In briefJaxolotl: a unified high-performance benchmark for LTL-based multi-task RL
SummaryThe paper introduces Jaxolotl, a benchmark suite for multi-task reinforcement learning that uses linear temporal logic (LTL) to specify instructions. It provides a modular JAX implementation of six algorithms and four environments, with new task suites and a standardised evaluation protocol. Precompiling symbolic task representations enables end-to-end speedups of up to 220 times. The authors find that general methods struggle as the number of propositions grows, while methods that scale better rely on environment-specific assumptions and suffer from myopia.
Mathias Jackermeier, Jacques Cloete, Alessandro Abate
In briefWorldLine: A Visual Simulator That Predicts Robot Action Outcomes
SummaryThis paper introduces WorldLine, an action-driven visual simulator that predicts how robot actions will play out before physical execution. It learns manipulation dynamics from over 10,000 hours of action-free robot videos and grounds them with more than 2,000 hours of action trajectories across over ten embodiments. Across held-out and out-of-domain settings, it maintains strong visual quality and robot-motion agreement, and it predicts trajectory success for policy evaluation.
In briefA latent guidance method that combines forecasts from several time series foundation models at inference time.
SummaryTime series foundation models give strong forecasts, but their accuracy depends heavily on choices such as lookback length, covariates, horizon and training data. The paper proposes Latent Inference-Time Guidance, which adaptively combines forecasts from a pool of these models through a time-dependent latent space with independent components. The framework includes identifiability and reconstruction guarantees and keeps the models usable out of the box. Experiments on datasets of various frequencies and domains show it is competitive with traditional ensembling methods.
Chlo\'e Hashimoto-Cullen, Amaury Durand, Laurent Bozzi et al.
In briefEVO-WAM: Evolving World Action Models via Video-Action Verification
SummaryThis paper proposes EVO-WAM, a framework that lets robot world action models adapt to new tasks by learning from their own generated video-action trajectories, without extra expert demonstrations or external execution. It adds state prediction and anchored multi-frame context for autoregressive rollouts, selects task-completing prefixes with a vision-language model, and checks video-action consistency with an inverse dynamics model before iterative training. On seven unseen RoboTwin 2.0 tasks, average success rose from 26.9% to 68.0% for Cosmos3 and from 28.5% to 46.4% for DreamZero.
In briefBeyond a Better Score: A Protocol for Long-Horizon Agentic ML on Physics Time Series
SummaryThis paper introduces SIDERIUS, an LLM agent system for developing machine learning models on physics time-series data, such as signals related to dark matter, neutrinos, and gravitational waves. It combines composable capabilities, a multi-layer scientific evaluation protocol that separates valid models from gaming ones, and resource-aware exploration. On the TIDMAD benchmark, it reports a 63.8% valid-model proposal rate and a valid denoising score of 2.744, compared with 1.6% and -0.647 for a Codex coding agent using the same GPT model.
In briefPow3R-SLAM: Real-time RGB-D SLAM Using 3D Reconstruction Priors
SummaryThis paper presents Pow3R-SLAM, a real-time RGB-D simultaneous localization and mapping system that uses Pow3R for tracking and mapping. It treats depth as a prior on the network's prediction and fills sparse depth regions using two-view photometric, depth, and intrinsic data. Evaluated on 24 sequences from TUM, 7-Scenes, and Replica against MASt3R-SLAM, it reports a 1.6x faster runtime, 15% lower mean trajectory error, and denser maps. The authors note it can struggle on a small set of self-similar scenes.
Christopher Kolios, Ishaan Mehta, Sasa Janjic et al.
In briefPRISM: Multi-agent topology design guided by a privileged sibling graph
SummaryThis paper studies how to generate collaboration topologies for LLM-based multi-agent systems, deciding which agents take part and how they communicate for a given task query. It notes that existing corpora list graphs that solved a query without ranking them, so continued supervised fine-tuning lacks per-query preference signals. PRISM uses privileged information: a teacher reads a sibling graph that also solved the query, and a student matches its per-decision targets on trajectories it samples itself. The authors report that PRISM improves over every topology-design baseline evaluated across six benchmarks.
WhyUses a sibling topology as privileged teacher information to fix unranked corpora; beats all baselines on six benchmarks with matched-teacher controls supporting the gain.
In briefImproving Function-Space Flow Matching with Kernel Optimal Transport
SummaryThis paper addresses generative models for function-valued data such as time series and PDE solutions. It proposes kernel Functional Flow Matching (kFFM), which replaces independent endpoint pairing with entropic optimal transport under a kernel-induced cost while keeping the FFM architecture. On time-series and PDE benchmarks, kFFM improves distributional matching over FFM and other baselines.
Fred Xu, Thomas Markovich, Barbora Barancikova et al.
In briefUserProxyBench: Testing how well AI user simulators follow their role in agent benchmarks
SummaryThis paper introduces UserProxyBench, an evaluation layer over the tau-bench family, and the User Fidelity Score (UFS), which measures whether a simulated user follows its private instructions, scored independently of agent success. Across 375 enterprise tasks with the agent held fixed, changing only the user proxy shifted mean task reward by 15.2 points, and 24.4% of successful episodes contained a user-specification violation, mostly premature disclosure. The authors also identify a cost-fidelity frontier across seven proxies to help select the cheapest simulator meeting a required fidelity level.
In briefStudy finds gender bias is common in LLMs but varies widely across models
SummaryThis paper examines gender bias in ten large language models released between April 2025 and June 2026 by nine vendors. It uses two tests: attributing stereotyped phrases to male or female writers, and making moral judgments about abuse or torture against a woman or a man to prevent a catastrophic outcome. The direction and size of bias differ greatly across models, with some behaving in opposite ways. The authors conclude that bias auditing should be an ongoing, multi-vendor process rather than a one-time assessment.
In briefExplicit solution to the finite-horizon five-expert prediction problem
SummaryThe paper gives an explicit solution to the five-expert prediction PDE over a finite time horizon. The formula shows the adversary's rank strategy (1,0,1,0,0) is globally optimal, while the COMB strategy (1,0,1,0,1) is optimal exactly where x1=x2 and x3=x4. The solution is derived from a companion paper's geometric-stopping result via a Laplace-transform link, with optimality checks computer-assisted and the proofs formalized in Lean.
In briefdoPlan: A Variable-Horizon Dataset for Multi-Stage Language-Conditioned Driving Planning
SummaryThis paper introduces doPlan, a publicly available, human-annotated real-world dataset built on nuPlan that studies passenger language as persistent task context for autonomous driving. It contains 5,154 passenger instructions covering immediate, deferred, event-conditioned, persistent, and multi-stage intent. Evaluating four language-conditioned driving models, the authors find that sensitivity to passenger language does not reliably produce behavior matching the requested direction.
Parthib Roy, Yash Tandon, Marcus Blennemann et al.
In briefDr. OPD: Learning Which Teacher Signals to Follow for Better On-Policy Distillation of Language Models
SummaryThis paper proposes Dr. OPD, a method that improves on-policy distillation for large language models by weighting the teacher's token-level supervision. The weights are chosen by solving a bilevel optimization problem that maximizes the expected reward of the resulting student. On math and code tasks, Dr. OPD outperforms all evaluated baselines, improving average math performance by 9.7 points over vanilla OPD in strong-to-weak distillation and letting the smaller student surpass its larger teacher.
In briefRetrieval-Augmented Skill Optimization via Cross-Harness Adaptation: reusing public agent skills to improve task performance
SummaryThe paper proposes RASO, a framework that uses an external corpus of public agent skills as prior knowledge to optimize skills for a target task. It adapts retrieved skills to the target domain and harness through Cross-Harness Adaptation. RASO has two stages: RASI builds a knowledge-grounded initial skill without agent rollouts, and RASU iteratively refines the skill using retrieval guided by execution feedback. Experiments on four agent benchmarks and two models reportedly show RASO consistently outperforms baselines lacking retrieval-augmented initialization and updating.
In briefPE-EK-PINN: Scalable Physics-Informed Neural Networks Using Evolving Kernels
SummaryPhysics-Informed Neural Networks (PINNs) enforce governing equations only through loss residuals, which makes highly oscillatory wave fields hard to learn. This paper proposes PE-EK-PINN, which freezes a converged subsystem field and reuses it as an evolved kernel to represent higher-level configurations without deriving new governing equations. According to the paper, the peak number of active kernels is independent of system size and cumulative training cost falls from O(N) to O(log N). Experiments on dipole arrays, composite line-source geometries, and cross arrays show large training cost reductions with reduced or comparable relative L2 error.
In briefAuditing long-term memory evaluation: repeated judging, reader variation and negative controls
SummaryThis paper audits how a long-term memory retrieval pipeline was evaluated on 500 LongMemEval-S development questions. It tests how scores shift with repeated judging, different reader configurations, and a negative control verifier. The authors conclude the findings do not establish a new leaderboard leader or a transferable memory advantage, since all questions were used during development and no untouched holdout was evaluated.
In briefBeyond Lip Sync: Reference-Grounded Oral Refinement for Audio-Driven Portrait Animation
SummaryThis paper presents RGOR, an audio-driven lip-sync framework that aims to render the actual mouth of the person being dubbed rather than a generic average mouth. It conditions generation on enrollment frames of the same person and high-definition mouth patches, and trains against a judge that rejects realistic mouths belonging to others. The authors also propose an evaluation protocol and compare against open-source and commercial systems on held-out identities, reporting best or second-best results on most metrics.
In briefPrompt-selection sampler for GRPO using a Gaussian curriculum in arc-length coordinates
SummaryGRPO learns only from prompts whose sampled responses disagree, since all-correct or all-incorrect groups give no gradient but still use rollouts. The paper identifies arc length ψ = arcsin√p on the Bernoulli Fisher–Rao manifold as a natural coordinate for pass rates, where the expected GRPO update is near-uniform. It proposes ARCUS, a sampler that tracks prompts with a Kalman filter and selects them with an objective-matched Gaussian kernel. Across six math reasoning benchmarks and three backbones, ARCUS improves GRPO accuracy by about 2.8 to 2.9 points and dynamic sampling by 1.1 to 1.2 points, using 48–57% fewer rollouts.
In briefBrain-SAD: Brain-Inspired Safe Driving Control with Dynamic Fear Constraints on Dual Policies
SummaryThis paper proposes Brain-SAD, a brain-inspired framework for safe autonomous driving that uses constrained reinforcement learning. It notes that existing methods use static action costs and fixed feasible-region boundaries, which limits their adaptability across scenarios. Brain-SAD generates a dynamic fear signal from the current traffic scene to choose between a long-term policy for regular interaction and a short-term policy for urgent collision defense. Experiments reportedly show higher success rates, shorter task-completion and collision-recovery times, and greater reliability across intersections of varying complexity.
In briefWhen mixing data helps scaling laws: a high-dimensional regression study
SummaryThis paper studies when mixing data from different domains genuinely improves scaling laws rather than just adding samples. It analyzes a high-dimensional mixed-data regression model, deriving minimax risk bounds and deterministic equivalents for ridge regression test error. For target and auxiliary domains with aligned power-law spectra, it gives explicit scaling laws showing regimes where mixing yields a faster rate than either dataset alone. Language model experiments show a similar qualitative effect.
WhyDerives minimax risk and ridge deterministic equivalents in high-dimensional mixed-data regression, explicitly characterizing when mixtures improve scaling; supported by language model experiments.
Diyuan Wu, Lehan Chen, Theodor Misiakiewicz et al.
In briefComparing simulated and diffusion-generated images to find the best data mix for object detection
SummaryThis paper compares two ways of generating synthetic training data for object detection: Unity simulation rendering and controllable diffusion generation (CIA), in settings where real data is scarce. Using a unified framework that mixes real, simulated, and generated data with identical model settings, the authors evaluate results with Precision, Recall, mAP, and custom Δ-metrics. Neither synthetic source alone performs best, but adding a small share of synthetic data improves generalization, while replacing too much real data causes domain drift.
Mohamed Benkedadra, Aissa Saoudi, Maxime Gloesener et al.
In briefHybridCUA: Teaching Computer-Use Agents to Combine GUI and Command-Line Actions
SummaryThis paper proposes HybridCUA, a method that lets computer-use agents combine graphical interface actions with command-line interface (CLI) commands. The authors build HybridCUA-8K, a dataset of GUI-only, CLI-only, and mixed trajectories, and train with supervised fine-tuning followed by reinforcement learning using CLI-aware rewards. They report that HybridCUA-9B reaches 53.6% accuracy on OSWorld, 14.8 points above its base model, and gains 4.0 points on WindowsAgentArena.
In briefHARISSA: Self-Checks for Efficient, Safe Local Language Models
SummaryThis paper proposes HARISSA, which uses a language model's own hidden states to decide whether a hard query needs extra computation or should be deferred to a human, rather than sent to a cloud model. The states before and after answer generation are trained to predict correctness, and a cascade tries cheaper answering methods first, deferring when the final answer is predicted wrong. On a single-model device, HARISSA reports accuracy within one point of chain-of-thought at 2.7 times lower latency.
In briefDiagnosing and improving LLMs' probabilistic reasoning for decisions
SummaryThe paper proposes a decision-theoretic framework that splits an LLM's decision loss into belief formation from evidence and translating beliefs into actions under a given utility function. Using a synthetic benchmark with known ground truth, the authors characterize probabilistic reasoning in frontier and open-source models. They find that reinforcement learning targeting one component redistributes decision loss, while jointly targeting belief and decision improves both, but only when training and evaluation formats match.
In briefNo Scale Left Behind: A Bidirectional-Attention Multi-Scale Autoencoder for Time Series Anomaly Detection
SummaryThis paper addresses time series anomaly detection, where anomalies span very different temporal scales. It proposes MSCAD, a semi-supervised framework with parallel autoencoder branches for different patch sizes, linked by symmetric bidirectional cross-scale attention blocks that let scales exchange information before reconstruction. On the TSB-AD benchmark (40 datasets, 530 series), the authors report VUS-PR scores of 0.57 (univariate) and 0.47 (multivariate), gains of 9.6% and 9.3% over the state of the art.
In briefChernoff bottleneck: compressing data while keeping error-exponent for binary decisions
SummaryThis paper studies an encoder that maximizes the Chernoff information between two conditional distributions, which governs the error exponent of decisions from many samples, subject to a rate constraint I(U;X) ≤ R. The optimal value strictly increases up to R = H(V) and then stays flat, and the authors give an alternating algorithm with proven monotone convergence. For large numbers of samples it yields smaller error probability than the classical information bottleneck. On 20 Newsgroups topic detection, compressing each word to 17% of its entropy retains 90% of the error exponent.
WhyReplaces the IB objective with Chernoff information to directly characterize decision error exponents, giving a structural optimum, convergent algorithm, and proven large-sample advantage over IB.
In briefBITEM at NTCIR-19 R2C2: Predicting Answer Confidence from Agentic RAG Pipeline Signals
SummaryThe BITEM team entered both NTCIR-19 R2C2 subtasks with one agentic pipeline that searches and records evidence over a movie corpus, with an orchestrator releasing only checked claims. Confidence is computed by the orchestrator from the run's recorded signals rather than asked of the model. Hand-written rules over these signals raised accuracy to 0.9375 and HMR to 0.6985, and the authors propose accHMR, accuracy multiplied by HMR, as a metric that rewards accuracy as well as calibrated confidence.
Julien Knafou, Luc Mottin, Alexandre Flament et al.
In briefAREX-2: Long-horizon reflective training for self-improving AI agents
SummaryThis paper presents AREX-2, a method for improving LLM agents' ability to repeatedly refine their solutions during testing. The approach combines reflection with long-horizon execution and trains on long improvement trajectories synthesized from machine learning and algorithmic programming tasks. The agent, built on Qwen3.8-27B, reports strong results on several benchmarks and continues to improve as the number of rounds increases. The authors conclude that long-horizon reflective data is an effective route toward self-improving agents.
In briefWhich attention heads resemble the human brain? Not the ones doing the computing
SummaryThe paper compares LLM attention-head representations with human EEG on an abstract pattern-completion task and tests how ablating those heads affects performance. Brain-aligned heads contribute to task performance, but removing them is substantially less disruptive than removing heads selected via attribution patching. The authors conclude that brain alignment captures how the model reads the stimulus rather than how it solves the task.
Christopher Pinier, Gustaw Opie{\l}ka, Hannes Rosenbusch et al.
In briefTabFM-Auto: Self-Evolving Data Pipelines for Tabular Foundation Models
SummaryThe paper presents TabFM-Auto, which pairs the tabular foundation model TabFM with a language model agent that iteratively refines the surrounding data pipeline, including cleaning, feature engineering, context selection, and post-processing, to reduce error. On all 51 TabArena datasets, five TabFM-Auto configurations take the top five overall positions, and the best raises TabFM's Elo from 1785 to 2013. The discovered pipelines also transfer to other frozen tabular foundation models without further search, and on MLE-Bench's 8 tabular competitions the approach ranks first overall among MLE agents.
In briefKV-Kaizen: Learning Context-Adaptive Cache Compression for LLMs
SummaryAs input context grows, the key-value cache of a large language model can exceed the memory used by its weights, slowing decoding. This paper introduces KV-Kaizen, which learns a selector that chooses a per-layer cache configuration based on context, using layer sharing, lower precision, and low-rank truncation instead of evicting tokens. On instruction-following and reasoning tasks, the selectors reach the accuracy-versus-cache-size Pareto frontier. On a 14B model, the decode-time cache is reduced 32x while preserving accuracy.
Joao Monteiro, Louis B\'ethune, Anastasiia Filippova et al.
In briefORMA: Optimization-Based 4D Reconstruction of Articulated Animals from Monocular Video
SummaryThe paper presents ORMA, a training-free framework that reconstructs articulated 4D models of quadruped animals from monocular videos. It uses predicted poses as references for optimization and generative 3D priors for accurate shape, registering the result to the SMAL+ parametric model. The authors also introduce PAW4D, a synthetic multi-species benchmark with ground-truth geometry and camera motion, and report improved accuracy and globally consistent motion on PAW4D, PFERD, and in-the-wild videos.
In briefS³: Spectral Null-Space Swap Makes Reasoning Models More Efficient
SummaryReasoning models trained with chain-of-thought are accurate but costly in tokens. The paper finds that much of the reasoning capacity lies in the part of the Thinking model's weights that falls in the null space of the Non-thinking model's dominant singular directions, and that removing the dominant-subspace component can improve efficiency without hurting accuracy. It proposes Spectral Null-Space Swap (S³), a training-free method that composes paired Non-thinking and Thinking checkpoints. Across 2B–30B dense and MoE models on 28 evaluation environments, S³ reduces inference token overhead by 27.4% on average while improving overall accuracy by 1.0 percentage point.
WhyFirst to exploit projection null spaces for training-free model composition, cutting tokens and raising accuracy across 28 environments and 2B-30B models.
In briefSocial choice foundations for efficient simulation-based answers to contested questions
SummaryThis paper studies simulation-augmented generation (SAGE), where models simulate individuals' viewpoints at inference time to give more representative answers to contentious queries. The authors formalize the efficiency problem using mPJR+, a proportionality axiom from proportional clustering, and prove that simulating a small number of individuals and routing to a still smaller subset per prompt retains approximate proportional representation for the whole population. In experiments on political questions and personal advice, the proposed routing algorithm achieves higher mPJR+ satisfaction rates than k-means-based and random selection baselines.
Sonja Kraiczy, Smitha Milli, Ratip Emin Berker et al.
In briefTraining masked diffusion language models with trajectory awareness
SummaryMasked diffusion models generate text by unmasking several tokens per step, but training and inference run under different conditions. The paper introduces PUMBA, which trains the denoiser on consecutive steps of model-induced trajectories, passes information between steps, and optimizes them jointly via backpropagation through time. When fine-tuning LLaDA-8B, PUMBA reaches matched performance with up to 22% fewer function evaluations in full-canvas generation and up to 26% fewer in block diffusion generation.
Manuel Madeira, Amitis Shidani, Alice Bizeul et al.
In briefPhysWAM: a physically consistent world-action model for self-driving cars
SummaryThis paper presents PhysWAM, a unified model for autonomous driving that jointly generates multiview video, metric depth, and ego motion within one flow-matching transformer. It introduces Coupled Point Projection (CPP), which turns generated depth into 3D points, transforms them with the generated ego motion, and pulls them toward LiDAR points transformed by the recorded ego motion, tying predictions to measured scene geometry. The authors evaluate it on NAVSIM v1 and v2 planning, zero-shot closed-loop transfer, and future video and depth prediction, reporting strong planning, zero-shot transfer, accurate depth, and gains from CPP.
In briefHistorical Anchors for Scientific Forecasting on Human-Aware Knowledge Graphs
SummaryThis paper studies how to predict which scientific connections, such as material properties, will form next, using embeddings trained on paths sampled from a publication hypergraph. It introduces Query–Anchor Agreement (QAA), a diagnostic that needs no future labels to choose the sampling budget and refine forecasts. Experiments on thermoelectric-material and COVID-19 drug-repurposing benchmarks show QAA-selected budgets come close to hindsight-optimal selection.
Siu Tin Wu, Ryan C. Barron, Phoebe Upton Sanderbeck et al.
In briefSoL-Refiner: One-Step Speed-of-Light Refinement for High-Resolution Video
SummaryHigh-resolution video generation is costly, so a common approach first creates a low-resolution video and then refines it. This paper presents SoL-Refiner, which converts low-resolution outputs into 4K video in a single denoising step, trained through high-resolution continual training, reinforcement learning, and one-step distillation. The authors also introduce Refiner-Bench, a benchmark built from outputs of different video generators. They report better VBench and UniPercept averages than external refiners at about 2K, and an 8.91x speedup in refinement latency over a three-step baseline.
In briefTabFM: A Zero-Shot Foundation Model for Tabular Data
SummaryThis paper presents TabFM, a 400M-parameter foundation model for tabular data that frames supervised prediction as in-context learning. It makes calibrated zero-shot predictions in a single forward pass without task-specific tuning. The model is trained only on synthetic tables generated from structural causal models. Across 51 TabArena benchmark datasets, zero-shot TabFM ranks first among default tabular foundation models and outperforms tuned AutoML pipelines.
WhyFrames tabular prediction as in-context learning, trained on synthetic SCM data, beating tuned AutoML across 51 TabArena datasets.
In briefKolmogorov-Arnold Classifier Systems: Rule-Based Learning as Universal Approximators
SummaryThe paper notes that Learning Classifier Systems suffer exponential growth in rules and parameters as input dimension rises. It proposes KACS, which uses the Kolmogorov-Arnold representation theorem to split the target function into one-dimensional subproblems, cutting worst-case rule count from O(m^n) to O(mn^2) with two parameters per rule. The authors give a constructive proof that KACS is a universal approximator on compact domains, and report competitive accuracy in many settings using 2% to 40% of the parameters of a direct partitioning approach.
In briefBrainNet Studio: a toolkit for building, analyzing, and visualizing brain networks
SummaryBrainNet Studio is an integrated toolkit for analyzing static and dynamic brain networks, covering network construction, feature extraction, predictive modeling, biomarker candidate identification, visualization, and interpretation support. It integrates 27 algorithms, including deep learning, graph neural networks, and spatiotemporal sequence models. A large language model generates summaries of functional connectivity, structural connectivity, and structure-function coupling that researchers can verify. The toolkit is publicly available on GitHub.
In briefTopological Coherence for Self-evolving Multi-agent Systems: a method keeping agent roles, handoffs and memory access aligned with task structure
SummaryThe paper argues that responsibility, handoff and memory boundaries in multi-agent systems should stay consistent with task dependencies, which it calls topological coherence. It introduces TOCOMAS, which ties a task graph to tool interfaces and jointly adapts agent, collaboration and memory policies during online self-evolution. Candidate changes are kept only if they satisfy structural constraints and improve evaluated reward. The authors report higher task success across BBEH, WorkBench, SWE-Bench-Verified and CoMemBench, with gains on CoMemBench in verified progress, handoffs and memory isolation over a self-evolving baseline.
In briefGaugeVLM: Teaching vision-language models spatial consistency with measured 3D interventions
SummaryGaugeVLM generates linked observations by making controlled object and camera changes in explicit 3D scenes, so spatial differences are measured. Its core objective, GaugeDPO, turns measured errors into preference margins to supervise consistent rankings across views. Across three VLM backbones, it improves all 10 established spatial metrics over supervised fine-tuning, with the 7B model gaining 15.0 and 18.9 points on MSMU distance and QSpatial+. The authors report gains extending to autonomous driving and embodied reasoning.
In briefLighting-consistent method for inserting advertising graphics into scenes
SummaryThis paper presents Ad-Relight, an inference-only method that transfers scene lighting onto a supplied advertising graphic without a banner-specific training set. It separates slowly varying shading from graphic structure and uses a pretrained diffusion relighter, probed with two nearly identical backgrounds, to isolate the target region's contribution. Across 560 generated placements, it improves structural similarity, perceptual distance, and illumination agreement over geometric compositing and direct relighting baselines. The study is image-based, and temporal stabilization is left as an open extension.
Rameshwar Mishra, Bishshoy Das, A. V. Subramanyam et al.
In briefVideo-RSI: Self-improving video agents by evolving their execution harness
SummaryThis paper introduces Video-RSI, a framework in which a video understanding agent uses its own language model to revise its execution harness, the program controlling what it observes and how. The model revisits training videos to test competing explanations for failures with additional observations, and a cost-aware process keeps revisions based on both accuracy and visual cost. On the evaluated video benchmarks, the evolved agent improves accuracy while processing fewer frames and reaches competitive accuracy-efficiency trade-offs against existing agents.
In briefProvable identifiability of drivers and dynamics in delayed physical systems
SummaryThe paper proposes a theory-grounded method for learning delayed physical systems. It proves that, under a set of permissive assumptions, the structural drivers and drift of stochastic delayed differential equations are identifiable. The method is reported to outperform other approaches on a driver identifiability benchmark and on a benchmark assessing the physical consistency of learned dynamics.
Julien Boussard, Antoine Debouchage, Th\'{e}o Saulus
In briefAn Efficient Machine Learning Approach for Predicting Degradation in AEM Water Electrolysis
SummaryThis study analyzes a new dataset of single-cell anion exchange membrane water electrolyzers (AEMWE) run at constant current across several heterogeneous experimental campaigns. Machine learning models of varying complexity, including linear baselines, LSTMs and CNNs, are trained to forecast the cell voltage degradation curve over the medium term. The models are evaluated within a framework designed for heterogeneous industrial data.
Marco Veneriano, Ani Gjergji, Sebastiano Bellani et al.
In briefDoes Local Video Understanding Transfer Across Encounters? The EgoGears Benchmark
SummaryThis paper introduces EgoGears, a benchmark that tests whether multimodal models can carry video understanding from one recording to another. It contains 567 single-video and 1,487 multi-video questions built from 126 egocentric recordings across 39 outdoor routes. Across 29 single-video and 31 multi-video model configurations, performance drops by 22.5 percentage points on average for multi-video questions, and the main bottlenecks are binding observations to evidence and tracking ordered route states.
WhyProposes a diagnostic benchmark separating single- and multi-video failures, backed by 126 recordings and dozens of model configurations.
In brief"Look Closer": Training AI-image detectors on small image patches
SummaryThis paper asks whether small patches of an image can reveal that it was AI-generated. The method trains a shared backbone on explicit crops, applying a separate loss to each patch and averaging patch probabilities only at inference, without handcrafted filters or a fusion module. On GenImage, the patch-wise variants improved average accuracy over whole-image versions across all four backbones tested. The authors note that some historical experiments used evaluation-based model selection, so their scores are not a uniformly selected leaderboard comparison.
In briefStatistically valid anomaly inference for Deep SVDD (PADI)
SummaryThis paper proposes PADI, a framework that adds statistically valid inference to a trained, frozen Deep SVDD anomaly detector. It uses Selective Inference, conditioning on the event that a test instance is flagged as anomalous, to derive valid selective p-values. The authors theoretically show that false positive rate can be controlled at a user-specified significance level α. The framework is also extended to Deep SAD for semi-supervised settings, and experiments support the theoretical findings.
Cao Le Cong Thanh, Dang Quang Vinh, Vo Nguyen Le Duy
In briefGRFBrain: graph-structured flow models for forecasting EEG brain connectivity
SummaryThis paper studies forecasting how functional connectivity between EEG channels changes over time. The authors propose a framework that separates a history-based prediction of the future connectivity graph from stochastic residual transport. A graph-structured Gaussian source, built from past connectivity, is transported to future residuals by a conditional velocity field. The study also identifies the conditions and controls needed to isolate the benefit of residual transport from that of deterministic prediction, learned representations, and sampling effects.
Haohui Jia, Zheng Chen, Jathurshan Pradeepkumar et al.
In briefLearning When to Update: A Near-Optimal Timing Bandit Approach
SummaryThis paper studies when resource-intensive systems, such as machine learning models and digital twins, should be updated in changing environments. It formalizes a "timing bandit" problem in which each candidate update interval has a fixed update cost and an unknown, stochastic degradation cost. The proposed Balanced Consecutive Arm Elimination (BCAE) algorithm achieves $\tilde{O}(\sqrt{T})$ regret, improving on standard bandit methods, and an optimism-enhanced variant keeps the same regret order while improving empirical adaptivity. Simulations show low and stable regret as the number of arms and update cost vary.
In briefLearning what a language model should remember over long interactions
SummaryThe paper introduces Memory Gain Policy Optimization (MGPO), which credits each memory rewrite for its marginal contribution to current and future downstream utility. Evaluated on document-level information extraction, MGPO improves extraction while reducing average memory length by nearly 80% relative to the initial memory policy. The learned memory policy also transfers across domains and to downstream models without further training.
In briefRollout-Marginal Distillation: a training method that keeps long autoregressive video quality high
SummaryThis paper addresses how autoregressive video diffusion models accumulate prediction errors over long generated sequences. It proposes Rollout-Marginal Distillation (RMD), which keeps the generated history for prediction but scores each chunk independently against a chunk teacher, so context artifacts do not distort the quality correction. Video-level distribution matching distillation is then applied to restore temporal coherence. The authors report that RMD keeps visual quality high beyond its training horizon and outperforms video-level DMD baselines.
In briefTime-Anchored Diffusion Language Models: Caching Latent States for Faster Generation
SummaryThis paper proposes time-based self-supervised anchoring, which learns and reuses latent anchors in diffusion language models without requiring supervised targets. It uses a two-stage design in which an anchor network periodically produces a cached latent state, and a lightweight denoising network reuses it across reverse steps through a fusion module. The authors report that the post-training variant improves throughput by roughly 49% to 79% on math, code, and STEM benchmarks, and the pretraining variant reduces Transformer-layer computation by up to 38%.
In briefEpiCon: Shared multimodal memory lets AI agents learn from each other's experience
SummaryThis paper introduces EpiCon, a shared multimodal memory framework that lets multiple AI agents reuse and build on each other's experience without updating the host models' parameters. It uses two independently trained 2B models: a memory controller that refines textual guidance and visual evidence across attempts, and a tree self-organizer that consolidates lessons and retrieves experience and rules for new problems. Evaluated on eleven benchmarks across four multimodal task domains, EpiCon improves macro-average scores by 1.7 to 4.9 points over No Memory and cuts memory-operation time by 67% to 74%.
In briefMaking OCR more faithful by adaptively weakening teacher guidance during training
SummaryVision-language models can rewrite unusual text in images into plausible-sounding wording, reducing OCR transcription faithfulness. The paper proposes GAD-RL, which adaptively regulates supervision from a frozen teacher model according to the student's current task reward and local token probabilities. On Qwen3.5-2B, GAD-RL reaches 59.92% Micro Recall on CHAOS-Bench and an Overall score of 91.18 on OmniDocBench v1.6.
In briefPattern Formation in Transformers: How Architecture Shapes Token Representations
SummaryThis paper uses pattern-formation theory to explain how a full Transformer organizes token representations, covering positional encoding, multi-head attention, and output-value geometry. The authors report that the architecture selectively amplifies a set of previously unreported patterns, including traveling and rotating waves, and characterize which components determine which patterns are amplified, stabilized, or coexist. They further show that choosing task-aligned positional encodings and weight initialization yields a controllable dynamical prior that improves data efficiency and optimization on controlled sequence tasks and on CIFAR-10 with ConViT.
In briefSYNCR: A simulator-based framework for diagnosing and training cross-video reasoning
SummarySYNCR is a simulation-based framework built on Habitat, Kubric, and CLEVRER that provides 4,000 evaluation and 15,960 training questions across eight cross-video reasoning tasks. Evaluation of 22 multimodal large language models shows persistent difficulties in physical comparison and scene integration that larger models do not consistently resolve. Supervised fine-tuning raised Qwen3-VL-8B's average SYNCR accuracy from 32.6% to 61.6%, and transfer to real footage improved temporal ordering accuracy by 9.0 to 20.5 percentage points.
Sara Ghazanfari, Siddharth Garg, Prashanth Krishnamurthy et al.
In briefSearch Dimension in Unlabeled Projection Pursuit: How Restricting the Search Space Scales
SummaryThis paper studies how the search dimension affects unlabeled projection pursuit, a method that looks for the direction in which data appear least Gaussian. When the observation space contains a large Gaussian component, the search can settle on a direction carrying no signal. The authors compare restricting the search to the column space of a known forward operator with estimating a principal subspace, and report an empirical scaling law that the paper describes as not proven to be asymptotically tight.
In briefRLX: A Rust Compiler and Runtime That Runs Machine Learning Models on Many Hardware Backends
SummaryThis paper presents RLX, a Rust codebase that combines a tensor compiler and a distributed runtime around a single three-level intermediate representation. It supports about fourteen runtime devices, several model file formats, and quantized arithmetic, and it can run across machines. The authors benchmark RLX against many existing frameworks on one host, reporting that it is fastest on some tasks, such as a MiniLM model on Metal, while keeping top-1 parity on reference checks.
In briefOvercoming Scaling Limits in On-Policy Self-Distillation for LLM Reasoning
SummaryThis paper studies on-policy self-distillation (OPSD) for improving reasoning in large language models. It finds that the correctness of the scaffold (reference solution) affects downstream accuracy more than the correctness of the teacher's context, and that unverified scaffolds create an imitation gap. The authors propose OASIS, which needs only final-answer labels and supervises on verified on-policy trajectories. On Qwen3-1.7B, 4B, and 8B across AIME and HMMT benchmarks, OASIS improves over the base model, and at 8B it outperforms OPSD by 3.05 points.
In briefUnequal Impact of Harmful Training Examples: Using Data Attribution to Tune Emergent Misalignment
SummaryThis paper studies whether every harmful example in a narrow fine-tuning dataset contributes equally to emergent misalignment (EM) in large language models. The authors use training data attribution to estimate each example's contribution and validate these scores by retraining on filtered data. Filtering by attribution scores or a black-box harmfulness score can substantially strengthen or weaken EM, with same-model filtering performing best.
Gon\c{c}alo Paulo, Louis Jaburi, Nora Belrose et al.
In briefGenerated query expansion still helps strong learned sparse retrieval
SummaryThis paper examines whether generated text can expand short scientific queries to bridge vocabulary mismatch with relevant papers. The authors compare four generated formats using SPLADE-v3 on NFCorpus, TREC-COVID, and SciDocs, holding the document index and token budget fixed. All methods improve aggregate nDCG@10, and eleven remain significant after Holm correction, with gains persisting when the original query keeps substantial weight. A corpus-induced typed concept graph, however, yields no consistent improvement.
Ryan C. Barron, Cade W. Trotter, Maksim E. Eren et al.
In briefFrom Pixels to Keys: Studying Spatial and Motion Cues for Inferring Game Inputs
SummaryThis paper studies Inverse Dynamics Models (IDMs) that infer player inputs from gameplay video frames, so that abundant unlabeled gameplay videos can be used for training. Under limited data, the authors examine how spatial motion features, model architectures, and training objectives affect performance, using per-key and balanced metrics such as macro F1. Experiments on Trackmania and Cyberpunk 2077 show that architecture and motion flow extraction matter, that unbalanced metrics can be misleading, and that camera motion, delayed effects, and imbalanced key-press frequencies cause ambiguities.
Abhishek Pillai, Ekta Prashnani, Joohwan Kim et al.
In briefBeyond Interaction Capacity: Recursive Estimator Scaling (RECAP) for CTR Prediction
SummaryThis paper examines click-through rate prediction in recommendation and advertising, noting that simply deepening cross networks to raise interaction capacity yields diminishing returns. It proposes estimator scaling, which uses multiple related estimators and their non-shared predictive variation, and introduces RECAP, a parameter-efficient recursive model that applies this idea through distillation, exponential moving averaging, and inference-route aggregation. The authors report new state-of-the-art predictive performance across multiple benchmarks along with a favorable performance-parameter trade-off.
Shivang Chopra, Fotis Iliopoulos, Zsolt Kira et al.
In briefChoosing an LLM Provider Can't Be Done From the Price List Alone
SummaryThe paper argues that for open-weight LLM inference, after choosing a model a client must also choose which provider serves it, and providers of the same model differ widely in quality, latency, availability, and price, with price not reliably predicting quality or availability. The authors propose a measured-map policy that routes to the cheapest healthy, quality-equivalent provider, and FACET, an online router that certifies per-task provider feasibility and falls back to an anchor when uncertified. Live tests show certification can shift traffic from a premium provider to a substantially cheaper certified one.
WhyIdentifies provider selection as a new routing axis, backed by live measurements and online routing experiments showing real cost savings and degradation avoidance.
In briefScaling zero-order pretraining by splitting models into independently trained shards
SummaryZero-order optimization trains models without backpropagation, but its gradient variance grows with model size, limiting large-model training. The paper proposes SOMA, which trains LSTM experts independently on separate data clusters using SPSA, exchanging no gradients, activations, or optimizer state. Experiments report that modest sharding improves compute efficiency over monolithic zero-order baselines, and the authors prove independent losses reduce relative gradient variance to about 1/N of a shared-loss estimator's.
Francois Chaubard, Mykel J. Kochenderfer, Chris R\'e
In briefGuide, Then Let Go: Adaptive Teacher Scheduling for Sparse-Reward AI Agent Training
SummaryThis paper addresses the cold-start problem in reinforcement learning for long-horizon AI agents, where sparse outcome rewards give little signal early in training. It uses on-policy distillation to provide token-level guidance from a teacher model on the student's own rollouts. The proposed GATS method adjusts the weight of this guidance according to the teacher-student performance gap, reducing and then withdrawing it once the student reaches the teacher's reference performance. Across ALFWorld, WebShop, and ScienceWorld, GATS reports the highest average success rate, improving over reward-only GRPO by 4.37% to 11.87%..
In briefIt's All Training: A Fully Synthetic Single-Stage Recipe for LLMs
SummaryThe paper introduces SYNTH, the first open-source synthetic corpus, built from 58,698 Wikipedia articles and merging pre-, mid-, and post-training into one stage through structured amplification of curated encyclopedic seeds. The authors trained models ranging from a 56M model to a 13B (1B-active) MoE to evaluate it. At equal compute, SYNTH outperformed filtered web data, and the models used 10–140 times fewer training tokens while achieving high factual precision. The SYNTH dataset and Baguettotron models are publicly released under a permissive license.
Pierre-Carl Langlais, Pieter Delobelle, Yannick Detrois et al.
In briefReCAP: Retrieval-guided reuse of skills for multimodal continual instruction tuning
SummaryThis paper proposes ReCAP, a framework that helps multimodal large language models learn new tasks in sequence while reusing previously learned abilities. At each stage, it uses external search and an LLM to build a knowledge base of domain, reasoning, and format knowledge, then selects and orders capability modules for each instruction. An adaptive subspace recycling method keeps reuse stable across stages by protecting important directions and recycling residual capacity. The authors report state-of-the-art results on MCIT benchmarks.
In briefVisual-Only Class-Incremental Learning for CLIP: Removing the Text Branch
SummaryThis paper studies class-incremental learning on CLIP, where a model must learn new classes over time without forgetting old ones. The authors argue that textual classifier weights from CLIP's text encoder can deviate from visual class distributions due to a modality gap. They propose VIS, which removes the text branch and builds the classifier entirely in visual space using a kernelized incremental least-squares SVM. The authors report state-of-the-art results without a textual branch.
In briefBehavioral Capacity Certificates: Measuring Quantized Language Model Complexity by Behavior, Not Weight Codes
SummaryThis paper introduces Behavioral Capacity Certificates (BCC), which measure the complexity of quantized language models by the aggregate prior mass of complete implementations that induce the same bounded loss, rather than by weight codes alone. It describes a three-step deployment workflow: forward-only screening to choose per-layer bit-widths, margin-certified identification of weights that can be pruned or sign-flipped without changing behavior, and bounding the population loss of the deployed model. Experiments show that, at equal cache memory, giving keys higher precision than values yields lower NLL and higher prediction agreement.
In briefBoids of a Feather Flock Together: how prey behaviour evolves against different predator attack strategies
SummaryThis paper extends a predator-prey boids model by combining six prey movement tendencies into one weighted acceleration update and reformulating wiggle as a sinusoidal manoeuvre. An evolutionary strategy optimises the six behaviour coefficients for collective prey survival against four predator hunting strategies. The authors find that prey survival was highest against attack-centroid and lowest against attack-nearest, and suggest optimal anti-predator behaviour depends on the interplay between escape tendencies and the predator's strategy.
In briefTopoEmbedX: a unified framework for embedding topological domains
SummaryThis paper presents TopoEmbedX, a framework for embedding topological structures such as simplicial complexes, hypergraphs, and cell complexes into Euclidean spaces. It combines five existing algorithms and introduces five new ones that extend graph embedding methods to higher-order settings using the augmented Hasse graph. Experiments show the generated embeddings support classification and regression on multidimensional data.
Florian Frantzen, Ibrahem AlJabea, Ines Henriques-Cadby et al.
In briefZero-shot dependency parsing via unsupervised cross-lingual bootstrapping
SummaryThis paper examines how encoder-based pre-trained language models can perform dependency parsing in languages without training data, a task that is hard to transfer zero-shot because it depends on syntax. The authors propose an unsupervised cross-lingual bootstrapping method to strengthen syntactic knowledge inside the model and improve generalization across linguistic typologies. The method significantly improves zero-shot parsing in low-resource languages. Parameter-free tree probing tests also yield higher scores, which the authors read as greater robustness in recognizing syntactic structures.
Lalita Lowphansirikul, Attapol Rutherford, Jian Gang Ngui et al.
In briefHow Many Labels Does a Language Need? Annotation Budgets and Cross-Lingual Pooling for African-Language Text Classification
SummaryThis paper studies how much labelled data African-language text classifiers need and whether data from other African languages can help. Using a lightweight character n-gram model, the authors test 28 language-task pairs covering news topic classification and tweet sentiment analysis. They find that topic classification needs about 400 labels in the median language, while sentiment often needs thousands, and that pooling other languages' data helps only when labels are scarce.
In briefHow Many Context Tokens Self-Attention Really Needs Under Competition
SummaryThe paper asks how many context tokens a language model actually uses by keeping only the highest-attention tokens at each head, layer, and query, without retraining. It measures the increase in negative log-likelihood as the selected set size varies, estimating the effective attention set size needed within a chosen loss tolerance. Attention-based selection substantially outperforms random selection, although the required set size varies across models.
In briefCo-PiLOT: Constrained Physics-Informed Latent Optimization for Target-Driven Inverse Design
SummaryThis paper presents Co-PiLOT, a method for inverse design that searches a latent space learned by a generative encoder-decoder, using the decoder as a validity prior. It is applied to designing magnesium alloy microstructure and texture, trained on about 80,000 EBSD-derived microstructures. Within a budget of 160 simulations, the ViT-FMDiT and MERIDIAN combination reduces relative target error by 3 to 22 percent compared with seven baselines.
In briefEndoPrior-GS: A Method for Dynamic 3D Endoscopic Reconstruction Using a Texture Prior
SummaryThis paper presents EndoPrior-GS, a pipeline for reconstructing dynamic 3D scenes from endoscopic video during surgery. It builds on 3D Gaussian Splatting and uses a joint texture prior, combining a tool-filtered tissue mask, a non-specular photometric filter, and anatomical structure salience, to guide primitive initialisation and density control. Tested on EndoNeRF and SCARED, the method reduces Flow Error by 27.7% and 25.8% over representative approaches while keeping competitive rendering quality and real-time speed.
In briefExceptionDrive: A Counterfactual Benchmark for Testing Autonomous Driving Planners on Rare Hazards
SummaryExceptionDrive is a benchmark that inserts rare, safety-critical hazards into real nuScenes driving scenes while keeping their original context. It contains 21 tasks across six safety families and uses reference-free metrics to judge planner outputs. Seven representative planners often intruded into hazard regions or kept insufficient clearance. The authors also propose a Reminder Agent that turns visual evidence into structured hazard records, which improved strategy accuracy and reduced under-warning in zero-shot tests.
In briefRemoving raindrops from a single image using learned synthetic photorealistic data
SummaryRaindrops on camera lenses or windshields can disrupt vision systems such as those used in autonomous driving, and realistic training data for them is scarce. The paper introduces the first photorealistic dataset of synthetic adherent raindrops, rendered with physics-based modeling of water dynamics, geometry, and photometry, covering various rainy scenes including driving. It proposes a detection network aware of raindrop refraction and blurring, followed by a removal network that recovers image structure. The authors report state-of-the-art performance in rigorous experiments.
In briefMasked Swingers: Improving Masked Autoencoders for Self-Supervised Learning with Data Augmentation
SummaryThis paper introduces Masked Swingers, a modification of masked autoencoders (MAE) that augments each image in two ways, masks and encodes each view separately, then swaps the global CLS representations between views before decoding the masked patches. This encourages a view-agnostic summary of the image for more efficient transfer. The authors report gains of 3 to 5 percent over MAE on ImageNet-1K kNN, large relative gains on fine-grained tasks, and a 64 percent relative error reduction on three new state-probing datasets.
Anthony Fuller, Scott C. Lowe, Daniel G. Kyrollos et al.
In briefLearning From Peer Models: Off-Policy-Aware Cross-Model Trajectory Exchange for RLVR (GRAFT)
SummaryThis paper addresses a limitation of reinforcement learning with verifiable rewards such as GRPO, where finite rollouts can yield all-failed groups that provide no reward signal. The authors propose GRAFT, which replaces these groups with informative trajectories from a peer model, using sequence-level compatibility weighting and token-level importance ratio clipping to manage cross-model mismatch. On three heterogeneous model pairs and five mathematical reasoning benchmarks, GRAFT improves both models over GRPO under the same rollout budget.
In briefOptimization landscape of low-tubal-rank tensor sensing and how local geometry depends on Fourier-slice ranks
SummaryThis paper studies the optimization landscape of low-tubal-rank tensor sensing using a balanced factorization. Under a tubal restricted isometry condition, it establishes a quantitative strict-saddle landscape with no spurious local minima for arbitrary Fourier multi-rank profiles. The authors show that local geometry depends on the Fourier-slice ranks rather than the tubal rank alone, and numerical experiments illustrate global optimization behavior and contrasting local geometries.
Eugene Agyei-Kodie, Longxiu Huang, Shuang Li et al.
In briefAgentBug-Smith: Automatically reproducing real-world harness bugs in AI agent systems
SummaryThe paper presents AgentBug-Smith, an automated method that finds and reproduces harness bugs from open-source agentic systems, and uses it to build Live-Harness-Bench, a extensible benchmark with 200 reproducible bugs. It reports 10.67% to 27.56% higher reproduction success rates than existing general-purpose bug reproduction techniques across different backbone LLMs. The authors show that the benchmark reveals the limited repair abilities of current software agents and that repair skills distilled from its fixes raise repair rates by 6.32%.
In briefIt's Not What the Image Shows: Irrelevant Context Destabilises VLM Judges Without Informing Them
SummaryThe paper introduces MIST, a stress test of 200 English sentences paired with an aligned image, a misleading image, or no image. Across thirteen vision-language model judges, any image, aligned or misleading, changed labels at similar rates, and the effect appeared in all judges. The authors conclude that whether a model can substitute for human annotators depends on the evaluation configuration as well as the model itself.
Nagham Omar, Mahmoud Jabarin, Kinan Ibraheem et al.
In briefOne Threshold Fails for Some Languages: Per-Language Deferral for Low-Resource Text Classification
SummaryThe paper examines whether the error-rate guarantee from split conformal prediction holds for every language in low-resource text classification. A single pooled threshold meets the 90% target on average but undercovers several languages, including Somali and Tigrinya. Setting one threshold per language brings each language to roughly 89% to 91% coverage, but the share of cases sent to human review varies widely across languages.
In briefStorage Is Not Strategy: Choosing Which LLM Parameters to Edit for Unlearning
SummaryThis paper examines how localized unlearning methods for large language models choose which parameters to update. Many methods pick a small parameter subset based on a localization signal and keep it fixed during optimization. The authors propose Intervention Score, which ranks editable groups by the predicted effect of the actual unlearning update, and a selective dynamic re-ranking method (DIR-R). The reported results show positive margins in most comparisons, though some are near zero, and the gradient-difference objective shows substantial field dependence.
In briefActive budget can reduce sensitivity in TopK sparse autoencoders
SummarySparse autoencoders are often scaled to wider dictionaries to extract fine-grained structure from language model activations. The paper examines whether features remain stable when the same meaning appears in different surface forms. It finds that scaling reduces the sensitivity of rare features while common features stay comparatively stable, and attributes the cause to the active budget k and the selection boundary. The authors propose pairwise rank stabilization, which targets ordering failures at the cutoff to improve rare-feature sensitivity.
In briefHandAnthro: Measuring Hand Size Automatically From One Phone Photo
SummaryThe paper presents HandAnthro, a method that estimates 44 projected hand dimensions from a smartphone photo of a palm-up hand placed on US letter-size paper, aiming to support protective-glove design. The pipeline rectifies the paper, whitens non-hand pixels, and locates 41 landmarks using a fine-tuned YOLO pose model. In controlled evaluation, mean absolute error against two trained operators' caliper measurements was 3.80 mm per dimension.
In briefAnthroDial: Benchmarking How Human-Like LLMs Behave in Autonomous Social Conversation
SummaryThis paper introduces AnthroDial, a framework for building and evaluating human-like social AI agents that decide for themselves whether, when, and how to communicate. It combines MindFlow, an interaction system using a dynamic Mind Buffer; CAPS-Eval, an evaluation method covering cognitive, affective, and behavioral dimensions; and a training approach pairing SEEDS with DiAPO. The authors also build evaluation datasets for everyday, game, and long-horizon character interaction. Reported experiments show gains in interaction autonomy and naturalness, and CAPS-Eval agrees with human rankings.
In briefDelta-Matching: A Method to Close the Final Gap in Native 8-bit LLM Training
SummaryThe paper shows that inconsistencies between forward and backward passes in FP8 attention produce stale deltas that distort training, with degradation becoming substantial at 1.67B and 5.29B parameters. It proposes Delta-Matching, which the authors prove restores the softmax gradient's zero-row-sum invariant under stated numerical assumptions. The method enables block-scaled FP8 in all attention-core matmuls, and across the tested setups it matches BF16/FP32 mixed-precision loss and downstream performance.
In briefFlowMap-OPD: Separating Rollout and Kernel for On-Policy Distillation of Fast Flow-Map Image Generators
SummaryThis paper introduces FlowMap-OPD, an on-policy distillation framework for few-step flow-map generators such as MeanFlow and consistency models. It separates acquiring student states from comparing teacher and student distributions. On ImageNet experiments across three teacher rewards, instantaneous-velocity distribution supervision with independently tunable student consistency performed best. In text-to-image experiments, the method consolidates multiple specialists and outperforms multi-reward Flow-Map GRPO in task performance and convergence speed.
In briefRelayVSR: Efficient Real-World Video Super-Resolution via Large-Small Model Collaboration
SummaryThis paper presents RelayVSR, a streaming video super-resolution framework in which a large generative model produces reference latents for sparse keyframes, and a lightweight Dual-Memory Video Transformer super-resolves every frame using them. The large model is trained with Video-Aware Reference Optimization, which uses reinforcement learning with system-level and reference-level rewards. At 1080p on a single A100 80GB, the dual-endpoint version with a 15-frame keyframe interval reportedly reaches 29.29 FPS and 13.82 GB peak memory, compared with 7.80 FPS and 24.447 GB for FlashVSR-Tiny.
WhyProposes sparse keyframe relay with dual-level RL rewards, a substantive advance for video super-resolution, backed by baseline comparisons and speed and memory measurements.
In briefIs manual software optimization becoming obsolete? This paper tests whether LLM agents can optimize scientific software on their own.
SummaryThe researchers tasked an LLM-based agent with optimizing software for t-SNE, single-sample gene set enrichment analysis (ssGSEA), and graphlet counting, while humans defined the scope, correctness criteria, and verification. The agent worked autonomously, in some cases for several hours, and code maintainers reviewed and verified the results. The optimized implementations were faster in all tested configurations, by up to two orders of magnitude over the fastest existing tools, and included a new graphlet counting algorithm. The authors argue that for well-scoped, verifiable problems, manual software optimization may be a thing of the past.
Pavlin G. Poli\v{c}ar, Martin \v{S}pendl, Toma\v{z} Ho\v{c}evar
In briefHow Evaluation Choices Change Biomedical ML Results: A Pediatric Pneumonia Benchmark Study
SummaryThe paper studies the widely used Kermany pediatric chest X-ray dataset with nine image classifiers under a controlled evaluation protocol. It finds that evaluation choices such as freezing versus fine-tuning the backbone and the decision threshold affect results more than the choice among pretrained backbones. A classifier using only file properties scores 0.992 balanced accuracy within the training pool but 0.496 on the official test split, showing the split differs from the training pool. The authors close with a seven-item reporting recommendation tied to these measured effects.
Bhanu Prakash Vangala, Sowmya Guda, Latha Peddi et al.
In briefScaling Influence Functions in LLMs with Compressed Gradients
SummaryInfluence functions estimate how individual training examples affect large language model behavior, but storing full training gradients is too costly at LLM scale. The paper proposes eigenbasis-corrected one-bit gradient projection (EOGP), which uses EK-FAC to reduce dimensionality, applies PCA within the retained subspace, and then one-bit quantizes the coordinates to fit more within a fixed storage budget. On GPT-2, the authors report that EOGP predicts retraining outcomes more accurately than the evaluated baselines while using one-sixteenth of their per-example storage. On OLMo 2 SFT models from 1B to 32B parameters, it remains competitive with baselines given over 100 times as much storage per example.
In briefDetecting Alcohol Impairment in E-Scooter Riders Using Onboard Sensors and Machine Learning
SummaryThis study had 25 participants ride an instrumented e-scooter sober and at blood alcohol levels of 0.05% and 0.08%. Entropy and standard deviation analysis showed that intoxicated riders shift from continuous small corrections to fewer, larger reactive corrections. An entropy-based classifier reached 85% overall accuracy under leave-one-participant-out validation, with a sober-vs-high AuROC of 1.00.
Rahul Rajendra Pai, Marco Dozza, Alexander Rasch et al.
