In briefPCB-MC: Analyzing Missing Components on Circuit Boards
SummaryThis paper introduces PCB-MC, a dataset for detecting components missing from printed circuit boards, built on RF100 with footprint-level annotations across 197 board types. The authors benchmark several supervised and unsupervised methods and propose board-type-aware cross-validation to prevent layout leakage. Supervised models show high false negative rates on unseen board designs, while unsupervised anomaly detection fails without spatial alignment to a board-specific reference. The authors conclude that detecting missing components across diverse layouts remains an open challenge.
Betsy Villa Brochero, Ian Gibson, Estefania Talavera
In briefAwakening of the Buddha: Subspace Learning During Population-Loss Plateaus explains how networks learn during flat-loss phases.
SummaryThe paper studies how a neural network can learn a more predictive representation while its population loss stays nearly constant. It gives theoretical conditions for two-layer ReLU and leaky-ReLU networks on Gaussian inputs, plus results for SwiGLU networks and population-moment experiments. The main result is that during a high-loss plateau, alignment between the teacher subspace and the leading AGOP eigenspace increases and refit error decreases, and training later drops below all plateau values.
In briefUsing language models to infer causal links between two event sequences
SummaryThis paper studies how to infer causal relations when only two sequences of events are available, rather than many observations or interventional data. The authors leverage the predictive power of Large Language Models (LLMs) to infer causality between the two sequences. The approach is validated on both synthetic and real data. It reportedly outperforms standard causal discovery algorithms on several time series datasets.
Nishchal Prasad, Eric Gaussier, Emilie Devijver et al.
SummaryThis paper studies whether task vectors from on-policy distillation (OPD) can be composed through model merging, as RL task vectors are. Across five domains and two architectures, merging OPD and RL vectors can outperform both constituent models, and OPD cross-task compositions score higher on average in seven of eight comparisons. Parameter analyses show substantial non-collinearity between OPD and RL updates.
In briefLocal-Entropy Credit Assignment for RLVR via Proximal Entropy Policy Optimization
SummaryThis paper addresses value-model-free RLVR methods like GRPO, which give every token in a rollout the same advantage despite unequal contributions. The authors propose proximal entropy, a token importance measure computed relative to neighboring tokens, and prove it is invariant to prompt difficulty and positional trends. The resulting PEPO method weights per-token advantages and outperforms GRPO and entropy-based baselines on mathematical reasoning with Qwen3-1.7B, Qwen3-4B, and Llama-3.2-3B-Instruct.
In briefOptimal Inventory Control with Censored Demand Using Biased Sample-Average Approximation
SummaryThis paper studies data-driven multi-period lost-sales inventory control, where stockouts reveal only that demand exceeded the stocking level. The authors propose a model-based framework for policy learning from censored data, built on a new cost decomposition for base-stock policies. They introduce a coverage condition and design upper-biased and lower-biased SAA algorithms, achieving near-optimal sample complexity offline and near-optimal regret online.
WhyProposes a new cost decomposition and coverage condition, with upper- and lower-biased SAA achieving near-optimal offline sample complexity and online regret.
In briefCan Earlier Problem Computation Help LLMs Solve New Ones?
SummaryThis paper examines whether keeping earlier problems in a conversation helps large language models answer later, independent questions. The authors find that retained history can either raise or lower accuracy on later turns, even within the same domain. They propose STAIR, which stores keys and values from earlier response generation in a fixed bank and trains current queries to read from it, while the base model stays frozen and only 12,288 parameters are trained. Across three Qwen models and four benchmarks, STAIR improves average later-turn accuracy by up to 11.67 percentage points over the unmodified model with history.
In briefFAER: An Auditable Trajectory Replay Method That Aligns Selection with Learning Utility for Language Model Training
SummaryThe paper observes that replay selectors often rank cached trajectories by format feedback, confidence, freshness, or length, which does not necessarily match the downstream learner's utility. It introduces FAER, an auditable full-trajectory replay framework with a training-free fixed selector and a learner-aware selector, FAER-UTILITY, fitted on disjoint calibration blocks. On GSM8K with Qwen2.5-1.5B-Instruct, FAER-UTILITY reaches a quality score of 0.6624 at about 62,844 target-run tokens, compared with 0.6329 for the fixed selector under the matched study.
In briefExperimental Experience Modeling for Autonomous Research: Reusing Past Experiments to Decide Which Ones Are Worth Running
SummaryThe paper introduces Experimental Experience Modeling (EEM), a framework that helps autonomous research agents decide which candidate directions merit costly full-scale experiments. EEM extracts decision-relevant records from past experimental trajectories, stores them in a reusable experience library, and runs targeted low-cost pilot experiments when historical evidence is insufficient. Experiments on autonomous research benchmarks show that EEM improves research performance while reducing model interaction overhead.
In briefDecoupled and Distilled: Task-Adaptive LoRA Teachers for Few-Shot Class-Incremental Learning
SummaryThis paper proposes TALON, a framework for few-shot class-incremental learning that learns new classes from very few samples while retaining knowledge of earlier ones. It assigns an independent LoRA teacher to each incremental task and distills these frozen teachers into a single unified LoRA student, weighting teacher contributions by feature-space similarity. The authors report comparable or better mean average accuracy across four benchmarks, with up to 33x fewer deployment parameters and a 41.70% reduction in average inference time per task relative to ASP.
Hongwei Zhao (School of Computer Science, Engineering, Beihang University) et al.
In briefUniWAM Technical Report: One Model for Navigating and Manipulating with Robots
SummaryMobile manipulation requires a robot to navigate to a suitable pose and then interact reliably with objects, two stages with different action spaces and visual needs. The paper proposes UniWAM, a unified mixed-stream world-action model with separate action encoders and output heads for navigation and manipulation, sharing one backbone, and introduces Manipulation Anchor Pose (MAP) supervision built from an automated pipeline over large-scale 3D scenes. On MAP-Bench, UniWAM outperforms the strongest external baselines by 30.1% in position error and 44.0% in heading error. Across 24 real-robot tasks, it reports leading results in MAP navigation and mobile manipulation, with competitive manipulation performance.
In briefBeyond Simulation: Retain-and-Repair Neural Operators for Real-World Adaptation
SummaryThis paper introduces R²NO, a framework for adapting neural operators pretrained on numerical simulations to real-world data while keeping useful pretrained structure. The pretrained operator is finetuned on real data and frozen to give a source prediction, while a shared repair module learns a sequence of refinements from the same observations. Using orthogonal Fourier projections, the method fits ridge regressions in cells of the Fourier domain and combines them with weights fitted on a held-out real-data split. The authors report that R²NO consistently outperforms full finetuning and iterative refinement across all RealPDEBench systems and six backbones.
In briefRIQE: A NIQE-Style Quality Model for CT Images
SummaryThe paper presents RIQE, a radiology image quality model that follows NIQE's statistical-distance approach but is fitted on 3,792 full-dose CT slices from 158 patients. It ranks simulated reduced-dose and noisier reconstructions as worse than their source images. However, its preferences among denoising filters do not reliably follow lesion signal, so the authors conclude it should not be the sole criterion for selecting or tuning denoisers.
In briefWhen, Not How Much: Testing Time-Series Foundation Models on Sparse Events
SummaryThis paper evaluates pretrained time-series foundation models on sparse-event data, asking which future periods contain activity rather than forecasting values. The authors rank positions within windows containing events and zeros across five datasets. Released point forecasts add little over simple training-free references, while lightweight event heads trained on frozen backbones rank events better and often outperform gradient-boosted trees.
In briefTTLab at Daleel 2026: STAR-Ar, a Sequence Tagging System for Arabic Argument Recognition
SummaryThis paper presents STAR-Ar, a BERT-BiLSTM-CRF model for detecting and classifying argumentative discourse units in Arabic, submitted to the Daleel 2026 shared task. It frames argument identification and classification as a joint token-level sequence labeling problem. The system scores an F1 of 72.69 on validation and 73.7 on test data. Models trained only on editorials perform worse than those trained on debates, which the authors mainly attribute to the smaller editorial dataset.
In briefFrom Search to Signal: Online Post-Training for Automatic Heuristic Design
SummaryThis paper studies LLM-based automatic heuristic design, where models iteratively propose and refine heuristics paired with executable code. The authors frame online post-training of small open-weight models as building context-dependent learning signals from program validity, task performance, and generation context. Controlled experiments use shared evaluated rollouts and matched update budgets to compare alternative signal mappings. Checkpoint, frozen-search, and resource-matched evaluations aim to separate gains from the model update itself from those due to accumulated search state.
In briefSkillFM: Generating Skills for LLM Agents via Latent Flow Matching
SummaryThis paper presents SkillFM, a generative framework that directly synthesizes task-conditioned textual skills for LLM agents without test-time skill retrieval. It encodes skills into a continuous latent space and trains a conditional flow model with improved MeanFlow, then uses an LLM-based decoder to turn sampled latents into guidance for a frozen agent. On ALFWorld and Search-QA, it achieves the best overall performance among compared vector-based skill approaches.
In briefEvoGen-Harness: Learning where and how to evolve image-generation harnesses without changing the generator
SummaryThis paper presents EvoGen-Harness, a generator-agnostic framework that evolves multiple external responsibilities, such as prompts, routing, and workflows, around frozen text-to-image generators. It also introduces Trace, which aggregates evidence across stochastic runs, uses failure attribution to guide candidate updates, and re-attributes residual failures. The authors report gains over the strongest baselines on GenEval2, T2I-CompBench++, and WISE.
In briefInfoAgent: Traceable, Evidence-Grounded Infographic Generation and Repair
SummaryThis paper presents InfoAgent, a training-free framework that generates and repairs infographics grounded in evidence. It uses a typed dependency graph to record factual content, evidence sources, and verification obligations, so repairs can target only the affected parts. On IGenBench it reports 93.0 Q-ACC and 59.0 I-ACC, and on the authors' InfoGraphicBench-Evidence the complete-checklist pass rate rises from 21.5% to 28.5%.
In briefEgoTools: Tool-Centric Reasoning in Real-World Egocentric Videos
SummaryThe paper introduces EgoTools, a suite for understanding tool use in first-person videos, comprising 100 hours of tool-centric recordings and a 1,000-question diagnostic benchmark. Gemini-3.1-Pro reaches 66.9% overall accuracy but only 51.7% on Perception & Grounding. Fine-tuning Qwen3-VL-8B-Instruct on the data raises its accuracy from 50.0% to 60.9% under strict source-video separation.
In briefGrasping Objects by Past Event Role, Using Active Viewpoint Selection
SummaryThe paper presents BeyondSCe, a zero-shot robotic grasping system that handles requests referring to objects by their role in past observed human-object interactions. When the target is occluded, the system combines an event prior from the history with current scene geometry to choose camera viewpoints likely to reveal it. It uses pretrained models without task-specific training and reports improved grasp success rates in real-robot experiments.
In briefWavelet Flow Matching for Multivariate Time-Series Generation
SummaryThis paper studies generating multivariate time series with flow matching in the wavelet domain, so that both multi-scale temporal structure and cross-channel dependencies are reproduced. The model operates on multilevel discrete wavelet coefficients, representing coarse structure and finer details at separate scales, paired with a channel-token transformer that models cross-channel attention. Across seven benchmark datasets and four sequence lengths, the method is best or tied on most dataset-metric combinations, with the largest and most consistent gains in Context-FID and discriminative score.
Lucas Poinsignon, Jorge da Silva Goncalves, Samuel Ruiperez-Campillo et al.
In briefEHR-RobustGym: A benchmark for training clinical AI agents to handle noisy health records
SummaryThis paper introduces EHR-RobustGym, an interactive environment built on MIMIC-IV hospital records for evaluating and training clinical agents that work with noisy electronic health records. It contains 5,486 Clean-Noise question pairs covering several noise types and clinical intents. Evaluated LLMs showed average task success falling from 62.2% on clean questions to 37.9% on noisy ones, and supervised fine-tuning and reinforcement learning improved performance, with gains extending to five external benchmarks.
In briefSimple parameter merging to transfer specialist skills into general language models
SummaryThis paper asks whether specialist models' task abilities can be moved into a general language model directly at the parameter level, without training or distillation. It applies two existing training-free heterogeneous merging methods, Intersection-Merge and Activate-Prune-Merge, to specialist-to-general transfer. Both methods improved the general recipient model across embedding, reranking, reward modeling, and MoE code-specialist transfer experiments.
In briefA fairer grid beam search for enforcing word constraints in text generation
SummaryThis paper studies lexical-constrained decoding for autoregressive text generation. It finds that grid beam search tends to satisfy easier constraints first and leave harder ones to the end, unlike DFA-constrained beam search. The authors propose fair grid beam search, which removes this bias while still needing only linearly many forward passes. Experiments on two constrained generation tasks confirm the bias and show the fair variant also finds higher-probability strings.
In briefA dynamical theory of how LoRA behaves in continual learning and forgetting
SummaryThe paper gives an asymptotically exact dynamical analysis of LoRA in a solvable two-task teacher-student model, deriving ODEs for macroscopic order parameters that describe generalization error. It finds that low-rank updates reduce interference with first-task features, but LoRA's initialization slows adaptation to the second task. A state-dependent masking strategy that freezes the strongest first-task hidden units and restricts adaptation to the complementary subspace markedly reduces forgetting, and the theory is qualitatively reproduced on sequential MNIST.
Th\'eo Marchetta, Filippo Alessandroni, Alessandro Breccia et al.
In briefCOBICount: Counting Objects in Satellite Images Without Training on Target Data
SummaryThis paper studies counting objects such as buildings, vehicles, and ships in overhead images using a model trained only on one source dataset, without target data for training. The proposed COBICount method separates generating candidate responses, accepting compact object-centered ones, and suppressing responses from repeated background structures like roads or parking grids. Trained on RSOC Building and tested directly on DOTA Large Vehicle, Small Vehicle, and Ship, it achieves the lowest average mean absolute error among compared methods, 174.132.
In briefReady2Blend: Turning Natural-Language Instructions into Blendable Alignment Prompts
SummaryThis paper introduces Ready2Blend, a method for continually aligning large language models to new requirements without forgetting earlier behaviors. Its AlignFormer component maps each requirement to a fixed-length alignment prompt stored in a modular prompt bank, keeping the backbone and prior prompts frozen. In two practical settings, it reaches 93.1% to 98.5% of a joint-training reference while using only a few prompt tokens and up to 4.3 times less training time.
In briefRethinking how multimodal models reorganize evidence across multiple images
SummaryThis paper studies how multimodal large language models organise visual evidence spread across several images. The authors introduce Mosaic, a general-purpose harness with ten composable image operations, and compare five re-representation settings on existing benchmarks and a new benchmark, MosaicBench. They find that visual re-representation helps most on tasks needing precise visual evidence, while semantic-heavy tasks show smaller gains. They also train MosaicAgent-8B with reinforcement learning using only accuracy and format rewards.
In briefLampAttention: Mixed-precision attention for dedicated AI chips that stays accurate by recomputing only sensitive parts
SummaryThis paper proposes a hardware-algorithm co-design that adds mixed precision to FlashAttention. It computes query-key products and their exponentials in 8-bit formats, then identifies sensitive sub-blocks and recomputes them in 16-bit. The authors also specify a dedicated accelerator designed to run this pipeline efficiently. Simulations with Qwen3 and Gemma 3 indicate that sending only a small minority of sub-blocks to high precision is enough to recover baseline model performance.
Stanislav Budzinskiy, Marian Gloser, Tolunay Yilmaz et al.
In briefAutoresearch for Mixed-Integer Linear and Nonlinear Programming
SummaryThis paper introduces AutoMIP, a reusable agent skill that organizes long-horizon automated research on mixed-integer programming problems. It uses a persistent idea pool and an algorithm tree to manage experiments and switch between research directions. On MIPLib and MINLPLib benchmarks, AutoMIP finds new best solutions for 31 of 60 and 52 of 60 instances, respectively, and ablation studies show both components contribute.
In briefWorking Around the Compute Ceiling: A Memory Layer That Makes LLM Reading a One-Time Cost
SummaryThe paper notes that LLM serving is stateless, so a second question about the same document recomputes its attention state from scratch. The authors present Galahad, a memory layer for vLLM, SGLang and llama.cpp, where Taliesin saves and reloads key-value state for repeated text and Blaise passes only the needed section to the model. In a recall test over a 97,000-token corpus, the combined system answered 100 of 100 questions at about 0.6 seconds each.
WhyPersisting byte-exact KV state to avoid recomputation is a clear advance; experiments span several runtimes but the scale and baselines are limited.
In briefMaking Asynchronous SGD Robust via Soft Throttling
SummaryAsynchronous SGD is a distributed learning method that applies each client's gradient as soon as it arrives, which speeds training but makes it vulnerable to attacks from fast clients. This paper introduces Throttle, a Byzantine-robust generalization that exponentially down-weights updates from faster clients by a factor q. The authors analyze its convergence rate and validate its robustness theoretically and empirically. Experiments also show improved performance over standard asynchronous SGD even without Byzantine attacks.
In briefComplexity analysis of preference-based bandit learning
SummaryThis paper studies preference-based bandits with general reward function classes, where a learner picks pairs of arms and observes binary preference feedback under the Bradley–Terry model. The authors introduce a new complexity measure, the locally sensitive eluder dimension, and propose the GINOP algorithm. They establish a first-order regret bound and argue that learning from preference feedback is as statistically efficient as learning from direct rewards, supported by empirical comparisons with baselines.
Ahmed Ben Yahmed (CREST, ENSAE Paris, FAIRPLAY) et al.
In briefHAPMoE: Automatic Parallelism Planning for Mixture-of-Experts Training on Heterogeneous Clusters
SummaryThis paper presents HAPMoE, an automatic parallelism planner for training mixture-of-experts (MoE) models that accounts for hardware heterogeneity. It uses a lightweight MoE-aware cost model and searches a six-dimensional parallel space to produce plans deployable on Megatron-LM. Experiments report up to 3.2× higher end-to-end training throughput than baselines on heterogeneous clusters, an additional gain of up to 78% from non-uniform pipeline partitioning, and a search completing within 1 minute.
In briefReusing offline LLM feedback to improve small language model reasoning without online calls
SummaryLarge language models reason well but are costly to access, while small models are easier to deploy yet weaker at reasoning. This paper proposes Reusable Latent Correction (RLC), which uses a black-box LLM only offline and converts its natural-language guidance into corrective experiences stored in the hidden space of a fixed small model. These experiences are kept in an external bank and retrieved according to the small model's current reasoning state. Experiments across several reasoning benchmarks and model scales show consistent gains without parameter updates or online LLM calls.
In briefWho spoke and whether it is remembered: evaluating persistent speaker identity across meetings
SummaryThe paper proposes SI-cpWER, a metric that scores meeting transcripts under one global speaker-ID assignment to test whether speaker identities persist across recordings. It benchmarks five commercial diarize-then-identify systems, two open academic baselines, and the authors' end-to-end system ThyVoice on the CHiME-8 NOTSOFAR 129-meeting set and CHiME-6. ThyVoice achieves lower SI-cpWER than every commercial system in all three conditions, with a full-panel mean of 47.13 versus 54.75 for the next system.
In briefGolden Path Hypothesis: Reusable Caching Schedules for Diffusion Models
SummaryThis paper proposes the Golden Path Hypothesis, which holds that under fixed inference conditions, prompt-independent cache schedules can reach final-output quality close to the best prompt-specific schedules. The authors test this across ten caching methods, four image and video models, and three cache ratios. They also analyze denoising trajectories and cumulative caching errors to explain why such schedules transfer across prompts and datasets.
In briefBelief-based maximum occupancy principle and how it compares to active inference
SummaryThe paper extends the Maximum Occupancy Principle to partially observable environments and derives a Bellman form of Expected Free Energy for Active Inference, with belief over hidden states included in the agent state. This allows offline value iteration over the belief-state space. In minimal experiments with uncertain food sources, MOP agents switch between food seeking and exploring different sources, while Active Inference agents mostly stay near one source.
In briefUnderstanding as No-Arbitrage: Bounded Dutch Books for Language Models
SummaryThe paper defines a language model's understanding as resistance to bounded Dutch books, meaning a computationally limited trader cannot guarantee profit by betting against its probabilities on logically related claims. It introduces Arbitr, a training framework with an adversarial trader and a calibration anchor, and reports five pre-registered experiments on Qwen2.5 and Phi-3.5. Arbitr reduces exploitability by orders of magnitude without losing task accuracy, though the authors conclude coherence is necessary but not sufficient for knowledge.
In briefElectrolyteFM: A Unified Model Predicting Multiple Electrolyte Properties by Sharing Knowledge Across Them
SummaryElectrolyte formulation design must balance several physicochemical properties, but existing models often cover only a few. The paper proposes ElectrolyteFM, a unified multi-property model that learns each property's representation separately and uses an expert pool and router to share useful cross-property knowledge through residual adapters. On Electrolyte12, it reduces normalized mean absolute error averaged over 12 properties by 14.8% relative to the strongest electrolyte-specific baseline. On an unseen sodium-electrolyte dataset, it reduces conductivity mean absolute error by 6.7% relative to the best baseline.
In briefContraSafe-VL: A contrastive safety preference dataset for vision-language AI models
SummaryVision-language models are increasingly used as general multimodal assistants, and their safety depends on how they interpret both text and images. Existing safety datasets mainly teach models how to respond, giving limited guidance on how visual context should shape safety judgments. This paper introduces ContraSafe-VL, which pairs visual contexts with the same text query to produce multimodal queries with different safety labels, along with response preferences based on harmlessness and helpfulness. Experiments across three model architectures show better cross-modal safety recognition and a improved balance between safety and over-refusal, while preserving general capabilities. The dataset will be publicly released.
In briefLearning Beyond Full Imitation: Task-Preserving Knowledge Distillation
SummaryThis paper argues that a student that matches its teacher's predicted probabilities more closely does not always perform better, since it may already separate the correct class more sharply than the teacher. The authors prove an exact separation between full imitation and conditional learning, and propose task-preserving knowledge distillation (TPKD), which keeps the label gradient intact and minimally corrects the conditional gradient. Experiments on CIFAR-100 and CLINC150 report gains of 0.47 and 0.35 percentage points over standard distillation.
In briefWinoTS: Wavelet-Based Self-Distillation Pre-Training for Time Series Models
SummaryThe paper proposes WinoTS, a self-supervised pre-training method for time series that uses invariance-based self-distillation with wavelet time-frequency augmentations. It reports better results than state-of-the-art baselines in long-term forecasting, cross-domain zero-shot transfer, and unsupervised anomaly detection. Ablations indicate the framework is architecture-agnostic across time series backbones.
In briefHow many samples are enough for learning across domains?
SummaryThis paper studies how many data samples each training domain needs for cross-domain learning to work well. Based on the learning bounds it presents, the authors set criteria for per-domain sample sufficiency. They report an inverse linear relationship between the number of training domains and the samples required per domain.
In briefTexTailor: Texture-Preserving Video Try-On with Adaptive Garment Conditioning
SummaryThis paper presents TexTailor, a framework for high-fidelity video virtual try-on built on a pretrained video Diffusion Transformer. It adds a timestep-adaptive modulation of garment features, frame-aligned positional encoding for garment-to-video correspondence, and a multi-source injection design to reduce interference between conditions. The authors report competitive results on several benchmarks, including the high-resolution Eevee dataset, for garment detail preservation, temporal consistency, and video quality.
In briefA serving system that speeds up speculative search for LLM test-time scaling
SummaryTest-time scaling improves LLM reasoning on hard tasks like math and coding by spending extra computation during inference. The paper identifies two challenges for speculative execution: an explosion of candidate reasoning paths and frequent fine-grained verification. It proposes SpecScale, a serving system that uses early pruning of low-quality paths, deduplication of computation across redundant paths, and deferred fine-grained verification. On the MATH and Olympiad benchmarks, the authors report higher throughput and lower latency than non-speculative and prior speculative approaches while preserving answer quality.
In briefNarrativeSteward: An Authoring Tool for Guiding AI Agents in Interactive Story Creation
SummaryNarrativeSteward is an authoring environment that links outlines, worldbuilding, and narrative graphs so AI agents can implement them and authors can guide the work. It offers agent dialogue, project-wide structural review, change records, and execution verification to help authors understand and assess revisions. In a 12-participant within-subject study, it made revision requests and change inspection easier and improved perceived understanding of changes and story structure compared with general-purpose agents. The code is open-sourced on GitHub.
In briefPatchKV: Compensating KV Cache Compression in Weight Space
SummaryThis paper addresses the linearly growing memory of the key-value cache in long-context LLM inference, which existing compression methods degrade at aggressive budgets. It proposes PatchKV, a training-free framework that pairs a compressed KV cache with a context-specific weight patch computed in closed form via ridge regression. Across long-context QA and math benchmarks on three model architectures, PatchKV consistently improves cache compression methods.
In briefCausal discovery from count data using discrete score matching
SummaryThis paper addresses causal discovery for count data, where standard score-matching methods lack derivatives. It proposes conditional curvature score (CCS) and off-diagonal curvature score (OCS) to recover directed acyclic graphs (DAGs), and introduces DISCO, an algorithm that estimates these scores using discrete diffusion. Experiments report accurate DAG recovery across Poisson, negative binomial, binomial, and mixed-family settings, scaling to 1,000-node DAGs on a single GPU.
In briefWorkGenesis: Building Work Environments to Train AI Agents
SummaryThis paper introduces WorkGenesis, a framework that builds executable occupational tasks from real-world public files, guided by O*NET occupational knowledge. It also uses execution feedback to iteratively repair tasks and rubrics until they pass an audit. A 35B model, Fx-Work-35B, trained by simple supervised fine-tuning on 20K synthesized tasks, reportedly outperforms comparable-scale baselines on three benchmarks and surpasses DeepSeek-V4-Pro-Preview.
In briefMotionWeave: motion-centered future prediction to improve robot action policies
SummaryVision-language-action models have started using world models to learn richer dynamics than sparse action labels, but predicting future images can capture appearance details irrelevant to control. The paper proposes MotionWeave, which uses an Action-Induced Motion Grounder to localize interaction regions per future timestep and a Horizon Residual Composer to inject temporal motion cues into action tokens. On six MetaWorld tasks it reaches a 75.3% average success rate, 8.6 points above π0 (66.7%), with especially strong gains on sustained-interaction tasks.
In briefHiWE: Hierarchical World Knowledge Model for Zero-Shot 3D Robot Path Planning
SummaryHiWE is a framework for generating robot demonstrations that links visual grounding and language planning through a point-based interface. PointVLM associates task-relevant objects with image coordinates, and depth data lifts these into a 3D semantic representation used by the 3DLLM planner to specify end-effector waypoints and gripper commands. A hybrid grasping module determines local grasp poses. The approach is evaluated on 14 simulated manipulation tasks and four physical-robot tasks, with ablation studies.
In briefTraining dynamics of GRPO for small language models
SummaryThis paper systematically studies Group Relative Policy Optimization (GRPO) fine-tuning on small language models with 1.5B to 7B parameters, using a single-node 8xA100 setup. It covers mathematics, coding, and science multiple-choice questions, examining how group size affects convergence, stability, and benchmark performance, along with LoRA target module choices. The GRPO-tuned models beat their base versions on about 80% of math benchmarks but show limited results on multiple-choice and code reasoning. The authors then refined their LoRA and reward-shaping settings to improve those domains.
Rajat Ghosh, Vaishnavi Bhargava, Henry Wong et al.
In briefOmniMed-Jev: Making Medical AI Confidence Scores More Trustworthy
SummaryOmniMed-Jev represents each medical decision, such as a diagnosis, finding, or cell count, as a full probability distribution over a candidate set rather than as generated text. Compared with a generative baseline trained on the same backbone, data, and schedule, its reported probabilities track observed correctness more closely, reducing calibration error by up to an order of magnitude. Counting tasks remain an exception where the generative baseline performs better. The authors state these results do not indicate clinical readiness, since the interface cannot be fully separated from training differences.
In briefRethinking generative image compression at extremely low bitrates: RAE-CoD keeps images recognizable
SummaryThe paper studies how generative image compression behaves as bitrate approaches zero, finding that existing codecs suffer semantic collapse and produce malformed or unrecognizable content. The authors propose RAE-CoD, a compression-oriented diffusion model built in a representation autoencoder space that directly aligns compressed and source representations. It preserves recognizable, naturally structured content for 256×256 images with as few as 16 bits. On MSCOCO-30K at 0.001–0.008 bpp, it reduces VFM feature MSE and Fréchet Distance ratio relative to the best competitors.
In briefFusing time-frequency EEG images to predict rTMS depression treatment outcomes
SummaryThe paper proposes montage and blending fusion methods and a lightweight CNN applied to EEG-derived time-frequency images to predict response to rTMS for depression. Segment-level 10-fold cross-validation reaches up to 99.90% accuracy, but under strict subject-disjoint validation most configurations fall to AUC 0.31–0.54. The authors conclude that segment-level results overstate performance.
Wael Korani, Md Fahimul Kabir Chowdhury, Mohammed Aledhari et al.
In briefReSAIL: Preventing Performance Collapse in Iterative Agent Self-Distillation
SummaryThe paper finds that iterative self-distillation with privileged information causes LLM agents to lose deployment performance across training cycles. It proposes ReSAIL, which selects interaction steps where privileged information most changes the teacher's predictions and regularizes the student's outputs toward the frozen teacher. On ALFWorld and TextCraft over three cycles, adding ReSAIL to self-distillation baselines raised final-cycle success rates by an average of 22.5% absolute.
In briefAutomatic harness evolution for visual-interface robot agents
SummaryThe paper studies using an off-the-shelf coding agent as a robot policy that observes a browser-based 3D interface via screenshots and controls a virtual gripper with tools. A separate optimizer agent automatically revises the agent's harness of prompts, tools, and control rules. The authors find that larger rollout batches per round improve generalization to held-out tasks, and that a Champion-Challenger rule accepting only improvements raises LIBERO held-out success from 51% to 67%. Distilling the combined system into a standalone policy reaches 62.2% on RoboCasa365, and the evolved harness on top reaches 71.2% on the leaderboard benchmark.
In briefGeneralized Geometry Block Proximal Linearized Method for Multiblock Nonconvex and Nonsmooth Optimization
SummaryThis paper studies a class of multiblock nonconvex and nonsmooth optimization problems. The authors propose a generalized geometry proximal linearized operator that builds block surrogate functions from arbitrary inner products and general admissible metrics, and develop the GGBPL method and its inertial version iGGBPL. They establish a unified convergence framework, proving objective value convergence, global convergence to a critical point, a convergence rate, and an O(ε⁻²) iteration complexity for ε-stationary points. Numerical results on sparse nonnegative matrix factorization and sparse nonnegative CP decomposition with ℓ0 constraints show better performance than several state-of-the-art methods.
In briefBMASH: Detecting Soccer Headers Using Ball Motion Cues
SummaryThis paper addresses soccer header spotting, which identifies moments in broadcast video when the ball touches a player's head. The authors first adapt Video Swin as an action-recognition baseline, then propose BMASH, a fusion framework that adds ball presence and motion features from a frame-level ball detector. BMASH improves clip-level AP and ROC-AUC over the Video Swin baseline, and in continuous full-video spotting it reaches comparable event-level F1 with a different precision-recall trade-off.
Ahmed Endris Hasen, Muhammad Shahzad Khan, Nikolaos Passalis et al.
In briefTree navigation without LLM summaries for long-document QA retrieval
SummaryThis paper studies hierarchical retrieval for long-document question answering. It introduces NavTree, which builds a deterministic balanced segment tree over text chunks with zero language-model calls at indexing, using the tree only to navigate and returning leaf chunks. In a matched-cost evaluation, NavTree significantly beats BM25 on multi-hop QA and outperforms a summary-based RAPTOR variant at every multi-chunk budget.
In briefMiniRep: Reputation-Based Aggregation That Resists Malicious Agents in Multi-Agent Debate
SummaryThis paper studies how reputation can be used when multiple AI language-model agents debate the same question and combine their answers. The authors propose MiniRep, which scores agents using both their behavior on the current task and their reputation over time, and stops groups of highly similar responses from dominating the final output. Tested on the MATH dataset, MiniRep reportedly beats conventional debate aggregation and reputation-based methods with or without attacks, and outperforms all baselines in all 28 attack conditions in a heterogeneous 10-agent setting.
In briefSemantic-aware joint source-channel optimization for video transmission that works with any encoder
SummaryThe paper proposes SAJSCO, a lightweight plug-in scheme that jointly tunes source compression and channel coding for digital video transmission based on inter-frame semantic importance and channel state. It uses a multi-actor proximal policy optimization algorithm to adapt the rates, and the learned policy works across different video encoders without retraining. Reported results show BD-rate reductions of 34.86% with H.265 and 18.01% with DCVC-RT.
In briefANI: Adaptive Numerical Injection Combining Semantic and Arithmetic Representations for Numerical Reasoning
SummaryThis paper proposes ANI (Adaptive Numerical Injection), a framework to improve numerical reasoning in large language models. Text tokenization often splits numbers into fragments, which hinders arithmetic, while numerical embeddings ignore the semantic role numbers play in context. ANI uses a context-aware gating mechanism to selectively inject numerical embeddings (specifically FoNE) into the latent space, preserving identifiers while enhancing quantitative operands. The authors report a 9.5-point gain on the MATH benchmark over the official reference model across various LLMs, with robust performance on general language benchmarks.
In briefPhysics-Informed Group Data Handling: Building Function Representations Adaptively for Navier-Stokes Equations
SummaryThis paper proposes PI-GMDH, a method that progressively builds functional representations of coupled physical fields during solving rather than fixing their structure in advance. It uses tensor-product Chebyshev functions and is tested on the 2D time-dependent Taylor-Green benchmark for incompressible Navier-Stokes equations. In the tested setup, the adaptive variant reached validation and test losses of about 5.3e-19. Comparisons with other construction policies and selected PINN and KAN configurations are presented, but the authors state this is an illustrative result on a controlled synthetic benchmark, not evidence of universal superiority.
In briefSTCFormer: Adaptive Transformer for Station-Based Weather Forecasting
SummaryThis paper proposes STCFormer, an adaptive spatio-temporal Transformer that dynamically groups weather stations based on their local evolution within each temporal patch. Its Cluster-Guided Attention Block combines local attention within clusters with global attention over regional state summaries. Tested on eight temperature and wind forecasting tasks across three real-world weather datasets, it achieves the lowest 24-hour mean squared error on all eight tasks.
In briefEngramBench: A Benchmark for Testing Whether AI Agents Truly Transfer Skills
SummaryThis paper introduces EngramBench, a benchmark for evaluating how autonomous coding agents evolve and reuse skills, designed to separate genuine capability transfer from copying of past solutions. It includes 30 learning tasks and 13 unseen transfer tasks, tested across 48 multi-hour execution trajectories with human-expert validation. The authors report that static skill banks do not solve exact code implementation limits, but they guide agents away from costly trial-and-error and cut coding time by over 55%.
In briefA Missing Latent, Not a Missing Simulator: Radius-Augmented Inference for Real JWST Retrieval
SummaryThis paper finds that amortized simulation-based inference trained on radiative-transfer simulators works on synthetic JWST spectra but fails on real WASP-39b spectra. The authors rule out forward-model physics as the cause and attribute the failure to a missing latent variable, the planet radius. They introduce MIRAGE, a radius-augmented flow-matching posterior, which yields a physical, literature-consistent retrieval of WASP-39b and transfers across two instruments and three real JWST targets.
In briefA Tilted Bowl Is Not a Slippery Slope: Compressing Looped Models
SummaryThis paper studies why compressed looped models, which reuse the same weight block many times, often collapse. Testing more than 30 models from five families, the authors find that rounding error accumulates only in loops that never settle, while in settling loops it merely shifts the resting point. Based on this, they propose a controller that stops when the halting head fires and finishes with 8-bit loops, outperforming fixed-depth inference on Sudoku-Extreme and Maze-Hard.
Steven Kolawole, Pearse Jim, Opegbemi M. Busoye et al.
In briefReTaCo: Residual-Target Control for Efficient Student Training from Teacher Feedback
SummaryThis paper studies on-policy distillation, where a student model learns from a teacher's token-level feedback on its own outputs while keeping teacher communication costs low. The authors show that the prior EOPD method, which renormalizes the teacher's top-k probabilities, keeps pushing top-k mass toward one, so the teacher is not a stationary point. They propose ReTaCo, which keeps the top-k tokens and groups the rest into one residual symbol, with a single-sample estimator whose expectation matches the full-vocabulary reverse KL. Numerical optimization supports the theory, and ReTaCo outperforms EOPD on most mathematics and code benchmarks across three teacher-student pairs.
In briefMegaAvatar: Controllable Talking Avatar Video Generation
SummaryThis report presents MegaAvatar, a framework for generating controllable talking avatars, built on the Wan2.2-TI2V-5B model. Beyond audio and reference-image conditioning, it adds SMPL-X-derived 3D guidance to control body pose and head motion, and extra audio and face cross-attention modules for expression control and identity preservation. It also includes an audio-to-SMPL-X model, so audio-driven generation works without user-provided SMPL-X frames.
In briefParallel island-model genetic algorithm for 3D trailer loading
SummaryThis paper addresses the Three-Dimensional Trailer Loading Problem, which seeks the placement and orientation of heterogeneous items in a trailer to maximize volume use while meeting logistical and safety constraints. The problem is NP-hard, so exact methods are impractical at industrial scale. The authors propose PANGEA, an enhanced Biased Random-Key Genetic Algorithm accelerated by an island-based parallel framework that promotes diversity and reduces computation time. The method was validated in a real trailer loading process.
In briefRW-Flow: One-step generative modeling on compact manifolds via Wasserstein gradient flows
SummaryThis paper presents RW-Flow, a framework for learning one-step generative models on compact manifolds, such as geospatial event locations and protein and RNA torsion angles, using Wasserstein gradient flows. It establishes a necessary and sufficient condition for identifiability on compact, connected Riemannian manifolds, showing that for symmetric, Lipschitz-continuous costs the Sinkhorn-induced velocity field is identifiable exactly when the Gibbs kernel is nondegenerate. The authors note that the squared geodesic distance does not always guarantee identifiability. Across benchmarks under fair comparison conditions, RW-Flow outperforms existing one-step methods in nearly all settings.
WhyEstablishes necessary and sufficient identifiability for Sinkhorn-based one-step manifold generation; multiple manifold benchmarks support it, though impact is mainly within this subfield.
Ualibyek Nurgulan, Seungwoo Yoo, Prin Phunyaphibarn et al.
In briefTime-aware receptive-field bag method for interpretable irregular time series classification
SummaryThis paper addresses classifying irregular time series, which have uneven sampling intervals, missing observations, and variable lengths. It extends the Bag-Of-Receptive-Fields (BORF) transform to irregular data using a time-weighted normalization that weights each observation by its time delta, along with a sliding-window recurrence that keeps linear time complexity. Benchmarks on PYRREGULAR datasets show competitive classification performance and human-interpretable explanations.
In briefPLRS-IC: A Dual-Calibration Method for Aligning Chest X-Ray Images with Text
SummaryThis paper addresses two sources of ambiguity when aligning chest X-ray images with text: frontal and lateral views look different for the same finding, and contrastive training wrongly treats different patients with the same condition as negatives. The proposed PLRS-IC framework uses a low-rank residual similarity (PLRS) to adapt patch-text matching to each projection view, and an information-content-calibrated soft false-negative suppression (IC-SFNS) to soften penalties on semantically overlapping negatives. Experiments on nine public zero-shot benchmark settings report consistent gains in classification, grounding, and segmentation.
In briefUniversal Attack Method Against SAM3 Concept Segmentation
SummaryThis paper studies the robustness of SAM3, a segmentation model extended to concept-level prediction, against adversarial attacks. The authors propose AdvPCS, which combines min-max prompt optimization, a global-local perception deception attack, and a temporal transition deviation attack. A single universal perturbation generated by the method generalizes across frames from different videos and across point, box, and text prompts. Experiments on four benchmarks show it reduces the average mIoU on the SA-CO dataset to below 5% under text prompts.
In briefConcept subspaces beyond the logit lens: a weights-only test for locating representations upstream of output readout
SummaryThe paper proposes a geometric diagnostic that uses only model weights to measure how much an extracted concept subspace overlaps with the dominant right-singular directions of the unembedding matrix. It tests a ten-dimensional subspace called FARS across 26 models and nine estimators, finding low readout energy compared with a same-layer next-token control. A complementary intervention study on four models reports model-dependent effects that remain well below full-vector replacement.
Aojie Yuan, Zhiyuan Julian Su, Haiyue Zhang et al.
In briefA First Look at Jev for Network Traffic Classification: Accuracy, Processing Time, and Cost
SummaryThis paper evaluates Jev, a general-purpose decision model, on classifying ten application labels in the CESNET-QUICEXT-25 dataset using only the first ten packets' sizes, directions, and inter-packet times. Adding 40 labeled examples raised Jev's accuracy from 9.80% to 28.42%, still below Random Forest and Extra Trees trained on 8,000 records. On a paired 100-record subset, the study finds no established accuracy advantage for either Jev or GPT-5.6 Sol, and timing differences reflect different service configurations.
In briefEffective Does Not Mean Useful: Conditional Functional Substitutability for Redundancy and Scaling in Transformers
SummaryThe paper proposes Conditional Functional Substitutability (CFS), which directly measures when intermediate computational states in neural networks are functionally redundant because they produce similar downstream responses. Unlike importance or similarity-based measures, CFS captures functional relations and reduction potential. Across modalities and Transformer families, CFS shows systematic functional reorganization with scale, and controlled experiments indicate performance gains need not track substitutability growth.
In briefFrom Benchmarks to Production: Transferring Anomaly Detection for Electricity Output Monitoring
SummaryThis paper tackles anomaly detection in EDF's short-term electricity production forecasts, looking for atypical intra-day patterns that may signal data quality or operational problems. It introduces TAMIS, a scalable and interpretable system that flags anomalous days by measuring deviations from historical patterns learned from past data. Top anomalies are surfaced in a daily automated newsletter to support expert review. Experiments on real industrial data show TAMIS achieves the best accuracy-efficiency trade-off against baselines, and the anonymized datasets are publicly released.
In briefSelecting clients and training data to speed up synchronized federated learning
SummaryThis paper addresses how slow clients and non-iid data across clients delay convergence of a global model in federated learning. The authors show that prior probabilistic client selection methods fail under non-iid data, especially when few clients are present. They propose a joint client and training data selection algorithm for faster convergence. Experiments on CIFAR-100 report faster convergence and higher model accuracy than prior works that handle non-iid data and heterogeneous computation.
In briefTrust the Critic More: Action-chunked actor-critic for LLM reinforcement learning
SummaryThis paper introduces AC2 (Actor-Critic with Action Chunking), a reinforcement learning method for language models that avoids rolling every trajectory to its terminal reward. It assigns credit to short action chunks, and a learned critic scores the state at the end of each chunk. Trained on Qwen3-4B with FineProofs-RL and evaluated on IMO-ProofBench, AC2 is reported to exceed GRPO's peak validation score of 18.5% while using 2.5x fewer decoding FLOPs.
In briefRight answer, wrong mechanism: detecting harmful divergence in causal interventions
SummaryThis paper examines how causal interventions such as activation patching and distributed alignment search can yield correct answers through hidden pathways that a model never uses on natural inputs. The authors plant such pathways in GPT-2 small and test over 100,000 interventions. Existing distance-based detectors perform below chance, while the proposed Hidden-Pathway Contribution test works well when pathways show up as out-of-regime unit activity but fails when all units stay within their natural range. Optimised interventions tend to exploit hidden pathways, and an on-manifold penalty substantially reduces this reliance at some cost to success rate.
WhyPlanting hidden pathways in pretrained models makes wrong-mechanism interventions testable, with a label-free detector and extensive experiments on GPT-2.
In briefA framework for making zigzag persistent homology differentiable using linear interpolation
SummaryPersistent homology can be differentiated and used in learning, but no comparable method existed for zigzag persistence, where topological structure changes non-monotonically over time. The authors build such a framework for sequences of simplicial complexes formed by thresholding time-dependent filtering values, assigning diagram endpoints the times at which linearly interpolated values cross the threshold. Under mild regularity conditions, they derive differentials almost everywhere and prove local Lipschitz continuity outside an explicit measure-zero set. The method is tested empirically on sensor network coverage optimization and dynamic graph classification.
Enrico Maria Ferrari, Clemens Bannwart, Matteo Biagetti
In brief4MT-VLM: How Coarse Is a VLM's Cognitive Map?
SummaryThis paper introduces 4MT-VLM, a dataset of procedurally generated landscapes rendered in five stimulus modes that remove appearance cues while keeping layout fixed. Sixteen open and closed-source models were tested using a 4AFC measure, the same used to grade humans. Models recognize places from the studied viewpoint but drop below chance at 135° rotation, while a human scores 85%. The authors conclude current VLMs have rudimentary cognitive maps but spatial resolution is too coarse to keep a stable 3D understanding after viewpoint changes.
In briefLimits of Robot World Models: When Predictions Remain Useful for Planning
SummaryThis paper studies action-conditioned world models built on five frozen self-supervised vision backbones, evaluated on Meta-World manipulation tasks and BridgeData V2 real-robot interactions. The models reliably guide action selection only when goals lie within or slightly beyond the imagined trajectory. Even with perfect prediction, success drops as the target moves further away. Within its plannable range, the world model can also raise a vision-language-action policy's success rate from 65% to 77%.
In briefWhat Streaming Anomaly Detection Finds and Misses in Industrial Time Series
SummaryThis paper examines how streaming anomaly detection performs on industrial sensor data, motivated by EDF's need to spot faults in its power plants as they occur. Since no single streaming method is universally best in unsupervised settings, the authors compare streaming methods with state-of-the-art time series anomaly detection models run online on a real nuclear power plant dataset. It also evaluates automated anomaly detection in a streaming context. Results show online anomaly detection models are more consistent, and ensembling strategies provide strong robustness.
Magali Parrino, Antoine Ajenjo, Emmanuel Remy et al.
In briefRAIM: Robust Aggregation of Cheap Models for Hallucination Detection
SummaryThis paper examines whether a panel of inexpensive open-weight judge models (4–9B parameters) can be aggregated to replace a frontier model when evaluating the faithfulness of generated text. The authors propose RAIM, which combines a cross-fitted stacked logistic regression with an admissibility test to decide when aggregation is worthwhile. Tested with ten judges across eight benchmarks, the panel retains a median 93% of Claude Sonnet's Cohen's κ, loses 2.9 points of balanced accuracy on average, and costs about one sixty-fourth as much to run.
In briefFYAN: A Human-AI Framework with Semantic Auditing for Formalizing Documents
SummaryThe paper presents FYAN, a human-AI harness for document-level mathematical formalization that covers specification, proof planning, logical review, Lean proof construction, and validation. Its central component is evidence-grounded semantic auditing, which checks whether formal statements faithfully preserve their informal specifications. Using the same model in every stage, FYAN proves 86 of 143 FormalTCS theorems under a strict Lean check, compared with 69 for a general agent harness. On ConsistencyCheck, its semantic audit detects more inconsistent statements than a direct LLM judge.
In briefLearning multi-view geometry from multiple views via self-distillation
SummaryThis paper proposes Poincar3, a self-supervised method that learns visual representations from multiple views through self-distillation rather than RGB reconstruction. It combines masked patch and image-level distillation, with a teacher that observes additional views, enabling training from scratch without explicit 3D supervision. The authors report that Poincar3 outperforms DINOv3, MuM, and Muskie on correspondence estimation, camera pose estimation, and 3D reconstruction.
David Nordstr\"om, Thibaut Loiseau, Vincent Lepetit et al.
In briefPredicting argument structure in online conversations by comparing modeling methods and task designs
SummaryThis paper studies how to automatically predict complete argument structures from dialogue, identifying argumentative units and the relations between them. Under strict schema constraints, the authors compare supervised fine-tuning with prompt-based large language models, using single-step and multi-step task architectures. They benchmark these on three dialogical corpora adapted from Inference Anchoring Theory. The results show that identifying argumentative relations is the main bottleneck, largely because dialogical argumentation is implicit and context-dependent.
Siddharth Bhargava, Sara Tonelli, Patricia Mart\'in-Rodilla et al.
In briefQATFactory: a framework for training LLMs to hold up under low-precision deployment quantization
SummaryQATFactory is an open-source framework for quantization-aware distillation and reinforcement learning of large language models. It simulates deployment-time quantization while running matrix multiplications in BF16, supporting NVFP4, MXFP4, and llama.cpp's Q4_K formats, and exports checkpoints directly to vLLM and llama.cpp. On Qwen3.5-9B, the method reports average accuracies of 68.9% under NVFP4 and 66.0% under MXFP4, above the best post-training quantization results.
In briefDC-SAE: A Deep Compression Autoencoder to Speed Up Diffusion Training
SummaryThis paper proposes DC-SAE, a decoupled semantic autoencoder for high-compression image generation that aims to speed up diffusion model convergence. It pairs a semantic encoder, which enables higher compression ratios, with a pixel-level encoder that preserves details for faithful reconstruction. On ImageNet at 512×512, it reports 32× spatial compression with 29.79 PSNR and 3.37 gFID, and also shows text-to-image results at 1024×1024.
In briefLearning process rewards from final outcomes by propagating reasoning states
SummaryThis paper introduces Reasoning State Propagation (RSP), which labels each reasoning prefix as valid or invalid and models transitions between successive states. By propagating these transitions, process annotations supervise intermediate states while outcome labels supervise the final state and pass learning signals back to earlier steps. Across reasoning search, response selection, and reinforcement learning, RSP outperforms representative process reward model baselines, improving over Qwen2.5-Math-PRM by 5.6% in beam search and 2.1% in reinforcement learning on average.
In briefIn a Streaming World, Should You Stand Still? A Benchmark Comparing Streaming and Static Time Series Anomaly Detection
SummaryThis paper compares streaming and static time series anomaly detection methods in a unified streaming benchmark, the first large-scale experimental study of its kind. The authors also introduce TSB-drift, a distribution-drift dataset of real time series designed to isolate settings where incremental updates are theoretically justified. The results show that, contrary to common assumptions, static methods significantly outperform streaming approaches in most streaming settings.
Magali Parrino, Antoine Ajenjo, Emmanuel Remy et al.
In briefOptimal-rate additive regression when the input distribution is unknown and dependent
SummaryThe paper studies additive regression where the covariate distribution is unknown, possibly non-product, and the number of covariates grows with sample size. Under joint-density bounds and dimension-growth conditions, it proves matching minimax upper and lower bounds for prediction. The classical additive rate holds with known marginals and, when marginals are unknown, whenever the densities are at least as smooth as the weighted components; otherwise marginal smoothness sets the rate. All additive components are also recovered with squared error of the same order as prediction error.
In briefConsensus and factual dynamics in populations of interacting language models
SummaryThe paper introduces RHEON, a physics-inspired framework that treats a population of language model samples from one frozen model as a spin system evolving across different interaction topologies, with sampling temperature as the source of disorder. Across 432 configurations it generates Eraclitus-4.7M, a corpus of 4.7 million tagged responses. Consensus gains are strongest within the first few update sweeps, and more neighbours per agent speed convergence on average. However, whether consensus is factually correct or hallucinated cannot be predicted from initial state alone, and the optimal low-hallucination temperature depends on coupling. Semantic agreement correlates with factual convergence, but unanimity does not certify correctness.
Emanuele Ricco, Elia Onofri, Vincenzo Sammartino et al.
In briefUniAE-MoE: A Unified Audio Encoder Using Mixture of Experts
SummaryThis paper presents UniAE-MoE, a unified audio encoder for large audio language models that uses a Mixture-of-Experts architecture to cover speech, music, and general audio. It integrates encoders from Qwen2-Audio and Audio-Flamingo 3, adds SwiGLU with shared experts, and applies two-stage instruction tuning plus task-specific data scaling. The authors report a score of 0.802 on the XARES-LLM benchmark, state-of-the-art results, and top-tier performance in the Interspeech 2026 Audio Encoder Capability Challenge.
In briefUruqi: Learning spatial understanding from continuous visual experience (paper)
SummaryThis paper addresses vision-language models' difficulty with tracking their own motion and mapping surroundings while moving, by training on synthetic continuous visual experiences. The authors synthesized 11,738 camera trajectories over diverse 3D scenes and evaluated on the Uruqi benchmark of 52k questions across 2.7k episodes. They report accuracy rising from 15.84% to 47.73% for URUQI-Syn-8B, and a 17.13% average relative gain over its InternVL3-8B backbone on three external spatial benchmarks.
