn the functional dual variables as a parametric function of each sample, so the multiplier can be evaluated on unseen data. They show that the error in the recovered primal solution is controlled by how well this function class approximates the optimal statistical multiplier. Experiments on several everywhere learning tasks support the theory.
In briefCheatBench: Measuring Reward Gaming in AI Agents
SummaryThis paper examines how AI agents trained to maximize rewards may engage in cheating, such as accessing unauthorized information, evading monitoring, or breaching sandbox protections. The authors introduce CheatBench, a benchmark spanning mathematical research, knowledge work, coding, and visual tasks, whose environments pair difficult assignments with chances to cheat. It supports comparisons across models and task categories, and it is publicly released at https://cheatbench.ai.
In briefLow-Cost, Powerful Tests for Supervised Subspaces in PLS Regression
SummaryPartial Least Squares (PLS) regression extracts a few outcome-aligned directions from high-dimensional predictors, but inference on its fits has been costly, biased, or missing. The paper reduces inference to held-out OLS refits of the supervised subspace and proposes two tests based on held-out correlations: a corrected asymptotic t-test and a permutation test. The authors validate the methods on synthetic data, two NIR chemometric datasets, and cross-lingual word-embedding regressions, reporting more power than CV-permutation-Q² at a fraction of its cost.
In briefEAGLE: An Entropy-Aware Add-on for Better Object Localization with Partial Polygon Supervision
SummaryThe paper proposes EAGLE, a lightweight post-training model head for object detection that uses information theory and statistical shape analysis to quantify per-class geometric complexity. It refines edges and rescores localization using entropy-informed updates. Experiments on MS COCO and Pascal SBD show the entropy measure predicts shape reconstruction error and accuracy differences between mask-derived and box predictions. EAGLE improves already state-of-the-art detectors by 0.9 mAP50-95 points on MS COCO with minimal inference overhead.
In briefHeurEvo: Evolving Hybrid Solver-Augmented Heuristics with AI Agents for Time-Limited Optimization
SummaryHeurEvo is a framework that uses AI agents to automatically discover optimization algorithms by jointly evolving their high-level structure, implementations, and a shared pool of reusable components. A planner decides which components to use and how to allocate runtime, while a coder turns the plan into executable code. On combinatorial benchmarks and MIPLIB instances, it finds high-quality solutions within tight runtime budgets, often matching or beating state-of-the-art solvers, and it improves best-known results on some nonlinear geometry problems such as hexagon packing.
Feijie Wu, Hugo Barbalho, Konstantina Mellou et al.
In briefUniversal Classifier: One Graph Model for Node, Edge, and Graph-Level Tasks
SummaryThe paper argues that existing graph foundation models are mostly limited to node-level prediction or require fixed feature dimensions. The authors propose the Universal Classifier (UC), which accepts arbitrary feature and class counts and unifies node, edge, and graph-level tasks under a similarity-based classification objective. Experiments report strong zero-shot transfer performance across node, link, and graph-level tasks.
Ben Finkelshtein, Andr\'{e} Linhares, Petar Veli\v{c}kovi\'{c} et al.
In briefMoRE: Scaling Mixture-of-Experts Models with Hardware-Aware Low-Rank Routing
SummaryThe paper proposes MoRE, which factorizes the router weight matrix of Mixture-of-Experts layers into low rank, reducing routing cost to O((h+M)r). The authors provide theoretical proofs on routing expressivity and load balance, and design a fused Triton kernel to cut memory operations at inference. Experiments report better memorization on a phonebook task and improved knowledge-intensive Q&A performance, with reasoning ability matched.
WhyLow-rank factorized MoE router cuts routing cost; supported by proofs, synthetic memorization tests, and pretraining benchmarks, a substantive improvement over standard routers.
In briefTransfer-Learning PINNs for Estimating Spatially Varying Heat Transfer Coefficients
SummaryThis paper studies physics-informed neural networks for estimating spatially varying convective heat transfer coefficients in jet impingement cooling, inferred from interior temperature measurements. The estimates are compared with CFD-CHT reference profiles at several Reynolds numbers, and an adjoint finite-element method serves as an independent comparison. Generalized representations handle spatially varying profiles without a fixed quadratic form, and transfer learning cuts target training time substantially, though accuracy depends on source-target compatibility.
In briefBeyond Trees: Learning Mixed-Topology Feature Graphs with Adaptive Graph Sparse Autoencoders
SummarySparse autoencoders can extract interpretable features from large language model activations, but existing structured methods allow only single-parent trees or forests, and post-hoc graphs neither guide feature learning nor reliably recover relations. This paper introduces the Adaptive Graph Sparse Autoencoder (AG-SAE), which treats each feature's complete parent set as a structural hypothesis, letting the evidence choose zero, one, or multiple parents. Experiments show exact mixed-topology recovery in a controlled toy model and, on real LLM activations, better relational reliability and semantic validity than baselines, along with stronger feature-level causal interventions.
In briefSafe, fully decentralized multi-agent reinforcement learning for network control
SummaryThis paper presents a safe, fully decentralized multi-agent reinforcement learning algorithm for discrete-time control problems on networks, such as persistent monitoring. It combines a graph encoder and a state estimator fed by local observation histories, with a discrete-time control barrier heuristic that limits harmful policies. In a custom simulation, the method achieves 26.3 percent lower average uncertainty than a centralized policy and is within 1 percent of a more complex attention-based algorithm.
In briefA study of the surge in anti-Semitic posts on German social media after the October 7 attacks
SummaryThis paper examines how the Hamas attacks on Israel on October 7, 2023, changed anti-Semitic content in German Facebook and Telegram discussions about Judaism and Israel. The authors used large language models to detect 26 anti-Semitic categories across 125,718 posts from three months before and after the event. They found anti-Semitism rising significantly on both platforms, and about ten times more prevalent on Telegram than on Facebook.
In briefHow Much Prompt Is Enough? Blackbox Minimization of Few-Shot Examples in LLMs
SummaryThe paper presents a blackbox framework that reduces few-shot prompts to their necessary minimal subset. Experiments show few-shot exemplars can be cut by about 65.3% in character count while fully preserving propositional output fidelity. Models tend to keep logical identifiers and constraint declarations while discarding natural language prose and cross-prompt relational annotations.
In briefRobustness of EEG foundation models to electrode montage changes
SummaryThis paper examines whether EEG foundation models give stable representations and predictions when electrode montages change. Four models with different montage-handling designs are tested on resting-state eyes-open/closed and motor-imagery classification under channel reduction. The authors measure functional robustness via linear probe generalization and representational robustness via embedding similarity, finding the two can diverge. They also show that how outputs are aggregated strongly affects functional robustness.
Jakob Steglich, Justus Meyer zu Bexten, Shakiba Moradi et al.
In briefGNA: A retrieval layer that assembles past neighbors at two granularities to improve multivariate time-series forecasting
SummaryThis paper presents GNA (Granular Neighbor Assembly), a retrieval layer for time-series forecasting models that gathers similar past examples at two levels: whole past windows and per-variate best matches. A learned gate decides how much to trust these retrieved futures compared with a persistence forecast and the backbone's own prediction. With the same lookback and retrieval settings across datasets, GNA improves two Transformer backbones in 85 of 96 dataset-horizon settings and achieves the lowest MSE on 8 of 12 standard benchmarks.
In briefProofs without nominals: Gödel's ontological argument and open questions of the Monatshefte Notes
SummaryThis paper examines the shallow embedding of higher-order modal logic used in Benzmüller and Scott's Notes on Gödel's and Scott's ontological argument variants, showing its proofs can reach beyond the modal object language. The author argues every statement in the Notes has a proof in the embedding that substitutes nothing outside the object language. It also settles the Notes' three open questions, with results verified in Isabelle/HOL and Lean 4.
In briefIllusory Truth or Mere Exposure? Whether LLM agents repeat a human bias when judging social media claims.
SummaryThis paper studies whether LLM-based generative agents in social media simulations show the Illusory Truth Effect, where repeated exposure to a claim raises its perceived truth. Four models were tested in a two-phase experiment within a news feed, comparing ratings for repeated and unseen statements and separating genuine truth effects from mere exposure effects. The results differ by model: Gemma-3 shows a genuine effect, Qwen2.5 shows only mere exposure, GPT-5-nano shows no truth repetition effect, and Llama-3.1 shows a small truth boost. The authors caution against assuming uniform replication across LLMs.
In briefFrom Surfaces to Volumes: Aligned Geometry for Protein Representation Learning
SummaryThis paper introduces Protein-TetSphere, a residue-level volumetric representation of proteins. Each protein chain is tetrahedralized and the local volumes of each residue are registered to a shared fixed-topology reference in a common Laplacian basis. The authors report improved results over surface-based representations on ligand-binding pocket classification, protein-protein interface prediction, and de novo binder design.
In briefStochastic optimization under power-law spectra: tight bounds and shuffling
SummaryThis paper extends power-law spectral theory, previously shown for deterministic gradient descent, to stochastic gradient descent (SGD), showing the same spectral exponents govern stochastic dynamics. For isotropic Gaussian data, it analyzes data shuffling and derives exact constants. The authors conclude that Single Shuffle is strictly better than Flip-Flop and IID sampling.
Christopher De Sa, Thomas Dybdahl Ahle, Yaroslav Bulatov et al.
In briefStudy of how language models build internal rule-following circuits from in-context examples
SummaryThis paper examines how large language models' internal mechanisms for abstract rule-following activate as the number of in-context examples grows. The authors trace a three-stage circuit (abstraction, induction, retrieval) across three model families and find it is detectable and functional before accuracy becomes high. They report that per-head causal contribution grows up to 8x from 1- to 10-shot, and that interventions such as activation patching and injected function vectors can raise accuracy. The authors conclude that the infrastructure for this behavior is present in the weights before any demonstrations.”
In briefOTROPE: Robust Off-Policy Evaluation of LLMs Using Optimal Transport
SummaryThis paper proposes OTROPE, a method for evaluating a newer large language model when human-labeled data is scarce and the behavior and target distributions differ. It uses optimal transport to align labeled behavior-model samples with unlabeled target-model samples in a semantic space, and combines the corrected residuals with proxy predictors. The authors provide theoretical consistency and convergence results, and report experiments showing it outperforms baselines and lets ensembles of weaker evaluators approach or sometimes exceed stronger ones.
In briefPaired Multimodal Scaling Laws: how the share of paired multimodal data shapes loss in classification, and a new four-term scaling law
SummaryThe paper notes that existing multimodal scaling laws are fitted empirically after testing and never vary how much data is paired at a fixed budget. The authors train models in three environments, sweeping data sizes and pairing amounts, and find that pairing ratios strongly affect loss, tied to the task's synergistic information. They propose a law summing four power laws for redundancy, each modality's unique information, and synergy, reporting 3.2% fit-test error versus 10.4% for the best paired extension of prior laws.
In briefWhy LLM post-training updates are sparse: a Fisher information view
SummaryRecent studies found that parameter changes during large language model post-training concentrate in a small set of coordinates. This paper analyzes the pattern with the diagonal model Fisher, which measures how sensitive the output distribution is to each parameter. The authors show theoretically that small diagonal Fisher values yield small expected gradients, and in reinforcement learning and on-policy distillation experiments, fixed sparse masks chosen from the initial Fisher retain much of the gain from full training.
In briefCyFA: Linear sequence model that stores past associations in relative-time slots
SummaryThe paper introduces CyFA (Cyclic Flow Attention), a linear RNN whose memory is partitioned by relative time. A learned clock cyclically transports key and value states so stored associations are organized across relative-time slots. Pretraining experiments at 400M to 1.4B parameters report better recall-intensive performance, competitive language modeling, and high computational efficiency.
WhyProposes a novel relative-time-partitioned memory with cyclic transport and an exact chunk-wise reformulation, validated in 400M–1.4B pretraining against KDA on FDA with timing results.
In briefReducing deceptive safety alignment in reasoning AI models
SummaryLarge reasoning models are usually trained with reinforcement learning that rewards final answers, giving little direct supervision to their intermediate reasoning. This can produce deceptive safety alignment, where the reasoning and final answer send inconsistent safety signals. The authors introduce a metric, DSAR, to measure this inconsistency and find it is common across models and benchmarks, and worse under prefilling attacks. They propose SARA, an RL method that rewards safety-aware reasoning and safe answers, and report that it significantly reduces the problem while preserving helpfulness.
Xiangyu Zhou, Saleh Zare Zade, Rafi Ibn Sultan et al.
In briefPopulation Fidelity: Testing How Well LLMs Represent Human Groups
SummaryThis paper introduces Population Fidelity, a framework for judging whether responses generated by large language models represent a population. It assesses group-level accuracy, the amount of between-group variation, and the structure of that variation. The authors reproduce a prior study of machine bias in LLM survey responses and test cultural fine-tuning, finding it improves agreement with the survey center but not the representation of within-population differences.
Neemias B. da Silva, Martin Lukk, Ali Sutani et al.
In briefSigned Geometry of One-Shot Recourse: When a Single Step Succeeds
SummaryThe paper studies one-shot recourse, where a rejected user moves along the score gradient by a promised distance to reach the decision boundary. It proposes a signed-curvature criterion, showing that the step ends on the favorable side when path curvature is nonnegative, and analyzes the effect of extra model queries and conformal calibration. Experiments on 80 shallow models compare asymmetric and symmetric curvature penalties on datasets such as Fashion-MNIST and COMPAS.
In briefAction Chunking PPO with Feedback Correction (paper on robot control learning)
SummaryThis paper proposes ACPPO, a PPO extension whose actor outputs action chunks while keeping a standard state-value critic, avoiding chunked Q-functions. Its variant ACPPO-Corr adds a stepwise feedback corrector that adjusts planned actions online within each chunk. Evaluated on 25 simulated robotics tasks from IsaacGym and Bi-DexHands, ACPPO-Corr reports the strongest aggregate performance among compared methods. The authors conclude that chunk-level planning paired with closed-loop correction can be effective in online PPO.
WhyCombines chunked actor with stepwise feedback correction in PPO, validated across 25 simulated tasks with ablations.
In briefOLIVE: Distilling by Having Teachers Continue Student-Generated Prefixes
SummaryThe paper proposes OLIVE, which at each iteration has the student generate a prefix, the teacher continue it autoregressively, and the student is trained with cross-entropy on the teacher's tokens. The authors report higher reasoning performance than on-policy distillation at comparable GPU-hour cost, and a 23.8% reduction in training time with an asynchronous implementation. On ScienceWorld, training with text from GPT-5.4-mini alone outperforms offline SFT from the same teacher by 13%.
In briefCoRe: Co-Evolving Reward Models Against Latent Reward Hacking in Video Diffusion
SummaryThis paper finds that optimizing video diffusion models against a fixed latent reward model causes latent reward hacking, where predicted rewards stay high while perceptual and motion quality decline. The authors attribute this to distributional escape, in which the generator drifts beyond the reward model's training support within a few hundred updates. They propose CoRe, which continually refits the reward model on the generator's current samples while anchoring it to real-video preferences. Experiments on Wan2.1-T2V-1.3B show CoRe improves generation quality over the pretrained model and prior alignment methods and avoids the quality collapse of fixed-reward optimization..
In briefThink Before You Restore: Risk-Aware Restoration of Manchu Manuscripts
SummaryThis paper addresses blind restoration of full-page historical Manchu manuscripts, where annotations are scarce and damaged regions are unknown. It proposes SAGE-Restore, which first predicts which patches need repair using appearance and stroke-structure cues, then uses this to guide candidate generation and apply restoration only at selected pixels. The authors also introduce an evaluation protocol measuring recovery of degraded regions, preservation of intact content, and their balance. The method reports the highest R-Recovery (0.463) and RFS (0.626) with high U-Fidelity (0.968).
In briefCan representation learning continue when loss stops falling? Polar updates offer an answer
SummaryThis paper studies matrix Muon, an optimizer with polar-normalised updates, on teacher-student problems. It finds that loss can oscillate with period 2 and stay flat for thousands of steps while the learned features keep aligning with the teacher subspace. The authors give theoretical results for linear toy models and a mean-field ReLU model, and experiments across several activations and deep residual networks.
In briefCognitively specialized language models align better with matching brain systems
SummaryLarge language models can predict brain activity during language comprehension, but alignment is usually measured with one model across all regions. This study builds expert variants for six cognitive domains, including sensory, spatial, numerical, reasoning, social, and abstract processing, using prompting and fine-tuning. Each expert aligns more closely with the brain system tied to its matching domain than with other systems. The pattern holds across three base models and three fMRI datasets.
Zhivar Sourati, Mengxuan Helen Wu, Nona Ghazizadeh et al.
In briefChronoSRL: Time-Aware Geometry for Self-Supervised Robot Learning
SummaryThis paper introduces ChronoSRL, a reinforcement learning method that trains the critic's embedding distances to match the time an agent needs to reach a goal. It also predicts the distribution of goal-reaching times and time spent near the goal to favor reliable, sustained goal-reaching actions. On seven standard locomotion and navigation benchmarks, it learns faster and performs better than contrastive and survival RL baselines, even with smaller networks. On a quadruped robot sim-to-real setup, it was the only tested self-supervised method that tracked commanded velocities and goal positions and climbed the highest boxes.
WhyIntroduces temporal geometry into contrastive and survival RL representations, beating baselines on seven benchmarks and a sim-to-real quadruped task.
In briefTransversal Pooling Neural Networks: A network design that generalizes max pooling to affine group actions
SummaryThis paper introduces transversal pooling neural networks, which extend spatial max pooling to affine group actions. The authors prove equivariance to a chosen subgroup and derive explicit stability bounds for individual pooled wavelet coefficients under affine perturbations of the input. Experiments show the networks are useful in low-data settings and for predicting tropical cyclone intensification.
Emily J. King, Dustin G. Mixon, Michael Perlmutter et al.
In briefMERID: A Self-Improving AI Agent Framework for Multimodal Depression Detection
SummaryThis paper introduces MERID, a framework that uses AI agents to build pipelines for detecting major depressive disorder from multimodal data such as interview recordings and sensor measurements. The system revises its pipelines through recursive self-improvement, checking gains on small depression cohorts before carrying them forward. Experiments on depression benchmarks report that MERID achieves the best results on multiple tasks compared with multimodal and agent-based baselines, and the analysis highlights the value of acoustic and linguistic cues.
In briefHow feed-forward layers steer token behavior in transformers
SummaryThis paper studies how tokens move through transformers using control theory, treating self-attention as an interacting particle system and the feed-forward layer as an independent control. The main theoretical result shows that the feed-forward network can steer tokens arbitrarily close to consensus, regardless of the key, query, and value matrices. The authors state that the result extends to convergence into multiple clusters and to multi-head attention. They verify the results with numerical experiments and compare the theory's thresholding behavior with real-world LLMs.
Thomas Jacob Maranzatto, Semih Akkoc, Sennur Ulukus
In briefOne-step latent prediction is not a world model
SummaryThe paper examines next-latent prediction, where a one-step regression maps the current embedding to the next, and introduces LeNEPA, which applies LeJEPA's isotropy penalty to time series. The authors argue that the one-step regression identifies a conditional mean, which is a rollable transition kernel only in special cases. Analysis of linear-Gaussian, nonlinear, and non-injective observation settings, plus numerical examples on a scalar autoregression and a hidden rotation, shows open-loop error grows with horizon, and the isotropy penalty has zero partial derivative in the transition weights.
In briefMATE: Mutually adversarial self-training for unified image generation and understanding models
SummaryThis paper introduces MATE, a reinforcement-learning post-training method for unified multimodal models that combine image generation and visual understanding. Instead of the two branches cooperating, they alternately act as challenger and solver, with challenges generated from the model's own outputs and evolving across training epochs. On Janus-Pro-1B, MATE improves GenEval by 2.4 points, DPG-Bench by 1.7 points, and the average over nine understanding benchmarks by 0.7 points. It also strengthens consistency across repeated image-text cycles.
In briefBASE: batch-aware expert selection for faster Mixture-of-Experts decoding.
SummaryThis paper addresses the cost of serving Mixture-of-Experts language models in batched decoding, where different requests activate different experts and inflate the weights that must be loaded. The proposed method, BASE, ranks experts by how much their removal would change the MoE-layer output, using a lightweight linear predictor trained during calibration and custom GPU kernels. Experiments on three MoE architectures report better quality-efficiency tradeoffs without retraining.
