In briefForging LLM Authorship Fingerprints via Targeted Rewriting
SummaryThis paper studies whether rewriting one language model's output can make attribution classifiers assign it to a chosen target model. The authors propose ForgePrint, which searches for rewrites that shift attribution toward a target fingerprint and distils them into a one-pass 4B student model. On CNN/DM summarization, the student reaches 70.2% target success, above its teacher (54.1%) and the best of six published baselines (39.3%). The authors conclude that fingerprint detectability should not be equated with source authenticity, since text-only attribution can be misleading under targeted rewriting.
In briefCombining diverse reasoning paths to improve LLM answers
SummaryMulti-path methods like self-consistency sample several reasoning paths from a language model and pick the most frequent answer, but their gains level off as more paths are added. The paper treats this as a diversity combining problem from wireless communications, showing that correlation between path correctness caps the effective sample size of the vote. It also proposes an Adaptive-K rule that uses a four-path pilot to choose the number of paths, keeping 96 to 103 percent of the accuracy of 32-path majority vote across Math, QA, and NLU benchmarks.
In briefMore Choices, Fewer Decisions: Ordinal-Scale Bias in Direct-Decision Models
SummaryThis paper studies a systematic tendency in JEV-like direct-decision models to underuse ordinal rating scales, concentrating final decisions on fewer labels than the gold distribution. Across 40 datasets and nine models, the bias is weaker on unordered semantic categories and intensifies with finer scales. The authors report that targeted training can effectively mitigate it, making scale utilization a reliability target alongside accuracy.
In briefDecoMoE: Faster Multimodal MoE Inference by Separating Visual Propagation and Expert Computation
SummaryMultimodal mixture-of-experts models are costly to run because long visual-token sequences repeatedly require attention, routing, and expert computation. DecoMoE compresses both dimensions: a Sample-Adaptive Visual Boundary predicts an input-dependent layer after which visual tokens are removed, and a Routing-Calibrated Expert Prefix keeps the shortest contiguous set of experts covering a target share of routed text-token mass. On Qwen3-VL-MoE, it retains 97.91% of dense-baseline performance while cutting computation from 27.06 to 16.73 TFLOPs and latency from 0.44 to 0.26 seconds, a 1.69x speedup.
In briefSkillSeek: Rethinking Agent Skill Retrieval at Marketplace Scale
SummaryThe paper presents SkillSeek, an open-source two-stage retriever for selecting LLM agent skills (SKILL.md directories), built from a standard IR recipe of a BGE-base bi-encoder and a small cross-encoder. On the 89-task SkillsBench benchmark, it reaches observed parity with an LLM-mediated query-refinement loop at much lower cost. Per-trial spend drops from USD 51.30 to USD 27.54.
In briefBARRAC: Adapting an English Aspect-Based Sentiment Analysis Method for Arabic Dialect Classification
SummaryThis paper asks whether methods built for majority languages like English can be adapted to Arabic tasks. The authors adapt an English aspect-based sentiment analysis framework into BARRAC, which swaps consumer-review attribute pools for Arabic linguistic devices and dialect markers, and replaces noisy self-training with two-stage training. On five Arabic dialect datasets it reaches a mean macro-F1 of 63.93%, beating the best few-label state of the art by 3% and GPT-4o on four of five tasks. The paper also provides error analysis of remaining challenges.
Ali Almutairi, Gelareh Mohammadi, Imran Razzak et al.
In briefStudy finds future video generation aligns better with human visual cortex than observed video.
SummaryThis paper compares how video diffusion models' internal representations match human visual cortex fMRI responses while people watch videos. It finds that representations used to generate future frames align better with the visual cortex than those of the observed video. Alignment with observed video reconstruction concentrates in lower-order visual areas, while future generation aligns with higher-order areas. A human behavioral experiment also found people preferred videos produced by amplifying the layers that aligned better with the cortex.
In briefWhose Voice Survives the Summary? An Audit of LLM Employee Feedback Summaries
SummaryThe paper proposes a Voice Retention / Representation Ratio metric to measure representational bias in LLM summaries of employee feedback. Applied to 2,586 English and German free-text responses from a global professional service company, it finds employees voice criticism far more often than praise, and the summarization pipeline filters by how frequently a theme appears rather than by sentiment. A concern raised only once was dropped 86% of the time, and short and German-only content was lost more often. A targeted prompt only recovered themes that were explicitly named.
In briefWhen Reasoning Goes Astray: How Attention Patterns Reveal Uncontrolled Looping in Reasoning Models
SummaryThis paper studies large reasoning models, which can fall into redundant verification and endless generation loops during long reasoning. The authors introduce RADAR, a method that identifies the current reasoning state in real time from attention patterns. They also propose Attention Realignment, which corrects abnormal attention distributions, reducing looping while largely preserving performance on normal tasks.
In briefYou're Hired: Strategic Model Selection for Multi-LLM Collaboration
SummaryThis paper studies how to choose which large language models join a multi-LLM collaboration system. The authors propose and systematically evaluate nine selection algorithms, including approaches based on model description diversity, capability-aware behavioral diversity, and LLM-based recruiters. Experiments on candidate pools of 10 and 32 models across math, coding, QA, and reasoning tasks show that successful selection algorithms outperform random or top-individual-performance heuristics by up to 36.1% in some settings.
In briefLearning Chaos Without Seeing Chaos: Autoregressive Transformers Extrapolating Global Dynamics
SummaryThe authors train small autoregressive transformers from scratch on trajectories from restricted parameter ranges of several nonlinear dynamical systems, including the logistic and sine maps, the Lorenz system, and the generalized Hopf system. Under closed-loop evaluation at parameters far outside the training distribution, the models reproduce self-similar period-doubling cascades, chaotic dynamics, and attractor structures. For the logistic map, the transformer reproduces period doublings up to period 128, giving a finite-order scaling ratio of 4.6687 that matches the Feigenbaum constant within 5×10⁻⁴.
In briefCan Terminal Agents Trust Their Own Verification? Diagnosing and Improving Self-Verification
SummaryThis paper examines how reliable self-verification is for terminal agents that solve tasks in command-line environments. Using a diagnostic framework with objective ground truth, the authors find that only 61.43% of incorrect candidate solutions are detected and 49.36% of detected errors are repaired. They propose Student-Conditioned Verification Distillation (SCVD), which improves Pass@1 on TerminalBench2.1 across three Qwen3.5 backbones.
In briefDCM-SAM: a method for segmenting metal additive manufacturing defects that runs on NPUs, using defect-specific LoRA expert mixtures.
SummaryMetal additive manufacturing parts are inspected with X-ray CT, where labelled data is scarce and key pores are only a few pixels wide. The paper proposes DCM-SAM, which adds a separate Conv-LoRA expert bank and mask decoder per defect class to a frozen Segment Anything backbone, trained only on synthetic slices while updating 4.4% of parameters. It reports 64.2% pore IoU on real NIST scans and describes making the model run in FP16 at 1024x1024 on a Qualcomm Hexagon NPU.
Md Mushfiqur Rahaman, Md Mahedi Hasan, Imtiaz Ahmed et al.
In briefCRAFT: Causal Tracing of Failure Modes in Medical Vision-Language Models
SummaryThis paper examines how medical vision-language models internally resolve conflicts between visual evidence and text, identifying two failure modes: text overriding correct image-based diagnosis, and committing to answers without sufficient evidence. The authors find these are driven by spatially separate attention head groups and introduce CRAFT, a method that localizes and causally verifies these heads. Experiments across several medical VQA benchmarks and model architectures show that targeted interventions reduce each failure without retraining.
WhySeparates arbitration and brake failures into spatially disjoint attention heads, validated by causal interventions across multiple medical VQA benchmarks and architectures.
In briefStateTree: Reinforcement Learning for Better Long-Term Dialogue Reasoning
SummaryThis paper proposes StateTree, a reinforcement learning method that trains language models to reason over long, multi-session dialogue histories. It builds a tree-structured path-tracing task from key-value records spread across sessions, using curriculum training with increasing tree depth. The authors report that models trained on 10K-token contexts generalize to 128K tokens, and StateTree-14B reaches 59.00% on LongMemEval, above QwenLong-L1-32B at 45.20%.
In briefBlackboard intelligence: a diffusion-based search that beats autoregressive models on globally constrained tasks
SummaryThis paper argues that next-token autoregressive inference struggles on problems with complex global constraints. It proposes blackboard intelligence, where a model searches and revises candidate solutions on a fixed canvas, implemented with diffusion language models. Mean confidence from the masked diffusion objective serves as a proxy for global coherence to guide search and revision. Across ZebraLogic, Nurse Rostering, and Job-Shop Scheduling, the method outperforms same-scale autoregressive baselines and, on some tasks, frontier LLMs.
WhyProposes blackboard-style global search over diffusion LMs using mean masked confidence to guide revision, with solid comparisons across three constrained tasks against autoregressive baselines.
In briefExplicit diversity in agent trajectories for RL post-training of LLMs
SummaryThis paper studies explicit trajectory diversity in reinforcement learning post-training of LLM agents. It defines diversity using user-specified, task-specific descriptors that map each trajectory to an interpretable behavioral representation, then measures set-level diversity over these descriptors. The authors propose TJPO, a single-policy framework that optimizes this diversity over sampled trajectory groups. Experiments on Sokoban and ALFWorld report improved task-specific trajectory diversity with competitive task performance.
In briefHow uncontrolled repetition forms inside model layers and a method to curb it
SummaryUncontrolled repetition can prolong text generation in large language models and enable resource consumption attacks. This paper studies, mainly in large vision-language models, how repetition signals arise in shallow layers and spread through the residual stream. The proposed Tokenwise Residual Comparison (TRC) method localizes these anomalies during generation and suppresses them, reducing loop rates by 57% on average. TRC also applies to large language and reasoning models.
In briefOverlap, Unique, and Conflict: Can LLMs Extract What They Can Recognize?
SummaryThis paper introduces Overlap-Unique-Conflict (OUC) extraction, a task that pulls out all overlapping, conflicting, and unique clauses from two narratives. The authors build a benchmark of about 22K narrative pairs and 140K OUC instances across factual, argumentative, and political texts. Testing 14 open-source LLMs, they find unique information is much easier to extract than overlap and conflict, and a fine-tuned Qwen-3-8B outperforms much larger models on several tasks, though overlap and conflict remain challenging.
In briefGraphCert: Training graph-reasoning AI agents using verified evidence checklists
SummaryGraphCert is a method for training AI agents that reason over knowledge graphs, reducing the need for manually built question-answer pairs and reasoning traces. It generates graph-based questions with marked supporting evidence, certifies them through execution, and turns the accepted evidence into rubrics that reward both correct answers and evidence alignment during GRPO training. The authors report that it outperforms much larger LLM agents and other post-training methods on five graph reasoning domains in GRBENCH. They also state that the learned policy transfers across different graph domains.
In briefTesting whether confidence estimates stay calibrated as a model's knowledge changes
SummaryAs AI agents continually learn, their confidence estimates must keep pace with changes in the model's knowledge without repeated supervision, which the authors call persistent calibration. They test confidence estimators trained on earlier checkpoints of open models to see whether they generalize to later checkpoints, using contrast sets of questions answered correctly by one checkpoint and incorrectly by another. Both inference-time and fine-tuning methods perform worse on these contrast sets than oracle methods trained on future checkpoints. Training across multiple checkpoints is reported to help.
In briefA method to recover missing training signals for faster speculative decoding
SummaryBlock drafters for speculative decoding are usually trained on text written by other models, so one off-policy token invalidates supervision for all later slots in the block. The paper proposes a rollout-based framework with two parts, Anchor-Label Relabelling (ALR) and In-Rollout Anchors (IRA), to restore full supervision without changing the corpus. On fixed vision-language and text corpora, it raises greedy accepted length by up to 36.5% over DFlash and outperforms erasing baselines. After three epochs, it matches training on target-regenerated responses.
In briefTraining LLM judges from language feedback using position-selective self-distillation
SummaryThis paper studies how to train large language model judges using natural language feedback, especially for subjective tasks. The same model, conditioned on feedback, acts as a teacher providing dense position-level supervision, and the authors identify context sharpening and context spreading through per-position entropy shifts. They propose masking higher-entropy-shift positions, which improves out-of-distribution generalization over naive self-distillation. The resulting judges outperform outcome-supervised RL by 2-9 percentage points on evaluated subjective subcategories while remaining competitive on objective ones.
In briefLEARN-TS: Language-Guided Time-Series Anomaly Detection with Normality References
SummaryThis paper addresses the difficulty of separating true anomalies from benign deviations using reconstruction errors in multivariate time-series anomaly detection. The proposed LEARN-TS method uses a frozen language model to create two semantic representations: window-specific observation semantics without exact numerical values guide masked reconstruction, while a fixed normality prompt aligns normal representations and estimates discrepancy. Across four benchmarks, it achieves the highest mean performance in 13 of 16 dataset-metric comparisons.
In briefPositive Ratings, Hidden Concerns: How Employees Voice Views to AI Listening Tools
SummaryThis paper reports a field study at a global management consulting firm that pairs an AI voice interview with a pre-survey on the same themes. In 44 first-session interviews, 20-41% of sessions showed a favorable rating alongside a later substantive concern, depending on the threshold. The authors find employees disclosed in guarded ways, and this protective behavior tracked how legitimate they perceived the listening structure to be, leading to a grounded model of bounded disclosure.
Thilo Tamme (1), Michael Saatkamp (1), Alma Bonte (1) et al.
In briefSpeech-based acoustic markers for detecting Alzheimer's disease
SummaryThis study examines whether acoustic features from spontaneous speech can reliably tell people with Alzheimer's disease apart from healthy controls. Using speech data from the Cookie Theft picture-description task, the researchers extracted prosodic, spectral, and temporal features and tested several machine learning classifiers. The best models achieved clinically meaningful separation between the diagnostic groups. The authors suggest these results support non-invasive, low-cost audio screening and call for further multimodal and longitudinal research.
In briefSame Loss, Different Gradients: Numerical Mismatch in High-Dimensional von Mises-Fisher Learning
SummaryThe paper shows that in probabilistic learning, the loss computed in the forward pass and the gradient supplied in the backward pass can describe different objects due to finite special-function recurrences, custom backward rules, and numerical clipping. Using high-dimensional von Mises-Fisher learning as the case study, it analyzes the mismatch in finite-start Bessel recurrences. It introduces AR/FR, a fixed-depth method that constructs a potential and its derivative jointly to ensure forward-backward coherence, with a uniform cubic-order error bound relative to the exact Bessel ratio.
WhyIdentifies forward-backward mismatch in von Mises-Fisher learning and proposes AR/FR, which jointly constructs potential and derivative with a uniform cubic error bound.
In briefHow Accurate Is Accurate Enough? A paper on how much numerical error learning systems can tolerate
SummaryThis paper asks how accurate a numerical approximation must be inside a learning system. It argues that identical error magnitudes can have very different effects on losses, predictions, and gradients depending on the learning state. For softmax cross-entropy, the authors analyze how class weights couple with errors and derive finite-error guarantees and certified numerical tolerances. Experiments in von Mises-Fisher learning show these tolerances vary by orders of magnitude across states.
WhyFrames numerical error impact via the learning objective, derives exact softmax cross-entropy loss extrema, and validates certified tolerances across many controlled learning states.
In briefDoes a shared temperature mean a shared angular scale in probabilistic contrastive learning?
SummaryThe paper studies the von Mises-Fisher probabilistic score used by ProCo in contrastive learning and shows that a shared temperature does not imply a shared similarity scale in high dimensions. The authors prove the score has a class-dependent leading angular gain, g_c = A_c/τ, which affects Softmax competition, decision boundaries and gradients. Experiments on CIFAR-LT, ImageNet-LT and iNaturalist support these predictions, and gain-equalized cosine prototype classifiers agree with the original predictions in 98.43–99.99% of 16 frozen settings.
In briefPersona and persuasive framing in AI voice agents: a 2×2 field experiment with children
SummaryThis study reports a randomized field experiment on a public German Santa Claus telephone hotline, routing children's calls to one of four LLM voice agents. The agents varied in persona (high-authority Santa or low-authority Helper) and framing (persuasive nudges toward prosocial wishes or neutral). Across 89 analyzed conversations, persuasive framing raised the probability of a prosocial wish from 11.6% to 45.7%. Persona authority had a near-zero effect on compliance, but children hung up on the Helper within the first minute far more often (65% vs. 39%).
Thilo Tamme (1), David Steck (1), Anton Hantel (2) ((1) Technical University of Munich et al.
In briefChartDensity-Bench: Testing How Well AI Models Recover Chart Numbers in Dense Visuals
SummaryThis paper introduces ChartDensity-Bench, a benchmark for testing multimodal large language models on rebuilding structured numerical data from compound chart figures. It varies how many charts appear at once (1, 3, 6, or 9) to measure how visual density affects performance, and uses a multi-dimensional evaluation covering structure, completeness, parseability, and numerical accuracy. Experiments on five recent models found that reconstruction generally worsens as density rises, though the extent of decline differs considerably across models.
In briefRAST: Transferring High-Resolution Audio Knowledge to Improve Low-Resolution Activity Recognition
SummaryThis paper addresses a mismatch in audio-based human activity recognition, where high-resolution audio is available for training but only low-resolution audio can be used at deployment. The authors propose RAST, a framework that compresses high-resolution teacher representations while preserving token-level information and neighborhood structure, then aligns them locally with low-resolution features. Experiments on the SAMoSA and AudioIMU datasets show RAST outperforms low-resolution-only training and direct teacher-transfer baselines. It improves low-resolution-only recognition by up to approximately 7.8%.
Ji Hwan Park, Gautham Krishna Gudur, Yufei Shen et al.
In briefAction-conditioned bisimulation for memory in GUI agents
SummaryThe paper proposes a rule for deciding when two web-page states are the same in a GUI agent's memory. Two states merge only when their shared actions lead to agreeing outcomes and successor blocks, based on an empirical predictive state graph built while the agent acts, without using observation similarity and without training. On MiniWoB++ the method raises success rate over a memoryless agent, while control variants (identical exploratory detours, the prior successor-representation merge, and the criterion without action conditioning) show no change.
In briefEfficient Training Data Attribution for Diffusion Models via Score Discrepancy Distillation
SummaryThis paper studies which training samples influence a sample generated by a diffusion model. It attributes influence using a local score discrepancy measure, estimated as a preconditioned gradient similarity with Kronecker-factored curvature, yielding the TID method. The authors then distill TID into TIDE, a forward-only student model. On CIFAR-10, ArtBench-10 and MS-COCO, TID matches or outperforms prior methods, and TIDE attributes samples in milliseconds, faster than generation.
In briefObjectTwin: Measuring the Asset Correspondence Gap in Digital Twins
SummaryThis paper introduces ObjectTwin, a protocol for checking whether generated 3D meshes that replace objects in captured indoor scenes reproduce the calibrated observations of the same object. It freezes the instance, cameras, scan-derived support and depth, and anchor, then swaps only the object representation to measure the asset correspondence gap. Across two indoor RGB-D domains and two generation backends, visible anchored assets lose 0.302 to 0.497 MaskIoU, and the generator ranking reverses across domains.
In briefLearning Under Forgetting: How Repeated Reinforcement and Forgetting Select What Neural Networks Retain
SummaryThe paper proposes Repeated Reinforcement with Persistent Forgetting (RPF), a minimal framework in which patterns that recur in training data are reinforced while persistent forgetting attenuates learned information, treating forgetting as a selection mechanism. The authors develop the theory in three layers: a survival law for independent features, spectral filtering in a shared-parameter model, and an implicit trade-off yielding MDL-like compression. Controlled experiments provide evidence for this reinforcement–forgetting selection mechanism.
In briefdattri-LLM: A Unified, Efficient Library for Attributing LLM Training Data
SummaryThis paper presents dattri-LLM, an open-source library for training data attribution (TDA), which estimates how much individual training examples contribute to a large language model's outputs. It improves efficiency with compact gradient representations and cost-model-based routing, and it collects per-example gradients from existing training loops without modification. The authors report an average of 3.2x the throughput of the fastest competing library and show multiple attribution methods scaling to 110B-parameter models on four H200 GPUs.
In briefPathAnchor: Path-Structured Evidence Retrieval for Scientific AI Agents
SummaryThis paper introduces PathAnchor, a bounded scientific reasoning system that organizes evidence as path-structured workspaces. It retrieves source-linked Material-Sensor-Signal-System trajectories that preserve each step's role, direction, and supporting evidence, then produces a claim-cited answer with an explicit evidence boundary. On 120 flexible-sensor questions, PathAnchor scores 82.6% and leads six evaluated systems.
In briefLasting Effects of Abstract Pretraining: Gains Beyond Perplexity
SummaryThis paper studies whether warming up small language models on abstract, algorithmically generated stack-manipulation data before natural language pretraining improves later abilities. Allocating about 1% of pretraining tokens to this data raised multi-hop question answering on MUSIQUE by up to 3.9 F1 points, with comparable language-modeling perplexity. The gains appear mainly on sequential reasoning chains and persist through billions of subsequent language tokens.
Zachary Shinnick, Hemanth Saratchandran, Damien Teney et al.
In briefAdaptive-GEPA: Adapting Execution Harnesses to Mixed Request Types
SummaryThis paper introduces Adaptive-GEPA, which learns both how to route heterogeneous requests and how to solve them under one search budget. It evolves a router and a library of specialist programs, aligning specialists by the requests they handle and inheriting their descriptions and code. On a mixture of four task families, the reported Qwen3-8B run's routing matches the task partition on all 651 test requests, and the family-mean score rises from 52.6 to 70.6.
In briefWhere Do Multi-Agent Systems Fail? Diagnosing Collective Mechanisms from Execution Evidence
SummaryThis paper examines how to tell, from an execution's records, whether a multi-agent system broke a collective mechanism, the rules governing how agents share and use information. The authors propose a diagnostic contract that classifies a violation as supported, ruled out, or unknown, and test it on four mechanisms by replaying executions. Broken mechanisms often still produced correct answers, so a correct outcome does not substitute for records of how the mechanisms operated.
WhyIntroduces a diagnostic contract framework for mechanism violations, validated by replay experiments, though benchmark scope is limited and transfer to another workflow degrades.
In briefHALO: Localized-observation heterogeneous allocation for routing robot fleets
SummaryThis paper proposes HALO, a hybrid method for the Vehicle Routing Problem that splits the task into an allocation step and a routing step so robots can compute onboard solutions in real time. Its allocation phase uses a heterogeneous graph neural network to separately learn spatial distributions and task-to-robot compatibility. On a partially observable online VRP variant, HALO clearly beats a heuristic baseline and nearly matches an all-knowing offline version. On the traditional static single-depot VRP, it outperforms architectures optimized for that setting by up to 14.06% while running fastest.
In briefEvent-driven refresh and recurrence memory to reduce stale grounding in video object segmentation
SummaryThe paper addresses a weakness of Sa2VA, a method that segments objects in videos from language queries: it grounds the target using only a few initial keyframes, which can cause stale grounding and false positives in long or changing videos. The proposed EDRRM re-invokes Sa2VA only at stable change points detected from tracking cues, and uses CLIP similarity to retrieve anchor frames when the target re-appears. Experiments on Ref-DAVIS17, MeViS, and ReVOS report comparable or better J&F scores with fewer refresh calls and fewer false positives.
Abu Hanif Muhammad Syarubany, Jaehyun Jang, Siwoo Lim et al.
In briefSelf-improving algorithm-design agents that escape in-context evolutionary plateaus via population-curated policy optimization
SummaryThis paper studies LLM-based algorithm-design agents, which tend to plateau when pure in-context evolution is used in domains needing specialized knowledge. The authors propose Population-Curated Policy Optimization (PCPO), which keeps and reuses high-quality, diverse self-generated algorithms from a global population to shift the policy toward stronger algorithms. Trained on only 4 chip cases for learning rate schedule design in global placement, PCPO outperforms OpenEvolve and ShinkaEvolve on average across 16 chip cases. The paper also reports reduced inference token cost and an average 8.27x speedup on four GPU kernel designs over PyTorch Eager.
WhyProposes PCPO with an Improvement Chain proposition; experiments on chip placement and GPU kernels with baselines give strong evidence, though impact is mostly within specialized EDA and kernel domain
In briefAgentic camera pose estimation that picks and checks the best estimator with learned ranking
SummaryThis paper introduces PoseAgent, a framework for estimating the relative camera pose between two images. It profiles each image pair, uses a learned ranking to choose which of several pose estimators to run first, and uses a learned verifier to predict pose error and fall back to other estimators when needed. The authors report that PoseAgent improves AUC@5 degree by up to 4.2% over the strongest single estimator on ARKitScenes, MegaDepth, ScanNet++, and RealEstate10K.
In briefWhere Evidence Lives: Auditing AI Companions' Self-Claims
SummaryThe paper argues that an AI companion's claims about itself, such as remembering users or being changed by the relationship, and users' experience ratings can only be checked where the evidence is observable by users, in behavior or in the users themselves. When evidence lies in internal machinery, fluent self-description and moderately positive ratings do not show those mechanisms ran. The authors present an audit procedure comparing an agent's self-descriptions with user judgements and implementation records, labeling each claim as supported, contradicted, or unresolved, and apply it to Lita, a proactive companion deployed for a month with nine colleagues.
In briefHearing the World in Stereo: Learning Dynamic Spatial Correspondence for Immersive Video-Audio Generation
SummaryThis paper proposes StereoBind, a framework that makes generated stereo audio's perceived sound location move consistently with the motion of its visual source. It coordinates visual motion and stereo sound through three mechanisms: Visual Motion Binding, a Spatial Track Encoder, and Residual Track RoPE. The authors also release StereoWorld-29K, a stereo audio-video dataset with motion tracks, and StereoWorldBench for measuring audiovisual spatial consistency. Experiments reportedly show substantially better spatial alignment than existing models while preserving overall audiovisual quality.
In briefConsensus-aware fusion method for detecting camouflaged objects using reference samples
SummaryThis paper addresses reference-guided camouflaged object detection, which segments targets that blend closely into their surroundings using auxiliary reference samples. The authors propose a framework that combines trainable PVTv2 and frozen DINOv3 features, selects foundation-model evidence via reference-conditioned correlation, and aggregates consensus across multiple references. The authors report that extensive experiments demonstrate the method's effectiveness.
In briefMatisse: a training-free method for active 3D reconstruction using evidence-based reasoning
SummaryMatisse is a training-free framework that helps a 3D reconstruction system decide which views to acquire and which keyframes to keep from partial observations under a limited computation budget. It uses evidence from a pretrained generative 3D model to estimate evidential uncertainty and an information gain measure that guides both view acquisition and keyframe selection. The paper reports Chamfer distance reductions of 12.7%, 3.8%, and 9.2% on GSO30, YCB-V, and Replica relative to the best baseline on each dataset, and a 1.50× end-to-end speedup on GSO30. In a keyframe selection experiment on GSO30, it achieved comparable Chamfer distance to Stream3D using 14% of the input views.
Xihang Yu, Kaichen Zhou, Lorenzo Shaikewitz et al.
In briefMolecular property prediction for structurally different compounds using tabular foundation models
SummaryThe paper studies predicting molecular properties for compounds whose structures differ from labeled training molecules. It introduces MolPAIR, which combines molecule-level and molecule-pair contexts, using a frozen tabular foundation model to refine initial predictions by modeling error differences between a query and labeled reference molecules. Across 58 MoleculeACE and Polaris tasks, MolPAIR further improves the predictor on 46 tasks, with gains across four molecular representations and three backbones.
WhyCombines molecule and molecule-pair contexts for tabular in-context learning, improving 46 of 58 tasks across representations and backbones, a solid composition-based advance.
In briefLearning Multi-Step Visual Retrieval for Agents
SummaryLLM agents need retrieval tools to access external knowledge, but standard single-step retrievers limit visual agentic search. The paper introduces VHOP, a data generation framework and benchmark, and VHOP-Router, a trained pipeline that turns a standard embedding model into an autoregressive multi-step retriever. VHOP-Router retrieves linked image chains in one tool call within visual latent space. The authors report retrieval performance rising from under 5% to 76.3%, and agentic task success improving by 52.7%.
WhyOffloads multi-step navigation to an embedding retriever returning image chains in one call; novel idea, but evaluated mainly on a self-built benchmark.
In briefCode to Control: Synthesizing Parameterized Reactive Controllers
SummaryThis paper introduces Code to Control, which uses an LLM to synthesize the structure of Python controllers and derivative-free search to fit their parameters from environment feedback. The learned controllers run directly as policies, requiring no LLM inference or planning at decision time. Across Atari, Flappy Bird, and MuJoCo tasks, the method outperforms planning-based program synthesis, is competitive with deep reinforcement learning using fewer environment interactions, and transfers across substantial changes in dynamics.
Zergham Ahmed, Joshua B. Tenenbaum, Chris Bates et al.
In briefDecentralized Decisions Among Heterogeneous Autonomous Vehicles Using an α-Potential Game Framework
SummaryThis paper studies non-cooperative games among different autonomous vehicles, where each vehicle uses a decentralized policy based on its own state and interacts with others through possibly asymmetric weights. The authors build a single auxiliary α-potential function whose minimizers give approximate Nash equilibria, and they characterize the approximation error α in terms of interaction asymmetry. They also introduce vehicle-specific scaling to reduce asymmetry, and derive social-efficiency guarantees for the selected policies. Numerical experiments cover collision avoidance, lane changing, overtaking, and priority-based intersection crossing.
In briefUniEvo-VL: A self-distillation training method for multimodal models to improve themselves
SummaryThe paper proposes UniEvo-VL, a framework in which one multimodal model acts as both teacher and student: the student sees only the plain question, while the teacher also receives the model's own critique, and training minimizes the divergence between their denoising diffusion distributions. Built on the open-source Qwen-image-2512, experiments report GenEval rising from 0.747 to 0.808 and GenEval2 Soft-TIFA from 32.97 to 35.53. The authors also note that gains are not uniform across tasks, with mixed results on text rendering.
In briefMetaSteer: A Context-Dependent Nonlinear Method for Steering Language Models
SummaryThis paper introduces MetaSteer, a method for steering large language models using nonlinear interventions applied to attention projection matrices, so that their effects vary with input context. It is trained once on a pooled preference corpus and transfers zero-shot to unseen concepts and out-of-distribution contexts. The authors report that MetaSteer matches or outperforms task-specific steering baselines on most aggregate comparisons across text-generation and agentic benchmarks.
In briefSoft Spatial Reasoning: A post-training method that replaces discrete reasoning tokens with continuous soft states in vision-language models
SummaryVision-language models often reason about space through chain-of-thought, committing to one discrete token per step, so early errors can propagate. This paper proposes Soft Spatial Reasoning, which mixes token embeddings into continuous soft states, with a controller called AdaptSoft that adjusts the degree of softness from hidden states and predictive uncertainty. The controller is trained with a gradient-alignment objective. The authors report that it outperforms hard and fixed-soft chain-of-thought baselines on the same backbone across diverse spatial benchmarks.
Rafi Ibn Sultan, Md. Sajid Alam Chowdhury, Saleh Zare Zade et al.
In briefSpatialCORE: confidence-aware grounded spatial reasoning for vision-language models
SummaryThe paper addresses the weakness of Large Vision-Language Models in spatial reasoning, noting that existing grounding-based methods mainly reward final-answer correctness. The authors propose SpatialCORE, a post-training framework that weights predicted bounding boxes by the model's coordinate-token confidence and gates grounding optimization with answer correctness. The authors report state-of-the-art results among open-source and specialized spatial reasoning models across diverse benchmarks, with effective zero-shot transfer to unseen data distributions.
Rafi Ibn Sultan, Xiangyu Zhou, Md. Sajid Alam Chowdhury et al.
In briefHard-Region Supervision: first place on the Waymo 2D video panoptic segmentation leaderboard
SummaryThis paper describes the winning entry to the Waymo Open Dataset 2D Video Panoptic Segmentation Challenge, which requires per-pixel semantic labels and identity tracking of countable objects across frames and five cameras. The authors build on the DVIS++ baseline and propose hard region supervision, which defines hard regions as where the baseline makes mistakes and adds a loss and an auxiliary head for them. The auxiliary head is used only during training. Three test-time steps are added: a two-model ensemble, merging the segmenter output into the final panoptic map, and cross-camera identity linking.
In briefStaying on Task: A paper testing whether AI models stay accurate during long, repeated operations
SummaryThis paper introduces Long-Transduction, a diagnostic that tests whether a model can keep performing context-dependent operations such as arithmetic, sorting, variable lookups, and table transformations throughout a long generation. The authors independently varied context length, input format, and local task complexity, evaluating seven open-weight models. They report performance decreases of 62.8% when context length scales from 4K to 128K, 36.5% when input format changes, and 39.9% when local task complexity increases.
Jeffrey Willette, Krishna C. Puvvada, Boris Ginsburg
In briefSCALE: Synthetic Calibration via Agreement Labeling in Embedding Space, a method to make pathology AI confidence scores more reliable
SummaryThis paper examines calibration, whether a model's predicted probabilities match real outcomes, across eight pathology foundation models. It finds calibration error is consistently higher on cases where pathologists disagree, a pattern hidden by aggregate expected calibration error. The authors propose synthetic agreement calibration, which interpolates between high-confidence class anchors in embedding space to create an ambiguity signal for retraining a linear probe. On MHIST it recovers most of the gain from real pathologist agreement, and it also improves calibration on PatchCamelyon and BreakHis while preserving discrimination metrics.
In briefBudget boundary effects in test-time math reasoning: when a token cap cuts off a solution mid-derivation
SummaryThis paper studies what a test-time controller should do when a cumulative token cap falls inside a math derivation: stop strictly at the cap, or let the current attempt finish. Using paired offline replays of 19,200 public traces covering 120 AIME, BrUMO and HMMT problems and two archive configurations of one model, the authors find that most advisory accuracy gains come from replacing abstention with a correct answer. They also report that more candidate coverage does not guarantee higher accuracy, and argue that budget curves should state the cap, realized cost, candidates, stopping rule and selector information.
In briefCascadia: A control-plane-free alternative to hyperconverged AI infrastructure for serving LLMs
SummaryThis paper presents Cascadia, a system for serving large language models across fleets of commodity Intel AIPCs using their CPU, integrated GPU, and NPU, with no dedicated routing control plane. Each node handles ingress, scheduling, and execution, and supports whole-model, replicated, and pipeline-sharded serving modes. In tests, a three-node NPU testbed achieved 3.10x the throughput of one node under ten concurrent requests, and a four-node deployment reached 4.06x direct single-node serving.
In briefAlways-On Experiments: Sequential Tests That Control False Discoveries as Treatments Are Added and Removed
SummaryGenerative AI has sped up the creation of new treatments, so experimentation platforms now often run continuously, adding treatments when ready and removing underperforming ones. The paper formalizes this "Always-On" setting and studies how to accept or reject each treatment while controlling the false discovery rate. The authors develop sequential tests with time-uniform Type-I error control under arbitrary stopping times and predictable treatment schedules. Built on the testing-by-betting framework, the test supermartingales for each treatment's average effect are shown to be growth-rate optimal in an almost-sure sense.
WhyFormalizes always-on experimentation and builds time-uniform sequential tests via testing-by-betting, with growth-rate optimality; novel but evidence is abstract.
Ricardo J. Sandoval, David Arbour, Avi Feller et al.
In briefNo Corners Cut: State-faithful transitions when switching prompts mid-stream in video generation
SummaryStreaming video generators let users change prompts during synthesis, but existing methods may skip necessary intermediate state changes. The paper proposes SEGUE, which uses a training-free planner to write transition prompts with roles and durations at each switch, and SPANDMD, a training method that keeps each prompt's DMD residual only within its assigned span. On the OpenTrans-360 benchmark of 1,800 switches, SEGUE ranks first on all eight transition metrics and raises the overall score from 0.866 to 0.887.
In briefGATE-ST: A Gene-Aware Text-Image Model for Predicting Spatial Gene Expression
SummarySpatial transcriptomics links gene expression to tissue morphology, but measuring it is expensive and slow. This paper introduces GATE-ST, which encodes text summaries of genes and fuses them with image features through cross-attention to predict spatial gene expression. The authors report that it outperforms random gene embeddings and several other image-text fusion architectures.
In briefEPIC: Epipolar-Consistent 360° Immersive Stereo Video Generation
SummaryThis paper presents a zero-shot pipeline that extends existing video diffusion models to generate 4K stereoscopic 360° videos for immersive headsets. Drawing on binocular vision and depth perception, the authors design an epipolar-aware 360° image matching metric that measures temporal and stereo geometric inconsistencies across views. They use this metric as a preference signal for direct preference optimization with limited training data. The work aims to provide a scalable way to generate content for immersive and mixed reality displays.
Debabrata Mandal, Dongdong Fu, Jonathon Miller et al.
In briefJusticeAxis: A Benchmark for Legal Judgment Balancing Rigid Rules and Unsupported Discretion
SummaryThe paper frames legal judgment as a reference-anchored task that must tie a decision to both the statute and the circumstances of the case. It introduces JusticeAxis, a benchmark of 256 real criminal cases from 18 countries with audio, image, and text evidence, each paired with three lawyer-written judgments. It also proposes JusticeAgent, a multi-agent harness whose skills are distilled from execution trajectories and admitted only under Bayesian credible bounds.
In briefConcept-Grounded Attention: Testing Graph-Injected Attention, Temporal Versioning, and Epistemic Status
SummaryThis paper introduces the Concept Lifecycle Model, which treats concepts as persistent, graph-grounded, time-versioned entities with provenance and epistemic status. It also proposes Concept-Grounded Attention, which injects concept-graph structure into transformer attention. Using controlled baselines with disabled mechanisms, the authors find that much of the graph-attention gain reflects added capacity, while explicit temporal representation and protocol-derived epistemic status improve answer accuracy and reduce unsupported assertions.
Sachin Dev Duggal, Pradyumna Swarnalatha Ramanna, Alexandros Vassiliades
In briefMotion-Based Representation for Cinematic Camera Trajectories
SummaryThis paper studies cinematic camera trajectories and finds that representing them as direction and speed, rather than per-frame poses, improves both trajectory-to-text alignment and text-to-trajectory generation. The authors introduce a new evaluation protocol, a generative model called CineGEN, and CineScript, a dataset of movie clips with scene descriptions and higher-level metadata.
In briefReGain: Restoring subject fidelity when personalizing on synthetic images
SummaryText-to-image diffusion models are personalized with DreamBooth fine-tuning on a few subject images, which increasingly come from other diffusion models rather than cameras. The paper finds that fine-tuning on such synthetic images degrades subject fidelity, causing oversaturated color and excess high-frequency detail, and traces this to classifier-free guidance. The authors propose ReGain, a training-free sampling-time correction that scales down each frequency band of guidance that is inflated relative to the base model, without needing real photos. On Stable Diffusion v1.5 it closes 51-64% of the subject-fidelity gap, and it also improves SDXL and SD 3.5 while preserving text alignment.
Shubhang Bhatnagar, Ishan Bhatnagar, Viraj Shah et al.
In briefKeeping Imagined Trajectories Within Offline Data for Model-Based RL
SummaryThis paper studies model-based offline reinforcement learning, where model errors can push imagined trajectories outside the offline data distribution. The authors propose in-distribution imagination (IDI), which estimates trajectory support in a learned representation space and truncates rollouts that leave the offline trajectory manifold. Combined with trajectory-regularized RL, IDI consistently improves performance in limited-data settings. Experiments show trajectory support predicts rollout failure better than transition-level uncertainty.
In briefProvable test-time scaling for beam search in LLM reasoning
SummaryThis paper studies the theoretical guarantees of beam search, a method that scores intermediate reasoning steps with a language model's own likelihood and uses an external reward model only on complete responses. The authors show that vanilla beam search needs at least Ω(C*(x)²) samples for the optimal response to survive, and propose confidence-filtered beam search (CF-Beam), which reduces this dependence to nearly linear. Experiments indicate beam search is more robust on hard instances and longer reasoning horizons.
In briefAuditing where scientific search agents fail to find or inspect target papers
SummaryThis paper argues that final-answer accuracy alone cannot show whether a scientific search agent never encountered a target paper, tried to inspect it, or gave an accepted answer after inspecting it. The authors propose decision checkpoints that record observations and tool actions during inference without benchmark labels, then join them with target identities and evaluator labels to assign outcome categories. In five conditions on 540 answerable AutoResearchBench Deep questions, keyword search reached 24.6% accuracy and raw search 17.8%.
In briefOff- vs On-Policy Distillation: Comparing Their Training Objectives
SummaryThis paper studies how a student model learns when distilled sequentially from multiple teachers. Forward KL divergence corresponds to a weighted arithmetic mixture of teacher targets, while reverse KL corresponds to a normalized weighted geometric aggregate. The authors develop learning algorithms for off-policy and on-policy feedback, with logarithmic regret bounds in the tabular setting extended to function approximation. The analysis explains both the benefits and the fragility of on-policy distillation.
In briefCollabFlow: Self-Improving Collaboration Among AI Agents
SummaryThis paper introduces CollabFlow, a system in which a trainable director builds teams of language-model agents, a frozen executor runs them, and each round's results are used to retrain the director. The authors add a flow-based training objective, Collaborative Trajectory Balance, that keeps several strong teams in use rather than concentrating on a few. On twelve datasets, the system outperforms all baselines and keeps improving across rounds.
WhyNovel trainable collaboration director with flow-based CTB objective; outperforms baselines on twelve datasets, giving solid evidence, with impact mainly within multi-agent systems.
In briefEvoSteer: Online self-evolving team orchestration for LLM agents, with a new credit-assignment method
SummaryThis paper proposes EvoSteer, an online self-evolving method that orchestrates LLM-based tool-using agents into communication graphs and repairs failing steps during execution. It introduces Anchored Trajectory Balance (AnchorTB), a flow-matching loss that assigns credit to each orchestration action by comparing subtrajectories against a frozen reference, and Validated Skill Admission, which tests candidate skills before promoting them. The authors report that EvoSteer significantly outperforms baselines on twelve datasets spanning question answering, math reasoning, code generation, and interactive decision making.
In briefBreaking Babel: A self-evolving multi-agent system for translating long subtitles
SummaryTranslating long subtitles requires understanding discourse and cultural context across episodes or whole series while keeping terminology and style consistent. The paper proposes SMART, a self-evolving multi-agent system that builds series-level memory during test-time training and translates sentences using a dynamic router, a Mixture-of-Agents layer, and tools for terminology and subtitle constraint checks. A judge-refiner loop updates agent prompts and routing policies without retraining the underlying LLMs. SMART achieves the best overall MQM score in all 15 Subtitle Arena directions, reducing average penalty by 6.9% over the strongest competing agent system.
In briefVmem-φ: Low-compute detection of unfamiliar inputs in spiking neural networks using membrane-potential statistics
SummaryThis paper studies spiking neural networks that process event-camera data and proposes using per-channel statistics of subthreshold membrane potential to detect out-of-distribution inputs. The authors introduce Gen1-C, a corruption benchmark built on the Prophesee Gen1 automotive detection dataset, with six histogram-level stress tests at five severity levels. Their Multi-Descriptor Deviation (MDD) method reportedly achieves an AUROC above 0.88 on five of six corruptions at the highest severity, using a 64-frame window.
Arul Rana, Agrim Tripathi, Shoaib Ahmed Dipu et al.
In briefMulti-armed bandits with several best options: worst-case optimal regret and why knowing the count matters
SummaryThis paper studies multi-armed bandit problems where several arms are optimal, a setting that fits many practical decisions with multiple correct answers. The authors sharpen the analysis of earlier sub-sampling algorithms, obtaining a minimax regret rate that nearly matches a matching lower bound. They also show that knowing the number of optimal arms up to constant-order factors is necessary to achieve near-optimal regret.
In briefTacit-TTS: Faster Voice Cloning Without Reference Transcripts
SummaryTacit-TTS is an efficient zero-shot voice cloning system distilled from IndexTTS2 that does not require transcripts of reference speech. It replaces autoregressive text-to-semantic decoding with masked non-autoregressive generation, adds training-free acoustic length estimation, and speeds up the flow-matching renderer via ReFlow distillation. The authors report competitive quality on two English and two Mandarin datasets, with generation over 10x faster than IndexTTS2 for utterances longer than 5 seconds.
In briefTERRA: Terrain-aware reconstruction and control for muscle-driven human walking on uneven ground
SummaryThis paper presents TERRA, an end-to-end pipeline that enables muscle-actuated virtual humans to move over non-flat terrain. From kinematic motion data alone, it combines terrain priors, estimated contacts, and free-space evidence to recover support geometry, while respecting anatomical, tendon, and contact constraints during retargeting. The authors train a single muscle-driven control policy on motion-terrain pairs from five datasets, totaling 9.4 hours of locomotion. They report improved terrain accuracy, fewer anatomical and interaction violations, and the highest completion rate across supported terrain families.
Merkourios Simos, Chengkun Li, Bianca Ziliotto et al.
In briefAgBench: Benchmarks for Evaluating Agentic AI on Personal Devices
SummaryThis paper presents AgBench, a benchmark suite and open artifacts for reproducibly evaluating agentic AI on personal devices. It compares local, hybrid, and cloud execution on task success, latency, cloud API cost, and data exposure. Results show local-only execution generally has lower task success and longer completion times than cloud-only execution, especially under higher concurrency, while eliminating cloud API costs and sensitive-data exposure.
WhyProposes a benchmark for local, hybrid, and cloud agent execution, comparing success, latency, cost, and exposure over 162 million data points; strong evidence with practical value for a subfield.
In briefUncertainty-normalized margins for improving preference-based AI training
SummaryThis paper addresses how Direct Preference Optimization (DPO) ignores preference strength and prompt-dependent uncertainty in human feedback. It proposes UNM-DPO, which combines strength-dependent margins with a learned prompt scale, along with two training objectives and a length-normalized variant, ULNM-DPO-WR. On HelpSteer2 and HelpSteer3, with Llama-3.1-8B-Instruct, the method achieves higher win rates against DPO and shorter responses on average, and it also reports improved length-controlled win rates on AlpacaEval.
Sadegh Khorasani, Petrus Mikkola, Matthias Grossglauser
In briefJoint supervised and self-supervised training with acquisition-robust techniques for faster 4D flow MRI image reconstruction
Summary4D flow MRI measures blood velocity in three directions over time, but its scans take a long time, and its diagnostic signal lies in phase differences between velocity encodings. This paper combines supervised and self-supervised training, using both magnitude and velocity data, and adds acquisition-robust and acceleration-conditioned strategies. On the CMRx4DFlow 2026 aortic dataset at 1.5T and 3T, the model lowers RelErr by 38–50% and AngErr by 7.0–8.7 degrees compared with a training-matched baseline at acceleration factors R=10 to 50, improving on every held-out subject. It also generalizes to out-of-distribution data, with SSIM up 7.2% and AngErr and RelErr down 38% and 31%.
In briefRecursive Organization Improvement: A Modeling Specification for Human-Agent Teams
SummaryThis paper proposes a modeling specification for how organizations improve recursively, combining an executable checker, a mapping to public records, and two simulation studies. The specification links actor-visible histories, organizational memory, decision rights, and the evidence used to keep changes. The authors find that stronger agents alone do not guarantee better organizations, since the retained work arrangements shape what evidence teams can learn from.
In briefChartRevise: A Dataset and Protocol for Exact Chart Editing via Code
SummaryThis paper introduces ChartRevise, a structured dataset and evaluation protocol for editing charts precisely through the code that draws them. The dataset, built on the grammar of graphics, contains 92,438 records covering 344 edit types, 20 chart types, and three plotting libraries. The reference-free protocol separately measures requirement completion, gratuitous changes, and missed coupled updates, combined with execution and rendering checks. Fine-tuning yields relative gains of 16% in mean requirement recall and 22% in mean exact-edit rate across five models and four external benchmarks.
In briefVision-language-action driving agent with language-based memory (AD-Memo)
SummaryThis paper proposes AD-Memo, a general-purpose vision-language-action driving agent that stores memory as language. It records driving-critical surrounding objects as an extension of its chain-of-thought, which becomes part of its future input, addressing the limited number of frames VLAs can process. The authors train the model with supervised fine-tuning and Da Capo, a semi-closed-loop reinforcement learning method, and report improved driving quality, scene question answering, and plug-and-play memory across scenarios such as all-way stops.
In briefComponent-aware feedback helps self-evolving programs learn from which edits worked
SummaryLLM-guided evolutionary search can find complex programs, but existing methods keep only candidate programs and fitness scores, not which component edits caused which metric changes. The authors propose component-aware feedback, which compares each program to its parent, logs changed components with their metric differences into an attribution memory, and evaluates them on LLM reranking across twelve Bright datasets. The method reached the strongest baseline's final quality after a median of one third of the search budget, ended 7.2% higher in held-out nDCG@10, and under a cost-aware objective found more accurate pipelines using 11% fewer tokens per query.
In briefSHIFT-Truck: A High-Fidelity Aerodynamics Benchmark for Pickup Trucks
SummaryThe paper introduces SHIFT-Truck, a dataset of 1,000 high-fidelity SA-DDES simulations of pickup truck geometries varied across 17 shape parameters, with surface and volume flow fields released. The authors benchmark four neural surrogate models (DoMINO, GeoTransolver, AB-UPT and SMART) on geometry-grouped splits. The dataset also includes controlled distribution shifts in operating conditions, surface discretization and vehicle archetype to test generalization.
Riddhiman Raut, Yin Yu, Aashwin Anand Mishra et al.
In briefCross-modal object tracking via feature-alignment domain adaptation between template and search frames
SummaryThis paper introduces TSDA-Track, a framework for object tracking when the template and search frames come from different sensing modalities. During training it reduces the modality gap using two alignment strategies: Pre-AFA applies adversarial alignment before template-search interaction, and Enc-CFA applies contrastive alignment to encoder features after interaction. Both variants use a shared inference pipeline without modality-specific branches. The method improves over representative state-of-the-art trackers on LasHeR and in zero-shot tests on RGBT234 and GTOT, with Pre-AFA reaching an SR/PR of 43.2/56.0 versus 36.8/50.0 for the ToMP-101 baseline under the modality-switch protocol on RGBT234.
Fereshteh Aghaee Meibodi, Amir Mehdi Soufi Enayati, Shadi Alijani et al.
In briefSTEPS: Diffusion-based scene text editing that keeps the original text style
SummarySTEPS is a diffusion model for replacing the text in images while keeping its original look, such as font, color, orientation, and background. It uses a style encoder that captures visual text style independently of the text content, combined with conditions on the target characters and rendered glyphs. The authors report that STEPS outperforms previous scene text editing methods in style preservation, readability, and subjective quality.
Nicolas Thiebaut, Nameer Hirschkind, Xiao Yu et al.
In briefTSGL: Teacher-Student Graph Learning for Compressing 3D Gaussian Splatting Models
Summary3D Gaussian Splatting (3DGS) is a popular method for novel view synthesis, but its models contain millions of Gaussian primitives, leading to large file sizes. This paper proposes TSGL, a compression method that works directly on a trained model without retraining or training images. It uses a teacher-student model to learn a geometry graph of similarities between neighbouring Gaussians, then applies a Graph Fourier Transform to compactly represent the remaining attributes. On three standard benchmarks, it achieves 27x to 33x compression with less than 0.6 dB PSNR loss.
In briefProbabilistic weather forecasting with a proper scoring rule-based diffusion model
SummaryThe paper proposes auxiliary conditional denoising tasks that predict future weather states from both the context and a corrupted version of it, to help learn the predictive distribution. It trains a single stochastic predictor by minimizing a proper scoring rule across noise levels, while still generating ensemble members in one forward pass at inference. Controlled experiments report improved one-step forecasts across architectures, with larger gains at longer horizons, and better calibration.
In briefFast, accurate multilingual tagger for removing personal information from text
SummaryThis paper presents a general-purpose multilingual named-entity tagger for privacy redaction, which removes personal information while preserving relationships in the text. It fine-tunes XLM-R-large on text in 35 languages annotated by frontier language models, plus mapped human-annotated corpora. On human-annotated test sets, it reaches 88.2 redaction F1, above GLiNER2 at 69.1 and Microsoft Presidio at 61.4, and runs 4.9 times faster than GLiNER2 on CPU.
In briefMarking Yorùbá Contour Tones: A Typographic and Computational Proposal
SummaryThe paper addresses the difficulty of showing rising and falling tones in Yorùbá spelling, which for some personal names omits tonal information and can even reverse the name's intended meaning. The authors propose using the caron and circumflex on single vowels to encode these contour tones, making them typeable and processable by computers. They report an implementation in the WriteYoruba keyboard and the TTSYoruba speech synthesizer, with architecture and listener evaluation covered in a separate report by Tubosun, Oluokun et al. (2026).
In briefInterpretable but Fragile? Concept Bottleneck Robustness Under Geometric-Semantic Perturbations
SummaryThis paper studies whether Concept Bottleneck Models, which use intermediate concept layers for interpretability, are more robust than standard classifiers. The authors propose a generator-based framework that compares the two under continuous geometric perturbations in latent space and discrete semantic interventions in concept space. They assess robustness empirically and with certification via randomized smoothing. The findings indicate that interpretability does not inherently confer robustness; concept bottlenecks instead redistribute where sensitivity appears.
In briefGeometry and physics confounding hinders learning PDEs across changing domains
SummaryThis paper argues that when domain shape varies, geometric effects and intrinsic physics become confounded, a shared failure mode for learning partial differential equations. The authors propose a de-confounding framework that explicitly models the known geometry-to-operator transformation. Across five operator-learning benchmarks it improves prediction and data efficiency, and in two evolving-domain systems it recovers the generating equations and cuts held-out PDE residuals by more than two orders of magnitude.
In briefEulerian Motion Reconstruction for Water Scenery: a 3D method that rebuilds moving water from one video as a looping, viewable scene
SummaryThis paper reconstructs water scenes from a single non-looping 2D video as a 3D looping 4D model that can be rendered interactively from new viewpoints. Motion is modeled by a static Eulerian motion field that moves Gaussian splats, which are reborn at fixed periods and trained with rendering losses. A non-periodic, time-varying residual term is added to capture stochastic dynamics, and the authors report more photorealistic water animation than prior methods.
In briefDo language models catch real conflicts between scientific findings?
SummaryThe paper studies whether language models can tell when two scientific findings actually conflict. The authors build an unsatisfiable XOR constraint system and express its constraints as scientific reports from different labs, asking models to choose between the assignment that satisfies more constraints and the one that better fits biological expectations. Models recover the best-supported assignment well when constraints are stated directly, but in scientific prose one model often favors the biologically expected answer, and removing that preference raises recovery substantially.
In briefHIGS: Hierarchical Implicit Grids for Joint Geometric and Semantic Scene Understanding
SummaryThis paper proposes HIGS, a hierarchical neural implicit representation that splits a scene into overlapping submaps and optimizes each hierarchically to improve scalability and efficiency. It embeds both geometric and vision-language features in the same map, supporting SDF construction and open-vocabulary object grounding. The authors report improved computation and memory efficiency with maintained accuracy on large-scale real-world benchmarks, giving robots spatial awareness.
