In briefImportance-aware feature sparsification to cut communication in wireless split learning
SummaryThis paper addresses the communication bottleneck of sending high-dimensional intermediate features in wireless split learning. It proposes ICS, which has the server rank feature channels using Grad-CAM scores from the true-class logit, aggregated into a class-balanced importance vector that clients reuse without extra computation. Experiments reportedly show ICS consistently outperforms baselines, with larger gains under severe non-i.i.d. data partitions.
In briefDynamics to decision: a math theory linking Lyapunov spectra to classifier decision boundaries
SummaryThe paper models a deep classifier as a finite, nonautonomous discrete dynamical system, with layers as time steps, and studies the Finite-Time Maximum Lyapunov Exponent (FTMLE) of data trajectories. For Gaussian classes, the authors prove that the probability-level FTMLE aligns with the decision boundary normal, while the link weakens at the logit and hidden-representation levels, depending on the loss function and structural conditions. They propose geometry-aware fine-tuning and conditions for concentrating high hidden FTMLE near the boundary. Numerical results are presented to support the theoretical findings.
In briefViLegalExpert: A Large-Scale Vietnamese Benchmark for Legal Retrieval and QA from Real Consultations
SummaryThis paper introduces ViLegalExpert, a large-scale Vietnamese legal benchmark built from authentic citizen–lawyer consultations. It contains over 172K questions across 34 legal domains, with professional answers and expert-verified legal evidence, and supports legal information retrieval, extractive QA, and abstractive QA. Experiments show substantial difficulty in evidence retrieval and grounded answer generation, while pretrained models perform strongly on QA and hybrid retrieval achieves the best retrieval results.
Dat Tien Nguyen, Nghia Hieu Nguyen, Anh Thi-Hoang Nguyen et al.
In briefAttention Function as a Built-in Bias: How Models Behave Differently in New Contexts
SummaryThe paper proposes MoFA, a Transformer modification that fixes the ratio of softmax and sigmoid attention heads before training. Using a 124M-parameter GPT-2 with five ratios and five seeds, the authors find the ratio has little effect on in-distribution data but large effects under zero-shot shift across 15 out-of-distribution domains. The best ratio tracks whether text is short and informal or technical and long-form.
In briefLow-discrepancy dither for quantized recurrent state caches
SummaryThis paper studies how to store the recurrent state of Mamba-style and hybrid language models at low precision without losing accuracy over long generations. It proposes a deterministic golden-ratio Weyl dither that needs no random numbers and keeps the quantized model closer to the full-precision model than stochastic rounding, across models, storage formats, and decoding lengths, at no extra cost. The authors also show that round-to-nearest accumulates error by discarding small updates, and provide a discrepancy analysis explaining this ordering.
In briefDetection transformer estimating per-fiber microstructure from multi-shell diffusion MRI
SummaryThis paper reframes estimating white matter fiber orientation and microstructure from diffusion MRI as an object-detection-like task. It adapts the Detection Transformer (DETR) to predict mean diffusivity, fractional anisotropy, fiber direction, and signal fraction for a variable number of compartments per voxel from standard multi-shell linear-encoding data. On synthetic test data with up to five compartments per voxel, the model reports R² of 0.95 for MD, 0.88 for FA, and a median angular error of 4.2 degrees.
Sebastian Endt, Marcus Wirth, Johannes Reinhold Schlund et al.
In briefAligning Reasoning and Answers in Vision-Language Models with Probability Rewards
SummaryThis paper examines whether a vision-language model's reasoning matches its final answer in Visual Intention Grounding, where the model infers a target object from a human intention and predicts a bounding box. The authors find that prior IoU-based reinforcement learning can produce correct boxes while the reasoning points to a different object, a failure they call thinking drift. They propose Rita, which uses thinking and consistency rewards derived from reference-answer probabilities, plus difficulty-aware data filtering. Experiments on EgoIntention and the new RefEgo-Int benchmark show Rita outperforms supervised finetuning and standard RL baselines.
In briefMEND: Label-free detection, localisation and correction of hallucinated predictions in latent world models
SummaryThis paper studies hallucination in latent World Models, where a predicted latent state decodes to a scene that never occurs, and the error compounds because the model feeds its predictions back autoregressively. The authors propose MEND, a conditional score network trained by denoising score matching on real transitions, whose score field detects hallucination, localises it to image patches, and gives an inference-time correction direction. On two navigation environments, MEND reaches an AUROC of up to 0.80 for detection without using actions and a per-token AUPRC of up to 0.87 for localisation, and its correction reduces single-step latent error.
Ali J Alrasheed, Aryan Yazdan Parast, Basim Azam et al.
In briefWhitening makes linear probes less reliant on spurious correlations
SummaryThe paper studies why linear probes, which are linear models fitted on the representations of pretrained networks, can depend on spurious features. The authors show these models favor directions with large covariance eigenvalues, and that whitening, which equalizes those eigenvalues, removes this preference. Experiments on a synthetic data process and standard spurious correlation benchmarks show whitening improves robustness, and it can also improve existing approaches when added to them.
Floris Holstege, Bram Wouters, Noud van Giersbergen et al.
In briefFaithful chart generation for research: adapting visual plans to retrieved evidence
SummaryThis paper addresses how analytical charts in multimodal deep research must faithfully reflect the quantitative evidence retrieved from the web. The authors propose Frame-Evidence Co-Adaptation (FECA), which iteratively couples visual chart frames with retrieved evidence and decides whether each frame is accepted, revised, or dropped. Experiments on 100 real-world research topics report substantially improved numerical fidelity while maintaining report quality and chart utility.
In briefHow support vector machines behave on high-dimensional data with few samples under a spiked covariance model
SummaryThis paper studies the asymptotic behavior of support vector machines (SVMs) in high-dimension, low-sample-size (HDLSS) settings under a spiked model. The authors show that the usual geometric representation of HDLSS data fails here, and that standard SVM misclassification rates do not tend to zero, so it is not consistent. They propose a spike-corrected SVM (SC-SVM), which is consistent as the sample size grows, and check the classifiers with numerical simulations.
In briefToken-level perception-based advantage estimation to strengthen multimodal reasoning in AI models
SummaryThis paper addresses reinforcement learning for multimodal large language models, where existing methods use only coarse, sequence-level rewards and lack fine-grained supervision on visually grounded reasoning steps. The authors propose TPAE, which scores each token by its statistical consistency with the vision-entropy patterns of correct reasoning chains and uses these scores to modulate the sequence-level advantage. Experiments on seven benchmarks reportedly show TPAE consistently outperforms strong baselines with more stable and efficient optimization.
In briefWhen single attention-head ablations can support causal claims, and when they cannot
SummaryThe paper studies whether zeroing a single attention head and measuring the resulting performance change can justify causal conclusions about language models. Using GPT-2 small, it finds that post-projection ablation is nearly uncorrelated with corrected pre-projection ablation and selects a disjoint top-5 set of heads, while binary accuracy hides effects near floors and ceilings. With matched random and layer-matched controls, the corrected ranking is stable across splits, but evidence for task specificity is not robust; replication on DistilGPT2 preserves the intervention and control findings.
In briefDeep learning for combined electromagnetic, analog and digital MIMO precoding
SummaryThis paper studies antennas that can switch between several radiation patterns, adding an electromagnetic-domain degree of freedom to multi-user MIMO precoding. The authors propose Tri-PNet, a Conformer-based network that jointly learns electromagnetic, analog and digital precoding without supervision by maximizing average sum spectral efficiency. Simulations under two radiation-pattern modes show it outperforms random EM precoding and hybrid MIMO without EM precoding, approaches a greedy search scheme with much lower online complexity, and stays robust to imperfect channel state information.
In briefPredictive 4D Belief for Navigating Environments That Change Unseen
SummaryThis paper studies navigation in evolving environments, where targets may move while unobserved, so remembered locations can be outdated on arrival. The authors propose EvolvingNav, which builds a time-indexed belief from timestamped 3D object histories, updates it with RGB-D evidence, and uses a frozen vision-language controller to act. They also introduce EvoWorld-Bench, with 54 scenes and 803,680 tasks. Simulation and real-robot experiments report improved navigation success and search efficiency over the evaluated baselines.
In briefQuanVI: Score-Based Variational Inference with Quantum Mixed States
SummaryThe paper proposes QuanVI, a scalable quantum-inspired variational inference method that minimizes the Fisher divergence between a variational distribution and a target distribution. It represents degenerate low-energy subspaces with a density operator (maximally mixed state) and compresses it using a matrix product operator quantum tensor network to avoid exponential parameter growth. Experiments show agreement with exact solutions in low dimensions and scaling to high-dimensional synthetic and Bayesian posterior benchmarks, including non-Gaussian targets.
In briefTraining-free fusion of neural and embedding-based speaker diarization
SummaryThis paper proposes Training-Free Affinity Fusion (TFAF), which integrates speaker structure from a neural diarizer into an embedding-based diarization system. It builds a continuous affinity matrix combined with embedding-based acoustic affinity, followed by one global clustering step, without additional training. Experiments on AMI and CALLHOME show consistent DER improvements over both constituent systems, and on AMI the fusion also improves speaker-attributed transcription. Ablations indicate the neural speaker partition accounts for most of the gain.
In briefTripleFlow: Training-free video object removal combining residual editing and generation
SummaryRemoving an object from video is hard because the model must reconstruct the hidden background from context alone. TripleFlow is a training-free framework with source, residual, and synthesis flows: the residual flow suppresses the object, while the synthesis flow rebuilds the occluded background and feeds it back into the editing process at every step. The authors report state-of-the-art results across five benchmarks in reconstruction fidelity and temporal consistency.
In briefDAGent: Evaluate-then-Grow Planning for Deep Research AI Agents
SummaryThis paper presents DAGent, a multi-agent framework that organizes deep research tasks as directed acyclic graphs (DAGs). Instead of fixing the whole plan in advance, an Orchestrator grows the graph incrementally, conditioning each batch of new sub-tasks on confidence and uncertainty signals from completed nodes. The authors also propose DAGRPO, a reinforcement learning adaptation of GRPO that uses graph-topology-based credit signals. Experiments on BrowseComp-Plus, GAIA, and xbench-DeepSearch report gains over open-source baselines.
WhyProposes incremental batch-wise DAG growth with topology-conditioned RL signals, supported by multi-benchmark comparisons and ablations.
In briefSimple linear memory can distil a world-model policy for robot air hockey
SummaryThis paper studies memory-dependent control in simulated air-hockey defence, where the puck is temporarily untracked. A DreamerV3 teacher beats a memoryless policy under tracking loss, and resetting its recurrent state sharply hurts performance, showing the task needs memory. Distilled into 64-dimensional recurrent students, a purely linear recurrent model matches the GRU baseline and the teacher across five seeds, and adding nonlinear innovation rank gives no measured benefit. Conclusions are limited to the simulated task, teacher, state size and blackout horizon studied, with nonlinear observation encoders and action heads retained.
In briefOP-CAD: Clean-audio teacher distillation for noise-robust audio-visual question answering
SummaryThis paper studies how omni-modal large language models answer audio-visual questions when environmental noise and competing speech are present. It proposes OP-CAD, which trains with a curriculum from mild to severe noise and uses a frozen teacher that sees clean audio and the verified answer to guide token-level supervision of the student's responses. Token weights are set by comparing teacher predictions under clean, corrupted, and visual-only inputs, emphasizing acoustically sensitive positions. The authors report that OP-CAD outperforms compared methods under all evaluated noise conditions and better preserves clean-correct answers under strong interference without an aggregate clean-accuracy penalty.
In briefRep2Skill: Agents Improve Their Text Skills by Reading Their Own Internal States
SummaryThis paper introduces Rep2Skill, a framework that lets LLM agents refine their external textual skills by analyzing their own internal representation trajectories. It localizes turns that deviate from successful execution and turns these signals into targeted skill revisions. Experiments on two agent environments with two open-source LLMs show it outperforms text-only self-evolution methods.
In briefDo Self-Evolving Skills Generalize to Held-Out Tasks?
SummaryThis paper asks whether reusable skills that AI agents refine during training tasks still improve performance on new test tasks. The authors compare five self-evolving methods and a one-shot skill across six benchmarks, finding that of 21 skills improved on training tasks, only 5 kept all gains on test tasks. They propose Generalizable Skill Optimization (GSO), which keeps only a guide for writing skills and creates a new skill for each task, and report it scores highest on all six benchmarks.
In briefMindWorldBench: Testing Whether Video Models Act on Characters' Mental States
SummaryThis paper introduces MindWorldBench, a benchmark that tests whether image-to-video models can generate actions driven by a person's beliefs, desires, and perception, without explicit action prompts. It uses 744 prompts with a zero-action and counterfactual design, and an automated pipeline that scores video quality, commonsense plausibility, and mental-state consistency. Across 11 evaluated models, visual fidelity and physical reasoning were strong, but behaviors often failed to match latent mental states. The authors name a failure mode, Omniscient Bias, where models default to the objective world state instead of a character's subjective belief.
In briefMADBench: A Benchmark for Testing Security in Multi-Agent AI Debate
SummaryThis paper introduces MADBench, a benchmark for evaluating the security of multi-agent debate, where several language model agents exchange and critique answers. The authors organize attacks into a layered taxonomy and test six attack families across 356 tasks and 3,958 test cases. They find that debate does not necessarily improve reasoning under attack: it can reduce attack effects on answer accuracy in question-answering tasks but amplifies unauthorized reads or writes in both question-answering and workspace tasks.
In briefBlackout vs. Freeze: How VLA Robot Models Fail Physically When Cameras Break
SummaryThis paper analyzes how the π0.5 and GR00T vision-language-action (VLA) robot models behave under camera faults such as image blackouts and frozen frames. Even when task success rates are similarly low, freezing produces more extreme joint movements, while blackout after gripper closure leads to more dropped objects, especially without proprioception. Two mitigation methods, camera-blackout training and training-free replacement of faulty visual embeddings, improve success in some conditions but can increase unintended contact, and real-robot trials show unintended physical interactions can still occur during successful execution.
In briefA sharp Gaussian approximation for the long-run distribution of constant-step SGD
SummaryThe paper proves a sharp Gaussian approximation for the invariant law of constant-stepsize SGD with bounded additive noise driven by a uniformly ergodic Markov chain. For smooth, strongly convex objectives with a Lipschitz Hessian and nondegenerate long-run noise covariance, the normalized centered iterate is O(√α)-close to its limiting Gaussian in 1-Wasserstein distance. The proof combines blockwise Gaussian comparison with long-run contraction. A four-state example shows the bound is tight.
WhyProves a sharp Gaussian approximation for constant-stepsize SGD invariant laws, with a matching lower bound example; strong theory but narrow scope.
In briefRefCon: Iterative Refinement and Contrastive Memory Extraction for Context-Evolving Agents
SummaryThis paper proposes RefCon, which combines sequential self-refinement and parallel self-contrast to extract higher-quality memories from agent experience without gold labels. Evaluated on AppWorld and BFCL-V3 across several context-evolving agent frameworks, it reports relative gains of 21.6% on ACE and 16.6% on ReMe over no-scaling baselines. The authors also note that RefCon keeps improving as more trajectories are used, unlike diversity-only scaling.
Ubaidillah Ariq Prathama, Bo Liu, Yeo Boon Hong et al.
In briefCan Text Replace Vision? Structural Limits in Diagram Reasoning
SummaryThis paper asks whether structured text can replace images for diagram reasoning and proposes a protocol to tell whether errors come from missing information or from a solver failing to use it. On 240 held-out public FlowGen diagrams, gold structure reached 87% accuracy, while direct vision and learned text both stayed below 30%. Retrying only invalid extractions left accuracy essentially unchanged, and topology relevant to the question predicted correctness better than whole-graph topology.
In briefSchema: Agent Learning of Unknown Environments via Program Induction
SummaryThis paper introduces Schema, an agent framework that helps LLM agents learn unfamiliar environments by inducing executable programs from interaction. The agent expresses its evolving understanding as programs stored in a persistent workspace, which are checked against history, planned within, and executed with step-by-step verification. The authors report that with the same base model, Schema raises ARC-AGI-3 RHAE from 58.7% to 99.2%, solves all public DiG-bench games, and reaches the median of top-50 human players on MazeBench.
In briefBELIEFRAG: Making Adaptive RAG State-Aware as Evidence Changes
SummaryThis paper proposes BELIEFRAG, a closed-loop controller that maintains an explicit state over evidence sufficiency, reliability, conflict, uncertainty, gaps, and acquisition cost during multi-step retrieval. It chooses among retrieving, rewriting queries, verifying, answering, stopping, or abstaining. On six QA benchmarks with gpt-oss-120b, it reaches a mean token F1 of 0.572 using about 3.89k tokens per question, beating fixed iterative retrieval (0.555) while using 39% fewer tokens. Main gains come from corrective re-retrieval, with calibrated answerability playing the strongest operational role.
In briefID Balancing: Stable Training of Very Sparse MoE Models Using PID-Based Load Control
SummaryThis paper addresses the problem that expert load imbalance in Mixture-of-Experts language models worsens as routing becomes sparser, hurting training stability. It recasts two existing auxiliary-loss-free balancing methods as incomplete PID controllers and proposes ID Balancing, an integral-derivative controller that strengthens corrections for large or worsening load errors. Across Top-10, Top-5, and Top-3 routing over 768 experts, the authors report lower worst-case imbalance and competitive language-modeling and downstream performance.
In briefWhen Harnesses Lose the Signal: Causal Evaluation of Recovery in LLM Agents
SummaryThe paper argues that LLM agents use external harnesses to pass information and recover from execution errors, but average success rates hide the fact that the same recovery step can rescue or harm a trajectory. The authors treat recovery as a causal decision problem and introduce the Causal Intervention Router (CIR), a lightweight policy that decides when to intervene. On long-horizon ALFWorld tasks with Qwen3-14B, CIR raises success from 70.33% to 73.33%.
In briefAsking the World: agentic world modeling and probing for general physical reasoning
SummaryThe paper introduces Asking the World (ATW), a general agent that builds a task-relevant executable world model from video and then queries, simulates, and intervenes on it to gather evidence for a question. It also presents PolyWorld Engine, a Warp-based multiphysics simulator covering rigid, soft, cloth, rope, and fluid interactions. ATW is evaluated on CLEVRER, ContPhy, and three real-world scenarios, where it outperforms direct Gemini-3-Flash and GPT-5.5 inference.
WhyProposes an adaptive agent that builds and probes physical worlds with a custom simulator, outperforming several baselines on CLEVRER, ContPhy, and real scenes.
In briefEgoVine: Post-training planners on branching experience from real videos
SummaryEgoVine converts annotated real-world egocentric household videos into executable symbolic environments, letting an agent explore alternative actions and compare their outcomes under partial observation. The model jointly trains action generation and consequence prediction, using verified repairs from failed attempts as recovery targets. On a shared benchmark, EgoVine-1.5B reports 81.8% success on main planning tasks and 66.7% on longer task compositions, outperforming all evaluated Flash LLM baselines.
In briefStealthy attack that fools vision-language models only after visual token compression
SummaryThis paper studies adversarial images that keep full-token inference correct but cause errors once visual tokens are compressed, a failure that full-token evaluation can miss. The authors propose Feature-Aware Token Attack (FATA), which combines attention suppression with cosine-based feature preservation on salient clean-image tokens, using only vision-encoder gradients. The reported results show higher full-token accuracy retention and conditional blinding rate than the CAA baseline, and the lowest detection rate among four attacks across three detectors at a 5% false-positive rate.
In briefUncertainty-aware consistency distillation for faster few-step video generation
SummaryThis paper studies distilling a multi-step video generator, which needs dozens of sampling steps, into a student that generates video in only a few steps. The proposed method, UACD, estimates a local uncertainty for each spatiotemporal region by comparing the student's prediction with a consensus target built from two independently perturbed teacher-guided paths, and it relaxes the consistency penalty where uncertainty is high. The approach also adds feature-space adversarial training with semantic alignment and uses LoRA on a 50-step Wan model. The authors report a 0.556 mean score on VBench 2.0 for 4-step generation and state that users preferred it over competing methods.
In briefEvaluating High-Bandwidth Flash as Extra Memory for LLM Serving
SummaryThis paper studies whether high-bandwidth flash (HBF) can add memory capacity for serving large language models, where weights and KV caches strain memory. It proposes an HBM-HBF-host storage hierarchy and buffered cache-aware scheduling, and uses trace-driven simulations to measure performance, energy, and HBF write lifetime. The fastest HBF-augmented systems cut completion time by 36.1-87.7% versus HBM-only systems, with modeled energy savings up to 59.1%. Buffered scheduling extends estimated HBF write lifetime from 1.21 to 14.82 years in the evaluated setup.
In briefModeling Perceived Color Differences with Machine Learning and Human Judgments
SummaryThis paper proposes a data-driven method that estimates color differences directly from human ratings. Seven observers rated 2,000 systematically generated color pairs on a four-point ordinal scale, and these ratings were used to train regression models on RGB, HSI, and COLIBRI fuzzy-category features. The study found that the choice of color representation mattered more than the regression algorithm, and LightGBM using the combined representation performed best, with an R² of 0.703.
Elnara Kadyrgali, Muragul Muratbekova, Adilet Yerkin et al.
In briefGeoGAT: Worldwide Video Geo-localization Using Bidirectional Sampling and Hierarchical Graph Attention
SummaryThis paper proposes GeoGAT, a method for inferring where a video was filmed worldwide, evaluated at city, state/province, country, and continent levels. It combines bidirectional temporal frame sampling with graph attention networks and a dual-constraint mechanism that removes conflicts between hierarchy levels. The authors also build GeoGAT10k, a benchmark of 9,720 multi-shot edited videos from 166 cities. Reported results show state-of-the-art performance on CityGuessr68k and GeoGAT10k, with accuracy gains exceeding 24 percentage points on the latter.<br>
In briefReducing Localization Ambiguity: Geometry-Semantic BEV Learning for Satellite-Ground Localization
SummaryThis paper addresses satellite-ground localization, which estimates a ground camera's planar position and yaw within a geo-referenced satellite image. The proposed GeoSem-BEV method adds radial depth and vertical height supervision plus shared semantic supervision to constrain BEV feature placement and improve descriptor discriminability. On VIGOR, DReSS-D, and KITTI-CVL, it reduces orientation error relative to state-of-the-art BEV localization models.
In briefCDMD: A Cross-Dataset Mixed-Type Diffusion Model for Tabular Data
SummaryThe paper introduces CDMD, a diffusion model for tabular data trained jointly across multiple datasets with different schemas and mixed numerical and categorical features. It defines diffusion directly over the mixed-type feature space and uses a shared schema-aware Transformer denoiser. On seven real-world datasets, a single jointly trained CDMD achieves the highest average generation quality among the baselines while using fewer total parameters. Pre-training on 337 datasets also improves generation on unseen datasets.
Mohamed Amine Ketata, Maximilian Schambach, Stephan G\"unnemann
In briefAuditable GP-Based Near-Optimality Certificates via Prequential E-Values
SummaryThe paper studies when sequential optimization of an expensive black-box function, such as hyperparameter tuning, can stop with a certificate that the best evaluated value is within ε of the global optimum. It selects among predeclared GP/RKHS upper envelopes using prequential e-values, deletes contradicted candidates, and certifies with the largest surviving upper bound. On a 512-seed noisy RBF sweep, it roughly halves false-certification risk relative to fit-then-certify at comparable power.
In briefDoes Better Teacher Supervision Suffice? Strengthening Student-Side Learning in Multimodal On-Policy Distillation
SummaryOn-policy distillation (OPD) improves reasoning by having a teacher give token-level feedback on a student's own outputs. This paper finds that the student's limited visual perception is another key bottleneck in multimodal OPD, since supplying correct visual facts still substantially improves students. The authors propose S-OPD, which adds Teacher-calibrated Policy Contrast and Policy Agreement objectives to strengthen perceptual learning. It can be plugged into existing OPD frameworks and shows consistent gains on eight benchmarks, with up to 4.25 points on LogicVista.
In briefDiagnosing self-distillation that helps language models reason
SummaryThis paper examines on-policy self-distillation (OPSD) for mathematical reasoning in language models from 0.6B to 8B parameters. Using controlled experiments and token-level analyses, the authors find that the teacher's signal depends on reasoning-mode alignment and the full teacher prefix, not only on privileged semantics. OPSD improves reasoning only in narrow compatible regimes, and otherwise causes ineffective length growth, stable degradation, or behavioral collapse.
In briefM²Weather: A Benchmark for Joint Multi-Station and Multi-Variable Weather Forecasting
SummaryThis paper introduces M²Weather, a benchmark for forecasting weather by jointly modeling multiple stations and multiple variables. After quality control, it collects 2,809 stations with five physically coupled variables across France, Europe, and the globe. It also proposes a lightweight plug-and-play adapter that adds missing station or variable relationships without retraining. Evaluating 16 models, the authors report that jointly modeling both relationships helps, and completing the missing relationships further reduces MSE on all three datasets.
In briefGRC-Pose: Tracking unseen objects' 6D pose from video without prior models
SummaryThis paper tracks the 6D pose of an unseen object from a single RGB video, without CAD models, reference images, or pose labels. It introduces GRC-Pose, which matches generated object geometry from SAM3D CAD with reconstructed scene evidence through learned correspondences and robust pose estimation. With SAM3D CAD on HOT3D, it reports state-of-the-art Average Recall and a 58% improvement in motion retention over prior work. It also remains highly competitive on YCBInEOAT and LINEMOD.
In briefModels share a taste for model-written text, but no own-model preference is detected
SummaryThis paper re-analyses the public data from Laurito et al. (PNAS 2025), in which language models chose between two descriptions of the same product, paper abstract or film. Using a two-way fixed-effects model and an exact permutation test, the authors find no detectable own-model premium on any dataset or pooled. The original finding that models prefer model-written text still holds.
In briefBeyond Local Linearity: Scale-Resolved Geometry of Learned Image Encoders
SummaryThis paper studies how image encoder features respond to finite input changes, introducing a scale-resolved statistic that compares measured feature displacement with its local linear prediction as perturbation size grows. Across diverse image encoders, the authors identify a characteristic plateau-rise-peak-decay profile called the bump. The bump is absent at initialization, appears early in standard training, and does not form under randomized labels or random-noise inputs, while its shape varies with training distribution and robustness objective.
Jakub Szymkowiak, Wojtek Pa{\l}ubicki, Kamil Adamczewski
In briefOceanMind: A Multi-Agent AI System for Ocean Diagnosis
SummaryThis paper presents OceanMind, a multi-agent AI system that couples large language models with time-dependent 3D ocean data to support ocean diagnosis. The system runs analysis through four stages: query routing, skill-based planning, tool execution, and evidence-based summary generation, guided by 63 reusable ocean-specific skills. On a benchmark of 240 computational-workflow queries, it achieved relative improvements of 41.2% in effectiveness and 21.5% in efficiency over general ReAct agents. The authors also report reproducing published oceanic diagnostics and supporting environmental decision-making.{}
In briefThe Row Normalization Puzzle in Muon: Theory vs. Practice in Optimizer Design
SummaryThis paper studies how row-wise renormalization affects the Muon optimizer, focusing on the gap between NorMuon's worst-case guarantees and its practical performance. The authors show that under operator-norm geometry, row normalization adds a dimension-dependent factor to worst-case iteration complexity, even with exact polar computation and any fixed momentum parameters. They establish an algorithm-dependent lower bound and a matching upper bound in deterministic settings, and extend the upper bound to stochastic settings. Experiments find NorMuon slower than Muon on synthetic problems but better on LLM pretraining.
In briefCamAgent: An LLM-agent framework that unifies multi-species camera-trap analysis workflows
SummaryThis paper presents CamAgent, an autonomous large language model agent framework that integrates camera-trap analysis into one system. Users express ecological goals in natural language, and the agent runs tools for species recognition, CamtrapDP-compatible data management, occupancy modeling, temporal activity analysis, and species co-occurrence networks. The authors state it reduces the manual programming required of conservationists.
In briefDeepJEPA: Scaling World Models by Deepening Internal Computation
SummaryThis paper introduces DeepJEPA, a weight-tied joint-embedding predictive world model that treats the depth of each transition as an inner test-time scaling axis. It learns when an extra recurrent update is worth computing for each candidate and rollout step, averaging only 1.00 to 1.26 updates per transition. Across five visual-control settings, it improves or matches the strongest fixed-depth planner, with extra computation concentrated at contact onset and sustained object interaction.
SummaryThis paper introduces Bongard, an open-weight model that treats machine intuition as a separately designed and trainable capability. It uses a T5Gemma 2 encoder-decoder that reads a situation once and then answers multiple questions about it, trained in three stages. On DecisionBench it reaches 78.05% accuracy and ranks fourth of 61 systems.
In briefT-Router: Parameter-efficient reinforcement learning for reasoning via thalamic-style routing
SummaryThe paper introduces T-Router, a method that improves reasoning in a pretrained model by learning to reuse earlier computed block contributions through a compressed addressable bank and a depth-recurrent controller. It trains only 41.73M parameters, about 0.466% of an 8.95B backbone. After GSM8K reinforcement learning, it reports higher MathAvg scores than full-parameter GRPO and LoRA, and mean AIME accuracy rises from 48.33 to 60.56.
In briefMASCRDM: A multi-agent system that detects and reduces compliance risks during large language model training
SummaryThis paper addresses the challenge of making large language models comply with rules against discrimination and bias, noting that existing methods mostly filter inputs and outputs after training. The authors propose MASCRDM, a multi-agent system built on AI laws and a compliance knowledge graph that monitors key model components throughout training and issues risk alerts and suggestions to developers. Experiments on discrimination and bias benchmarks show improved compliance while maintaining reasonable semantic performance.
In briefReserve-aware certificates for conservative bandits with uncertain baselines
SummaryThis paper studies conservative bandits, where a policy must improve on an incumbent without exhausting a set performance budget, and the incumbent's reward is uncertain. Separately bounding candidate and baseline rewards double-charges shared estimation error, so the authors build Reserve-C4B around the baseline-relative contrast, using a shared confidence set and a reserve ledger. For linear rewards, self-normalized confidence sets give simultaneous validity over time and adaptive candidates. Experiments show large reductions in baseline fallback.
In briefFalse Frontiers: Catching Co-Cheating in Self-Evolving Search Agents
SummaryThis paper studies co-cheating, a failure in self-evolving search agents where the question-generating proposer and the answering solver agree on shared errors, raising internal reward without improving real correctness. The authors propose multi-sample verification (MSV) and CrossFit, which scores proposer questions with an auxiliary solver trained on disjoint source documents. CrossFit lowers false-agreement rates for Qwen3.5-4B and 9B and improves average results across seven search benchmarks over standard coupled self-evolution and Search-R1.
In briefBeyond Prediction: Retrospective World Modeling to Guide VLM Agents
SummaryThe paper proposes Retrospective World Modeling, which lets agents infer backward the action most likely to have caused an observed state transition, rather than only predicting future outcomes. The authors define a Self-Consistency Reward (SCR) that measures the probabilistic agreement between the policy's action and this retrospective explanation, and use it in reinforcement learning to give dense transition-level feedback. Experiments across diverse agentic tasks report improved policy robustness and generalization over prospective-only world modeling baselines.
In briefMoRA: Pruning Mixture-of-Experts Models via Router Bias Learning and Expert Approximation
SummaryThis paper proposes MoRA, a framework that removes experts from Mixture-of-Experts (MoE) models to reduce memory use. It learns a router bias for each expert, optimizing a language-modeling loss together with a routing-diversity regularizer, and after pruning approximates removed experts' outputs with affine transformations of the remaining experts. The authors test on Qwen3-30B-A3B, DeepSeek-V2-Lite, and Moonlight-16B-A3B, removing 25% and 50% of routed experts per layer, and report better results than prior pruning methods on nine zero-shot benchmarks.
In briefWhat Should an Agent Remember? Separating Retention from Retrieval in Limited-Memory Evaluation
SummaryThe paper argues that a persistent agent must decide both what to keep as information arrives and what to surface when a query comes, but existing memory evaluations often confound the two. The authors build a streaming-recall benchmark crossing retention and selection rules, testing every condition on the same 300 seeded episodes. With access held fixed, query-aware selection improves required-fact recall by 15.5 points, and under bounded retention all observed failures stem from eviction rather than ranking errors. Recall drops to 0% as target facts recede further into the past.
In briefManifold-constrained initial noise optimization for efficient generative model alignment
SummaryThis paper proposes ZeNOVA, a gradient-free method for aligning deterministic one- or few-step generative models by optimizing their initial Gaussian noise. It combines annealed soft-value guidance, manifold-constrained hyperspherical Langevin dynamics, and Metropolis-Hastings jumping to handle black-box rewards. Experiments on image and video generative models report that ZeNOVA raises rewards more stably than all evaluated zeroth-order baselines.
In briefA Generalisation Signal Is Not Always a Model-Selection Signal
SummaryThis paper examines whether properties of trained networks can select models when validation data does not reflect deployment conditions in computational biology. The authors propose a forward-only proxy motivated by the Hessian norm and compare it with common Hessian measures on molecular property, protein fitness, and drug-response tasks. The proxy correlates with generalisation gap on most tasks, but it does not reliably identify the best model for deployment.
Aditya Nagarsekar, M P Ashish Bhat, Aadi Nesarkar et al.
In briefDiFF: Doppler-Guided Flow Matching for Human Motion Estimation
SummaryThis paper presents DiFF, a generative framework for estimating human motion from sparse, noisy 4D millimeter-wave radar point clouds. It combines Doppler velocity priors with a Kolmogorov-Arnold Network-based conditional flow matching model to regularize this ill-posed problem. The authors report state-of-the-art results across several real-world datasets, with 3D endpoint error reduced to the millimeter scale on the mmBody benchmark.
In briefDeCoPrune: Denoising-Consistency KV-Cache Pruning for Autoregressive Video Diffusion
SummaryAutoregressive video diffusion models store a growing key-value cache of generated history, which raises inference cost. The paper proposes DeCoPrune, a training-free method that keeps tokens whose intermediate clean prediction differs strongly from their final denoised value and prunes the rest. The authors also introduce CMBench, a benchmark of 58 context episodes and 116 recall-oriented continuation tasks for testing long-range information retention. Experiments report that DeCoPrune prunes over 85% of historical tokens while keeping long-range recall close to full cache and speeding continuation generation by more than 4 times.
In briefA new efficient method for learning constrained long-run average-reward decision processes
SummaryThe paper studies infinite-horizon average-reward constrained Markov decision processes under a weakly communicating assumption. It proposes what the authors believe is the first computationally efficient algorithm achieving Õ(√T) regret and cumulative constraint violation with high probability in the tabular setting. The method adds cumulative violation to the state and uses differences of a Huber potential to reshape the reward. The approach keeps the √T rate without worsening dependence on T.
WhyGives the first computationally efficient algorithm with near-optimal sqrt(T) regret and constraint violation via state augmentation and Huber-potential reward reshaping.
In briefDeep Learning for Railway Anomaly Detection: A Structured Survey
SummaryThis paper surveys deep learning methods for detecting anomalies in railway systems. It organizes existing approaches with a unified taxonomy based on anomaly location, data representation, sensing modality and temporal characteristics. The survey also covers data challenges, evaluation metrics and deployment aspects such as edge-cloud architectures. It closes with a decision-oriented framework linking anomaly and data properties to suitable detection paradigms.
In briefSteepest Guidance: a practical, principled way to align flow and diffusion models at inference time
SummaryThis paper studies how to align flow- and diffusion-based generative models with desired objectives during inference. The authors treat alignment as a sequential optimization problem over probability measures and propose Steepest Guidance, which maximizes local improvement in the objective. The paper provides a theoretical analysis of the method and reports extensive experiments demonstrating its effectiveness.
Shokichi Takakura, Akifumi Wachi, Rei Higuchi et al.
In briefA strategic roadmap for building a sustainable atomistic machine-learning ecosystem
SummaryThis paper is a strategic roadmap summarizing discussions held at CECAM in Lausanne in January 2026. The discussions covered algorithms, models, software and hardware infrastructure, and applications of AI in atomic-scale simulations. The authors propose long-term goals and concrete actions to establish a healthy, sustainable and impactful atomistic ML community.
J\"org Behler, Michele Ceriotti, Cecilia Clementi et al.
In briefUGOD: Uncertainty-guided opacity and dropout to improve sparse-view 3D Gaussian splatting
SummarySparse-view 3D Gaussian Splatting tends to overfit because many Gaussian primitives are weakly constrained by limited views. UGOD estimates a view-dependent uncertainty score for each Gaussian and uses it to attenuate high-uncertainty primitives' opacity before compositing, with a detached soft-dropout branch during training. Experiments on Mip-NeRF 360 and LLFF report improved sparse-view novel-view synthesis and more compact Gaussian representations than compared methods.
In briefSCIC: A speaker-adapted TTS method for controlling expressive prosody using codebook-aware instructions
SummaryThis paper addresses limited clause-level prosody control in long-form live-streaming text-to-speech by introducing speaker-relative inline controls for pitch, energy, speed, and pause. The authors analyze Qwen3-TTS residual codebooks, finding energy concentrates in early codebooks while pitch accumulates in deeper ones, and propose SCIC, which combines a temporal instruction router with tag-specific codebook weighting. Reported results show improved pitch and energy control over standard instruction fine-tuning and a more distinct paragraph-level expressive hierarchy in long-form synthesis.
In briefTrustworthy Runtime Error Healing in Real-World Repositories: Benchmark and Guardrail
SummaryThe paper studies using LLMs to generate code that repairs a crashed program's live runtime state so execution can continue. The authors build HealBench, a benchmark of 265 runtime errors from 18 real repositories, and HealGuard, which uses static and dynamic taint analysis to check whether healing-changed state reaches developer-protected operations. The best setting resumes execution in 38.11% of instances and passes the target test in 28.68%. HealGuard flags 17.4% of passing executions as risky, and on controlled cases detects all unsafe ones with a 68.42% false positive rate.
In briefMRI super-resolution and segmentation with task-aware enhancement
SummaryThis paper studies whether super-resolution and quality enhancement of 1.5T brain MRI improve tissue segmentation, and for which segmentation models. The authors propose an unpaired, physics-guided training pipeline for a lightweight recurrent enhancer. On ABIDE, enhancement significantly improves the wavelet segmenter, degrades U-Net, and leaves Swin-UNet unchanged. On IXI, enhancement lowers Dice for all pairings, mainly due to spatially implausible CSF voxels in the labels.
In briefShared Phase and Retention Control for Efficient Adaptive Spectral Recurrence (SPARC): A Sequence Model Using Two Shared Signals to Control Recurrent Memory
SummaryThe paper notes that Transformers' computation and cache costs grow with context length, while fixed-state recurrent models offer constant-memory inference, though existing spectral recurrences use static transitions. The authors propose SPARC, which uses just two input-dependent scalar signals to coordinate retention and phase rotation across all complex memory modes. SPARC reports relative gains over second-best methods on Walker-P and FordA, and reduces recurrent-mixer training latency by 18.2%-34.2% on an NVIDIA Blackwell GPU.
In briefUsing AI coding agents on coding theory problems, with lessons on verification errors
SummaryThe authors spent five weeks using an LLM coding agent on open coding theory problems: finding large sets of four-letter words, such as DNA barcodes, that stay far apart in edit distance. Restricting the search to codes with a prescribed symmetry raised the best known code of length 6 and minimum edit distance 3 from 114 to 120 words, and the same pipeline improved twelve further lower bounds. The paper also reports failures, where intermediate results were treated as facts without rechecking, causing search to stop early, and notes that checking only final outputs did not catch these errors.
In briefCross-session motor decoding via association-profile conditioned set-temporal transformer
SummaryIntracortical motor decoders lose accuracy across sessions as recorded units change and persisting units shift in how they relate to behavior. The paper proposes APST, which keeps all network weights frozen and uses a few labeled calibration trials to compute a four-dimensional association profile per unit, which conditions a set-attention encoder and a causal transformer. On held-out DANDI688 sessions from two monkeys, velocity R² reaches 0.78 and 0.81, above a neural-activity-only variant. On FALCON private held-out evaluation, R² is 0.65, 0.42, and 0.44 on M1, M2, and H1.
In briefHow parameter symmetries shape representation geometry in overparameterized nonlinear networks
SummaryThis paper studies how the geometry of representations in neural networks relates to the computations they perform. The authors show that a broad class of parameter symmetries can change representational geometry while leaving the network's function exactly the same. These symmetries act on representations through three basic feature transformations: addition, duplication, and scaling. The paper also shows that implementation-level selection rules can resolve the resulting degeneracy, so that features are weighted by their contribution to the function.
WhyReduces parameter symmetries to three feature transformations and gives a closed-form decomposition of representational geometry, supported by rigorous derivation.
In briefMulti-LLM Collaborative Alignment via Stackelberg Games: an adaptive instruction curriculum for language models that learn from each other
SummaryThis paper proposes Stackelberg Alignment, a leader-follower framework in which multiple language models learn from one another's responses. A leader EXP3 bandit allocates a fixed sampling budget across instructions, updating its distribution using a reward based on instruction difficulty and response discriminability. Follower models answer, judge each other's responses, and learn via DPO or GRPO. Experiments on three model pools and 12 benchmarks reportedly show the highest macro-average, with gains over the strongest training-time baseline of up to 7.4%.
In briefHO-FL: Hybrid-order federated learning for memory-limited edge devices
SummaryThis paper proposes HO-FL, a federated learning framework for memory-constrained edge devices that trains the bottom layers of a model with zeroth-order optimization and the top layers with first-order optimization. Each device chooses its split point based on its memory budget. The convergence analysis identifies a trade-off: clients with larger first-order segments give more accurate updates, but favoring them can underrepresent other clients' data, leading to a sampling optimization problem. Experiments on language tasks show that hybrid-order local training retains much of full first-order performance while substantially reducing client memory requirements.
In briefBeyond Text: LLM-Based Dimensional Emotion Evaluation in Dialogue
SummaryThis paper proposes a large language model framework that performs discrete emotion recognition and Valence-Arousal-Dominance (VAD) dimensional evaluation on the IEMOCAP dataset, adding acoustic cues as natural language descriptions. Six models from the LLaMA, GPT, and Qwen families are compared under zero-shot, few-shot, and LoRA fine-tuning settings. LoRA fine-tuned LLaMA models outperform prompt-engineered GPT models on both tasks, and the best model reaches a Valence CCC of 0.7822.
In briefLexReward: A Taxonomy-Driven Reward Framework for Legal Language Models
SummaryThis paper introduces LexReward, a reward modeling framework for legal language models that scores legal responses along three dimensions: Style, Element, and Chain. Each dimension has rubrics defining evaluation criteria and quality levels, and the rewards are used to build preference data for Direct Preference Optimization and reward-model training. The authors report that DPO on this data improves performance across all three dimensions. The learned reward models, LexRM, can improve policy performance in their corresponding dimensions through reinforcement learning without requiring reference answers.
In briefCORE: Conflict-Oriented Reasoning Elimination for Checkable Language-Model Search
SummaryThe paper presents CORE, a search controller that asks a verifier for a certified conflict core, backjumps to the latest decision in that core, and caches the conflict to avoid repeating it. Under sound verification and finite branching and depth, the uncapped search is shown to be complete and never prunes a valid solution. On planted graph-coloring instances, CORE uses fewer verifier calls than chronological repair, and on five reasoning tasks it achieves higher mean success than Tree of Thoughts with two Qwen models.
In briefSparseEngine: A Sparse-First Inference Engine for Long-Context LLM Agents
SummaryLong-context LLM agents accumulate large interaction histories that burden KV-cache memory and attention computation. SparseEngine is a sparse-first inference engine with a shared lifecycle contract that lets each method manage its own KV representation while coordinating with common serving infrastructure. It supports 15 methods and introduces Chain Cache and Prefix-Cache Pruning for cross-request state management. The authors report higher throughput, faster decoding than vLLM, and end-to-end speedups on agent benchmarks while maintaining method quality.
In briefArgus: Real-EKS Study of When Predicting Spot Interruptions Beats Simple Checkpointing
SummaryThe paper presents Argus, a Kubernetes operator for handling AWS EC2 Spot instance reclaims, and compares interruption prediction with simple periodic checkpointing on a CIFAR-10 testbed. On real EKS, Argus survived a Spot drain via a graceful SIGTERM checkpoint and resumed from epoch 8. In an 80-trial benchmark, predictive scheduling drove wasted compute to zero, but an oversized lead caused excessive migration, while periodic checkpointing served as a strong ML-free baseline.
In briefAgent-based tool-augmented reasoning for explainable image forgery detection
SummaryThis paper proposes ATAR, a framework that combines 22 specialized forensic tools across seven domains to detect, localize, and explain image forgeries through multi-turn reasoning. It uses a dual-stream approach covering high-level semantic anomalies and low-level forgery artifacts, trained with a forensics curriculum learning strategy. The authors report 78.5% average image-level F1 across six zero-shot IMDL benchmarks, outperforming the strongest MLLM baseline by 11.8 points, while remaining competitive on other tasks.
WhyProposes multi-turn forensic tool reasoning with curriculum learning, reaching 78.5% average zero-shot F1 and beating the strongest MLLM baseline by 11.8 points.
In briefA 3GPP-compliant benchmark dataset for RIS-aided beyond 5G networks
SummaryThis paper presents a large-scale dataset compliant with 3GPP TR 38.901 for research on reconfigurable intelligent surface (RIS)-aided millimeter wave networks, spanning 20 controlled scenario variants. Each sample includes optimal RIS phase configurations found through brute-force codebook search, along with channel state information, channel decomposition, phase matrices, and CQI labels. The authors also propose a CSI-to-CQI mapping as the primary benchmark task, framing link-quality prediction as scalar classification.
In briefThe Missing Coefficients: Bayesian Pairwise Merging for Personalizing Models
SummaryThis paper proposes Bayesian Pairwise Merging (BPM), which converts a user's pairwise output choices into the coefficients needed to merge reward-specialized experts. It treats unknown reward weights as latent variables, infers their posterior from choices and reward-score differences, and uses the posterior mean as merge coefficients. Experiments on radiology summarization, image captioning, and story generation show personalization without per-user policy training, along with calibrated uncertainty about plausible trade-offs.
In briefTSMD: Modality dropout for video highlight detection under missing input streams
SummaryThe paper addresses video highlight detection when visual, audio, or text streams are partly or fully missing. It proposes Temporal-Stream Modality Dropout (TSMD), which simulates structured missingness and trains with a joint loss combining pointwise MSE, per-video Pearson correlation, and peak-oriented RankNet terms. Reported results on the MoSu and Mr. HiSum datasets show TSMD variants outperforming TripleSumm under temporal and complete-stream removal conditions.
In briefCriteria Extraction vs. Chain-of-Thought for LLM Depression Severity Rating
SummaryThis paper compares two ways an LLM can estimate depression severity from social media posts: rating severity directly with chain-of-thought, or marking clinical criteria that a simple rule converts into a label. The authors test three models on two Reddit corpora using PHQ-9 and BDI-II, measuring agreement with quadratic weighted kappa. Criteria extraction shows no significant gain, and in nearly all comparisons it misses more severe posts.
In briefStructure-aware reinforcement learning for guiding protein evolution in the lab
SummaryThis paper presents StructEvo, a reinforcement learning framework that uses protein structure to guide directed evolution. It approximates mutant structure features through feature differences and splits the large mutation space into smaller subspaces with a hierarchical action network. The authors report gains of 9.2% and 16.3% over prior state-of-the-art methods on two benchmarks, and an experimentally validated epistasis pattern in GFP.
In briefBadAction: Backdoor Attacks on Interactive Video Generation Using Action-Guided Triggers
SummaryThis paper presents the first systematic study of backdoor attacks against the interactivity of interactive video generation (IVG) models. The proposed method, BadAction, embeds predefined motion patterns into action sequences and links them to a static target video, so triggered models produce frames that stop responding to later user actions. The authors also test multimodal triggers that poison action, text, and image inputs together.
WhyFirst systematic backdoor study on interactive video generation, achieving 91.0% success with action triggers and evading existing detectors.
In briefRSIGame: Autonomous Game Development with Recursive Self-Improvement
SummaryThis paper presents RSIGame, an agentic framework in which language models autonomously develop games and recursively improve them through local explore-diagnose-improve loops and a global quality-tracking loop. Across 140 GameCraft-Bench tasks, two game engines, and five generators, it reports consistent quality gains under matched development budgets. Training the generator to internalize successful development experience lets Qwen3.8-27B score 61.38 on Godot and 58.53 on Phaser, above GPT-5.5 one-shot scores, while using 11 times fewer generation tokens.
In briefWhether Validators in a Live AI Agent Can Serve as Deployment Gates
SummaryThe paper screens 13 validators in a deployed generative agent to see whether their firing separates builds that work for users from those that do not. It evaluates them against 550 runtime and 350 static builds, reporting separation J with confidence intervals and exact tests. Only two checks survive multiple-comparison correction, and many others are indistinguishable from zero. Skipped probes were almost entirely on broken builds, and because skips are recorded as passes, the study finds a detection ceiling of about 84%.
In briefCycle-aware autoencoder for detecting abnormal railway door behavior
SummaryThis paper treats each complete door opening, dwell, and closing cycle on a train as one unit for unsupervised anomaly detection, training only on normal cycles. The proposed TCAA-CS model combines dual-stream encoding of physical and logical signals with temporal attention, and scores anomalies using reconstruction error, latent deviation, and cross-signal consistency. On real data from a passenger train in commercial service, it reports 93.8% recall, 97.3% precision, and a 0.5% false-alarm rate, and its feasibility for real-time onboard use was assessed on an NVIDIA Jetson AGX Xavier.
Ammar Bouketta, Smail Niar, Hamza Ouarnoughi et al.
In briefSwitching Linear Attention: A sequence layer that boosts linear attention's expressivity while keeping a fixed-size state
SummaryThe paper introduces Switching Linear Attention (SwiLA), a new sequence layer. Standard softmax attention needs a key-value cache that grows with sequence length, while linear attention has constant memory but weaker expressivity. SwiLA keeps the fixed-size recurrent state of linear attention while improving representational capacity, deriving its update from a test-time regression framework as online expectation-maximization in a mixture of linear regressions. On associative recall, in-context language learning, and language modeling benchmarks, SwiLA performs strongly, narrowing the gap to softmax attention and even surpassing it in several settings.
Hyun Dong Lee, Xavier Gonzalez, Nicolas Zucchet et al.
In briefTED: Text-Axis Evidence Decomposition for Prompt-Based Anomaly Localization
SummaryCLIP is a vision-language model not designed for fine-grained defect localization, so CLIP-based anomaly detectors adapt it with prompts or lightweight modules. The paper finds that under domain shift, such models often flag both true defects and visually complex normal regions as anomalous. It proposes TED, a post-hoc scoring method that checks whether ambiguous responses are better supported by source defect patches or by normal patches mistaken as anomalous, without changing the backbone or prompts and without target-domain training. Reported results show improved pixel-level localization, with the largest gains under stronger hard-normal competition.
In briefEarLog: Passively sensing ear physiology during headphone playback for personalized hearing risk awareness
SummaryHeadphone users may suffer irreversible hearing damage during high-load listening, while current protection relies mainly on sound-pressure limits. This paper presents EarLog, which uses playback audio as an opportunistic probe and applies content-conditioned residual sensing with a sparse FIR model to detect the weak acoustic reflex during everyday listening. Evaluated with 81 participants, 16 headphone models, and over 400 hours of recordings, it reaches 99.0% sensitivity and 98.7% specificity on a clinically referenced benchmark. It records listener-specific acoustic reflex events, providing data for future longitudinal hearing-risk studies.
WhyRepurposes playback as an opportunistic probe with residual sensing and sparse FIR modeling to detect acoustic reflex; validated on 81 participants and 400+ hours, with moderate impact for personalize
In briefMaking AI Reviewers Robust to Wording: Benchmarking and the SciCore Method
SummaryThe paper argues that AI reviewers may score manuscripts differently based only on wording, rewarding rhetorical polish over scientific substance. The authors define rhetorical robustness and introduce RobustReview, a benchmark of 1,260 manuscript versions used to evaluate 30 reviewer configurations. They also propose SciCore, a dual-branch reviewer that averages full-manuscript judgment with judgment on a structured science core, which in their GPT-5.5 comparison showed a leading stability-discrimination profile.
In briefSearch Shapes Conclusions: Auditing Evidence Selection Bias in Deep Research Agents
SummaryDeep Research agents write cited reports, but a well-cited report can still reach a misleading conclusion because the documents read are a selective sample shaped by adaptive search. The paper proposes Causal Evidence Selection Correction (CESS), which corrects the candidate-pool average of evidence direction using logged selection probabilities. Tests on the MS2 systematic-review benchmark and public Open Deep Research trajectories report reduced estimation error and sensitivity to document ranking. The authors also argue that estimating a pool average differs from measuring the effect of a search policy, which requires intervention.
WhyFormulates deep research evidence selection as adaptive sampling with causal correction, validated on MS2 and 4,800 paired-intervention trajectories, a notable advance.
In briefA rank-based metric for measuring fairness in algorithmic decisions by examining the distribution of prediction errors across individuals.
SummaryThis paper proposes a rank-based framework for assessing fairness in algorithmic decisions affecting individuals, such as credit scoring, by examining the distribution of model prediction errors. The framework includes Rank Graduation Fairness (RGF), its integrated measure AURGF, a centered Cramér–von Mises permutation test, and a feature removal procedure for explainability. Applied to HMDA mortgage data across four models, it produces model rankings that differ from classical fairness criteria, with tree-based models performing best on accuracy and rank-based fairness. The fairness null hypothesis is rejected for all four models.
