In briefAnyAct: A Universal Action Layer for Self-Evolving AI Agents
SummaryThis paper presents AnyAct, a universal action layer that organizes available tools into a self-evolving action space so AI agents can work reliably in large, changing tool ecosystems. It selects task-relevant actions through hierarchical retrieval, removes unreliable ones during testing, and unifies different feedback formats. The authors report better performance than baselines on LiveMCPBench and on OSMCP, a new benchmark they created, where it reaches 77.27% overall success using about half the steps of most competitors.
In briefLanguage as the Interface: Foundation-Model Contrastive Learning Links Gene Expression and Neuron Electrophysiology
SummaryThe paper introduces LangPatch, a contrastive learning framework that aligns gene descriptions and verbalized electrophysiological profiles through a shared frozen text encoder, trained on paired Patch-seq data. It achieves the highest mean transcriptome-to-electrophysiology prediction correlation among evaluated methods across mouse and human cortical cohorts. The model also transfers from mouse visual cortex to motor cortex and human cortex with partial correlation retention.
In briefContractRL: Constrained policy optimization that repairs malformed tool-call JSON step by step and can be audited
SummaryStructured tool calls often fail when a few fields break a schema or execution contract, and regenerating the whole object makes repairs hard to audit. The paper proposes ContractRL, which treats verifier-guided JSON repair as a bounded sequential decision process, using a contract-derived action mask to filter invalid operations before a deterministic validator applies each change. Under identical verifier information, it reports a semantic success of 0.9362 with about 34.4 generated tokens, outperforming the Patch-SFT and full-regeneration baselines.
In briefPhysics-Informed Multi-Agent Coordination for Hospital Patient Flow Optimization
SummaryThis paper proposes a multi-agent framework that embeds empirically calibrated BCMP queueing networks as physical priors within decentralized multi-agent reinforcement learning to coordinate patient routing and service scaling across hospital departments. Departmental agents cooperate by exchanging localized action fingerprints along network edges, which mitigates non-stationarity while limiting communication overhead. Evaluations on real MIMIC-IV patient trajectories report substantially lower cumulative system delay than static Markovian approximations, heuristic dispatching, and independent multi-agent baselines, while maintaining clinical safety constraints.
In briefLatCom: Cross-Agent Latent Compression for Efficient Multi-Agent Collaboration
SummaryThis paper proposes LatCom, a framework that compresses latent representations exchanged between LLM-based agents in a multi-agent system. Multiple sender latents are mapped into a fixed number of receiver-readable, task-relevant slots, optimized for the receiver's task utility rather than full reconstruction. Training occurs in two stages: single-sender readability learning, then multi-sender fusion to remove cross-agent redundancy. On Qwen3-4B across multiple benchmarks, LatCom reports an average 2.46x inference speed-up over LatentMAS and 70.3% fewer output tokens with comparable average accuracy.
In briefBack2Struct: Recovering editable vector code from structured images
SummaryThis paper presents Back2Struct, a method that converts images of diagrams, charts, and flowcharts back into editable SVG/XML code. Instead of low-level pixel vectorization, it predicts object-level code encoding text, shapes, topology, and layout, which can be imported into tools like PowerPoint. The model is trained with supervised fine-tuning and then refined with a composite reward for syntactic validity, length consistency, and structural or semantic similarity. The authors report improvements in accuracy, editability, validity, and user alignment over baselines.
In briefRBF-GNN: Graph neural network using rational basis functions for coordinate-based graph convolution
SummaryThe paper proposes RBF-GNN, a graph neural network that uses Euclidean, spherical, or angular pseudo-coordinates to add spatial inductive bias. It follows the SplineCNN architecture but replaces B-splines with rational Padé basis functions, avoiding the exponential growth of B-spline count with dimension. The authors add a spline-subspace initialization and variance-preserving weight rescaling for training. Replacing only the SplineCNN layers with RBF-GNN yields improved results on semantic keypoint matching, shape matching, and event-camera vision tasks.
In briefOn-board two-time-point building damage assessment for reducing satellite data transfer
SummaryThis paper presents a bi-temporal building damage assessment pipeline based on a YOLOX-derived siamese detector, aimed at rapid post-disaster assessment. Pre-disaster reference images are compressed and uplinked to the satellite, where they are compared with new post-disaster imagery so that only object-level bounding boxes and damage classes are downlinked. A latent-space shift estimation module is added to correct co-registration errors, and hardware performance is reported on Xilinx Versal VCK190 and NVIDIA Jetson AGX Orin.
Thomas Goudemant, Benjamin Francesconi, Marjorie Bellizzi et al.
In briefCADOC: Cache-Aware Dynamic Object Context for Long-Horizon Agents
SummaryLong-horizon agents struggle because their full history is resent with each request, which limits task length and degrades reasoning. CADOC replaces structured objects with compact Cards while keeping their original contents retrievable on demand. It schedules these replacements in batches by balancing accumulated waiting cost against shared cache-reconstruction cost, using a rule derived from an economic order quantity trade-off. The method reportedly lowers aggregate input cost by about 40% on average while keeping task performance close to full context.
In briefAn LLM-based agent framework simulating diverse evacuation behavior of people fleeing a moving threat in a plaza
SummaryThis paper proposes a framework in which pedestrian agents powered by large language models make evacuation decisions under a moving threat. Each agent sees only its own private view and keeps a memory-based knowledge graph built from its observations, with persona-conditioned prompts guiding choices. In a simulated public plaza, 89.5% of agents who knew a usable exit evacuated, compared with 1.05% of those who did not.
In briefWorld2Motion: Converting Video World Models into 3D Human Motion Generators
SummaryThis paper presents World2Motion, a framework that generates scene-aware 3D human motion and matching video from one image and a text prompt. The authors adapt the video world model Cosmos 3 into a single-stage 3D motion generator, using a dataset that mixes synthetic video-motion pairs with real videos paired with estimated 3D motion, and a shift-decoupled noise schedule to reduce motion jitter. On a multi-source interaction benchmark, it shows better motion-text alignment and scene interaction than the evaluated baselines, matches the two-stage baseline's interaction success rate, and runs about 3.3 times faster.
In briefVesselBench-800K: A Large-Scale Benchmark for Multimodal Vessel Detection, Counting and Density Estimation
SummaryThis paper introduces VesselBench-800K, a global-scale benchmark for vessel perception in multimodal remote sensing images. It contains 800,000 image pairs of 512x512 pixels, combining optical and SAR imagery, for detection, counting, and density estimation. The authors evaluate several state-of-the-art models on the dataset and state that it will be publicly released.
In briefVisual Parallel Search: Learning to Search High-Resolution Images by Inspecting Tiles in Parallel and Zooming Adaptively
SummaryHigh-resolution visual question answering often fails when models cannot obtain the small, localized evidence a question requires. The paper proposes VPS, a framework where a main agent first inspects image tiles in parallel via grid_search and then adaptively zooms into precise regions via zoom_in. It supports both training-free inference and post-training, and reports improved mean accuracy over zoom-only search in most same-model comparisons across five benchmark splits and three model sizes.
In briefFaster video generation via structured sparse attention
SummaryDiffusion Transformers produce high-quality videos but are slow at inference because full spatio-temporal attention is computationally expensive. The paper finds that video DiT attention shows periodic diagonal stripe structures along temporal and spatial dimensions, and proposes PSA, a parameterized stripe attention run by a single hardware-efficient CUDA kernel. A training-free offline search sets per-head sparsity under a specified error tolerance. On HunyuanVideo and Wan 2.1, PSA reports end-to-end speedups of 1.57x and 1.37x over FlashAttention-3 baselines, with acceptable visual quality degradation.
In briefSurvey on challenges and AI support for integrating SysML models across organizations
SummaryThis paper reports an online questionnaire survey of 29 MBSE stakeholders involved in cross-organizational collaboration. Respondents rated eight integration challenge categories and six AI-supported task types on five-point Likert scales. The results indicate that model integration is perceived as a multi-dimensional alignment problem spanning semantics, behavior, traceability, and exchange interoperability, with views varying by role and involvement frequency. AI was rated highly useful for analysis tasks such as semantic structure analysis and inconsistency detection, and respondents predominantly preferred human-in-the-loop use with mandatory verification.
In briefImbalancE: Inference-time latent search to fix degree imbalance in link prediction
SummaryKnowledge graph embedding models predict missing links, but their representation quality varies across the graph. The paper finds that this problem correlates with degree imbalance between entities in test triples, and is especially severe when the target entity's degree is much smaller than the anchor's. The authors propose an inference-time latent search method that builds on a pre-trained model, exploring the embedding space at evaluation time and blending known and out-of-band information. On imbalanced triples from common benchmark datasets, the method outperforms conventional approaches.
Alberto Bernardi, Luca Costabello, Christophe Gueret
In briefSalt++: Context-Aligned Post-Training for Fast Streaming Audio-Video Generation
SummaryThis paper proposes Salt++, a two-stage post-training framework for few-step streaming audio-video generation. It combines Causal Self-Flow, which aligns a noisy-history student with a clean-history teacher to improve semantic extraction and cross-modal alignment, and context-aligned autoregressive DMD, which matches generation and scoring contexts. At 480p under the same 4-step causal setting, the authors report 57% and 45% gains in visual and motion quality over OmniForcing on JavisBench. A scale-wise stage extends it to 4-step 1664×960 generation, outperforming bidirectional LTX-2 on six of seven reported metrics.
In briefCypherTurn: a multi-turn benchmark for conversational text-to-Cypher and the autonomy divergence
SummaryGraph databases are increasingly queried in natural language, but existing benchmarks only test isolated single-turn queries rather than multi-turn analysis sessions. This paper introduces CypherTurn, comprising 721 sessions and 5,927 turns across 7 knowledge graphs and 13 conversational phenomena. Evaluating 15 models under guided and fully autonomous protocols, the best model reaches only 64.7% execution accuracy and session-level correctness stays below 5%.
WhyFirst multi-turn conversational Text-to-Cypher benchmark, revealing autonomy-driven leaderboard reordering; supported by 721 sessions and 15 models, useful to the subfield.
In briefCF-LoRA: Clustered Federated LoRA Fine-Tuning with Separate Factor Aggregation
SummaryThis paper proposes CF-LoRA, a federated framework for fine-tuning LoRA adapters when client data is heterogeneous. It shares a global A factor while keeping per-client B factors, then clusters clients by cosine similarity of their B factors and aggregates B within each cluster with A frozen. Experiments on four language tasks and four vision datasets using RoBERTa and ViT report the highest average accuracy in both modalities, with only one LoRA factor communicated per round.
In briefAbductive World Modeling with Causal Representation Learning
SummaryThis paper proposes Abductive World Modeling (AWM), which learns structured causal representations by inferring latent causes from predicted futures, so models can reason about why and how the world changes. It realizes this through the Hierarchical Abductive State Pyramid (HASP), which organizes world states into Entity, Dynamic, and Relation components. Experiments report that compared with V-JEPA, AWM improves physical prediction AUROC by 10.7%, causal reasoning accuracy by 16.8%, and action Top-1 accuracy by 68.0%.
In briefREALHOP: Auditing Whether Multi-Hop Reasoning Benchmarks Truly Require Each Step
SummaryThis paper argues that multi-hop question benchmarks judged only by answer accuracy cannot show whether models actually use the evidence for each intended reasoning step. The authors introduce the Behavioral Necessity Rate (BNR), which measures how often removing targeted evidence prevents correct answers, and find panel-mean values of 16.6% to 48.9% across five existing benchmarks. They then present REALHOP, a framework that builds questions verified for per-hop dependence, raising panel-mean BNR on MuSiQue from 27.4% to 94.4%.
WhyProposes a behavioral necessity metric and benchmark framework, validated across multiple benchmarks and models, mainly influencing multi-hop evaluation methodology.
In briefSRJudge: Using selective reasoning in large language models to tag educational content with fine-grained knowledge concepts
SummaryThe paper proposes SRJudge, a three-stage Select-Reason-Judge framework for tagging educational content with specific knowledge concepts. A fine-tuned small model such as BERT first narrows candidates to a top-K shortlist, a lightweight LLM then reasons over them using improved reinforcement learning, and a larger LLM judges the rationality of the reasoning. The authors also build biology and physics datasets, and report that the method outperforms state-of-the-art baselines.
In briefUltraMatch: Fast, Low-Memory Image Matching via Transport Path Routing
SummaryUltraMatch is a semi-dense image matching framework that routes only a small set of candidate matching paths instead of computing full token-level matching. A lightweight router ranks candidate blocks, and a sparse global Dual-Softmax is applied over the selected candidates. The paper reports accuracy competitive with other semi-dense matchers, 1.67 times faster than SuperPoint+LightGlue, and 0.44 GiB peak inference memory. It also reports inference at up to 6K resolution on a single RTX 3090.
WhyReplacing dense token matching with sparse routed block candidates is a substantive improvement, validated by speed, memory, and transfer to EDM and ELoFTR.
In briefAMU: Admission and Memory Update for Personalized Conversations, a structured memory control method guided by small language models
SummaryThis paper addresses how personalized AI assistants maintain persistent user memory across long-term conversations, focusing on controlling what gets written into memory. The proposed AMU framework uses structured memory filtering to decide what enters memory, and a small language model guided storage manager to decide whether a record is stored separately, discarded as a duplicate, or fused as an update. Experiments in a controlled memory writing and retrieval setting show that AMU keeps personalized memories cleaner and more retrievable.
In briefDual-track label curation and classification framework for wheat area estimation in Punjab, India
SummaryThis paper estimates 2022 Rabi-season wheat area in Patiala district, Punjab, using Sentinel-2 NDVI time series. An 849-sample reference dataset was built through iterative rule-based bootstrapping, and four classifiers were compared. XGBoost achieved the highest overall accuracy at 78.82%, and its predicted wheat extent deviated only +2.99% from the official tabular target, compared with +25.11% for the government spatial mask.
In briefRepetition, Not Length: Why Speech Models Miscount Repeated Phrases
SummaryThis paper studies why text-to-speech models loop, truncate, or lose count on text that repeats a phrase many times. The authors pair each repeated text with a control of matched sentence and word count that has no back-to-back repeated words. Six models rendered the controls almost perfectly but got only 18.2% of repeated texts exactly right at k ≥ 6, versus 94.3% for controls. They conclude repetition itself, not length, is the cause.
Kirill Borodin, Vasilii Kudryavtsev, Maxim Maslov et al.
In briefStructured visual targets for EEG-to-image retrieval across subjects
SummaryThis paper addresses cross-subject EEG-to-image retrieval, where EEG representations learned from source subjects must align with visual embeddings for an unseen subject. Instead of focusing on the EEG side, it preserves spatial information from the Perception Encoder, turns its patch grid into learned visual views, and aggregates them with a content-dependent router. Under leave-one-subject-out evaluation on THINGS-EEG2, it reports 35.3%/65.6% Top-1/Top-5 accuracy, and a training-free refinement raises this to 48.1%/77.1%.
Salini Yadav, Taveena Lotey, Micka\"el Coustaty et al.
In briefEqually Good, Yet Different: Benchmarking Rashomon Sets in AutoML Packages
SummaryThe paper introduces ARSA ML, an open-source Python framework that measures Rashomon set structure and predictive multiplicity within AutoML pipelines. Benchmarking AutoGluon and H2O on 28 binary classification datasets, the authors find AutoGluon yields larger, diverse sets with stable explanations. H2O produces compact sets with higher prediction divergence and explanation instability, making its users more exposed to x-hacking.
Katarzyna Wo\'znica, Katarzyna Rogalska, Zuzanna Sie\'nko et al.
In briefPrior-driven 3D Gaussian Splatting with normal and depth regularization
SummaryThis paper addresses geometric inaccuracies and visual artifacts in 3D Gaussian Splatting (3DGS) for complex scenes. It regularizes optimization with surface normal priors, which align Gaussian covariance with local surface structure, and dense depth priors combined with SfM initial points for better per-pixel depth. The method is tested on street-view and highly reflective datasets across multiple SfM pipelines, and the results report improved geometric accuracy and visual quality.
In briefTaskBridge: Linking Unsupervised Tabular Anomaly Detection and In-Context Learning via Virtual Tasks
SummaryThis paper presents TaskBridge, a framework for unsupervised tabular anomaly detection that reuses pretrained general-purpose tabular foundation models. It builds virtual supervised tasks so that anomaly detection becomes in-context inference, where anomalies receive lower support than normal samples. Across 790 real-world datasets, TaskBridge reportedly outperforms 30 baselines, including state-of-the-art foundation-model methods, without anomaly-specific pretraining or dataset-specific optimization.
In briefJudgeCast: Time Series Forecasting with Experience-Based Covariate Judgments
SummaryThis paper introduces JudgeCast, a framework for time series forecasting with covariates. A frozen time series foundation model produces a base forecast, and a frozen LLM adjusts it using current context and past experience, first making explicit judgments about each covariate. After observations arrive, the method reconstructs alternative judgments from the residual and keeps the best-performing decision as validated experience. The authors report that it outperforms strong baselines across diverse real-world datasets.
In briefChinese-Jev: Applying a System One Model to Chinese-Language Tasks
SummaryThis paper presents Chinese-Jev, a System One model for Chinese-language decision tasks that uses a lightweight encoder-only backbone to score candidate answers. It is pre-trained on a 10-million-example general corpus and then fine-tuned separately for medical, legal, and financial domains, and the authors introduce the CJ-Bench evaluation benchmark. According to the paper, it beats the closed-source Jev model by 1.24% on general tasks with a 20.3x speedup, and gains 4.0% accuracy in medicine with about 15 ms latency per example.
In briefBeyond Readability: Measuring Whether Task Information Can Still Be Recovered
SummaryThis paper argues that direct visual readability and recoverability of task information are different measures. The authors propose an evaluation framework that makes observation, query, target, and available knowledge explicit, and compares the overlap of success sets across recovery routes. In a controlled book-cover study with a fixed title-author catalog, entity-linked recovery (identifying the title visually, then looking up the author) succeeded across the tested vision-language models beyond the direct optical frontier, while a substantial region remained recoverable only optically.
In briefJoint sparse-quantization co-design for trillion-parameter MoE models
SummaryThis paper presents a hardware-software co-design framework that compresses Mixture-of-Experts (MoE) expert weights into low-precision sparse formats that run on Sparse Tensor Cores. It relaxes semi-structured sparsity selection via continuous reparameterization, allowing differentiable joint optimization with quantized weights, and adds a custom grouped sparse GEMM kernel. Reported results on MoE models from 30 billion to one trillion parameters include up to 4.35 points of accuracy gain over prior joint methods, retention of 96.09% of original performance, and on NVIDIA B200 GPUs, kernel speedups up to 1.65x, 1.18x higher serving throughput, and up to 4.03x lower end-to-end latency.
WhyContinuous reparameterization enables joint sparse-quantization optimization plus a custom grouped sparse GEMM kernel, evaluated with measured speedups on 30B–1T MoE models.
In briefPurlin: Separating Orchestration from Data Movement in GPU Collective Communication
SummaryThis paper presents Purlin, a scale-up communication framework that separates the semantics, orchestration, and datapath of GPU collective communication. It introduces a shared orchestration protocol called SNAC and a hardware-specific datapath called Atom, so new hardware mechanisms can be adopted while reusing orchestration. Evaluated on A100, H200, and B200 GPUs across seven collectives, Purlin achieves up to 5.14x latency speedup and 4.50x bandwidth improvement. Integrated into SGLang, it improves LLM serving throughput and interactivity, and reduces diffusion image generation latency by up to 1.13x.
Osayamen Jonathan Aimuyo, Swapnil Gandhi, Christos Kozyrakis
In briefCool the Sampler, Not the Learner: Lowering sampling temperature to tolerate stale samplers in GRPO
SummaryThis paper studies reinforcement learning for language models when the sampler falls behind the learner and is refreshed only occasionally. On Qwen2.5-Math-1.5B and GSM8K, importance-corrected GRPO refreshed every 192 updates degrades severely after about 180 steps. The authors propose decoupled cooling, which samples at temperature 0.8 while the learner and importance weights stay at temperature 1, and report that cooled runs remain stable and match more frequent refreshes at the same update budget.
In briefER-JEPA: Experience Replay for Better Joint-Embedding Predictive Learning in Language Models
SummaryThis paper proposes ER-JEPA, which adds an episodic replay path to LLM-JEPA, a method that aligns different views of the same knowledge in language models. Training pairs are stored in memory, and relevant stored data is retrieved at each step to provide additional supervision for token prediction and representation alignment. Experiments on NL-RX, GSM8K, Spider, and NQ-Open show that ER-JEPA consistently outperforms LLM-JEPA.
In briefDiffusion-based simulation inference that handles simulator misspecification and composes observations
SummaryThis paper proposes diffusion-based simulation-based inference methods that combine heterogeneous observations, preserve hierarchical latent structure, and address mismatch between the simulator and observed data. It introduces a continuous-time diffusion coefficient, Hierarchical Blockwise Diffusion Sampling (HBDS), and path-regularized fine-tuning, using Girsanov's theorem to quantify path divergence between pretrained and fine-tuned models. Applied to 940 measurements across four cell lines in a Bone Morphogenetic Protein signaling model, fine-tuning improved posterior-predictive accuracy over the pretrained model.
In briefFine-tuning on self-generated, reward-weighted data: convergence and off-policy benefits
SummaryThis paper studies how fine-tuning a policy model on self-generated, reward-weighted data behaves, focusing on RE(S), a generalized REINFORCE that updates its rollout distribution every S gradient steps. It views RE(S) as a stage-wise optimization that minimizes KL distance to a fixed reward-weighted rollout distribution. For softmax policies in multi-armed bandits, the authors prove global convergence for any fixed S, a tight Θ(1/T) suboptimality rate, and show that starting from a weak policy, RE(1) can stall near suboptimal policies while a suitable S converges faster.</summary>
In briefRethinking weakly supervised chest X-ray localization: null controls, protocol dependence, and PACE
SummaryThe paper argues that weakly supervised localization on chest radiographs is usually reported as one overlap accuracy, which is not comparable across papers. The authors use position-null controls to measure what those scores contain and find that several rarely stated protocol choices shift results substantially. They introduce PACE, a readout combining label-trained and frozen-backbone evidence, and report its gains alongside its costs and limitations.
In briefDual-Channel Robust Group-Relative Policy Optimization: A Method Using Advantage and Sequence-Weight Estimation
SummaryThis paper addresses instability in group-relative policy optimization caused by outlier rewards and token-level log-ratio perturbations. It proposes RoVR-GSPO, with a reward channel using robust reference estimation and bounded residual credit, and a ratio channel using differentiable SoftRoVR aggregation for robust sequence weights. The authors provide stability and efficiency analyses for both channels. Experiments on mathematical reasoning, long-context summarization, and tool-call annotation show consistent gains over GSPO, and controlled perturbation studies indicate greater robustness to reward contamination and token-ratio anomalies.
In briefNeural networks built to be safe by design, using energy-based learning
SummaryThis paper proposes a neural architecture that combines energy-based modern Hopfield networks with a port-Hamiltonian neural ODE to guarantee safety by design. It constructs barrier functions that yield explicit sets of admissible inputs and quantitative robustness radii. Tested on several benchmarks, including a 12-dimensional nanodrone model, the authors report state-of-the-art performance and certified invariant sets more robust to external disturbances than comparable methods.
In briefDispFlow-GS: Motion-Supervised Monocular Dynamic 3D Gaussian Splatting
SummaryThis paper addresses dynamic scene reconstruction from a single camera for robotic perception. It argues that existing Gaussian-flow supervision aligned with optical flow has limited benefit due to a domain gap, and proposes Displacement Flow supervision with motion disentangling via intermediate-view rendering. The authors also introduce Deformation-Rendering Consistency (DRC), a motion-aware metric. Experiments report motion localization and motion-rendering consistency gains of up to 39% and 6%, while image-based metrics change by only about 0.1%.
In briefSCA: Spatial Credit Assignment for Training GUI Agents with Reinforcement Learning
SummaryThis paper proposes Spatial Credit Assignment (SCA), a training method for GUI agents that operate on screens. It uses the screen coordinates of sampled clicks to refine group-relative rewards, adjusting credit with prediction residuals when successes and failures are mixed, and ranking by distance to the target when all clicks fail. The spatial information is used only during training, and the deployed policy stays unchanged. The authors report improved GUI grounding and action-prediction results across benchmarks.
In briefWeLike2Party: In-Context Motion Transfer for Multi-Person Animation
SummaryThis paper presents WeLike2Party, a framework that transfers motion from a driving video to multiple people in a reference image, using direct video conditioning without explicit pose or mesh extraction at inference. It adds Reference Asymmetric RoPE Conditioning to preserve appearance details and Identity Binding Supervision to link each identity to its motion trajectory. The authors also build MotionTwin, a synthetic dataset of 14.4K cross-identity video pairs totaling 84.3 hours, and MotionTwin-Bench, an evaluation benchmark. Experiments report gains over recent methods, especially in multi-person scenes with heavy occlusion.
In briefCoEM: Commit-on-Evidence Memory for Better Long-Context Reasoning
SummaryThis paper introduces CoEM, a memory method that helps language models reason over long inputs. It keeps potentially useful source excerpts in a pending set, then promotes, retains, or discards them as later context clarifies their relevance, with a frozen verifier accepting only evidence-backed facts. The policy is trained with reinforcement learning using step-level evidence and final answer rewards. On 6,400-document inputs, the authors report CoEM beats the strongest memory baseline by 10.4-11.4 F1 points on Qwen3.5-9B.
In briefVLALight: A Vision-Language-Action Model for Controlling Traffic Signals
SummaryThis paper presents VLALight, described as the first vision-language-action model for end-to-end traffic signal control using multi-view roadside videos. It maps camera observations directly to coordinated signal actions, without relying on manually engineered traffic states or separate perception modules. The model uses two-stage supervised cold-start training followed by cooperative reinforcement learning, and can switch between fast and slow reasoning modes. Experiments on seven real-world traffic-flow datasets across three urban networks reportedly outperform transportation-based, RL-based, and LLM/VLM-based baselines.
WhyFirst VLA model for end-to-end signal control from multi-view roadside video, with adaptive reasoning; outperforms diverse baselines on seven real datasets across three networks.
In briefLearning from the Gap: Pruning Advantages and Adaptive Sampling for Faster GRPO Training
SummaryThis paper addresses the high computational cost of GRPO-style reinforcement learning, where multiple rollouts per question and low-information trajectories slow learning. It proposes FastRL, a plug-and-play framework that prunes trajectories to keep high-advantage ones while maximizing gradient diversity, and adaptively adjusts rollout sampling across training stages based on historical pruning. The authors report an average 2.07× training speedup on Geometry3K and GeoQA8K-R1V and about a 1.64% gain in average accuracy on visual reasoning benchmarks.
In briefDating the Model: how a hidden date in system prompts changes LLM evaluation results.
SummaryThe paper finds that the current date is silently injected into the system prompts of some LLMs, where users cannot control it and it changes daily. Testing 9 recent LLMs on 6 datasets, the authors show that changing only the date shifts performance by up to 6% on multiple-choice QA, 14% on math reasoning, 7% on code generation, and 2.84 BLEU on machine translation, and can alter model rankings. The effect exceeds other non-determinism sources, and standard prompting methods do not reduce it, so careful evaluation protocols are needed.
Mario Sanz-Guerrero, Minh Duc Bui, Manuel Mager et al.
In briefSFE-VGGT: Source-free distillation of VGGT for event-camera depth estimation
SummaryThis paper proposes SFE-VGGT, a framework that transfers geometric priors from the VGGT vision model to event cameras without any paired RGB data. It reconstructs surrogate frames from the event stream to serve as a frozen teacher, and adds reliability-aware distillation and a cross-frame consistency loss to handle imperfect supervision. The authors report that on MVSEC nighttime sequences, it reduces average 10 m depth error by 15.3% compared with EventVGGT.
In briefNeuro-Symbolic Computer Use: Reusable Learned Policies for Repeated Computer Tasks
SummaryThis paper proposes running recurring computer workflows with learned, executable code policies instead of having an agent re-plan every step. Stable decisions such as ordering and loops are fixed in code, while observation-dependent steps like grounding and state checks are handled by neural models. The policies are learned through neuro-symbolic policy iteration and guarded by a pre-action verifier during deployment. The authors report the highest Pass^3 across four settings on OSWorld-Verified and ScienceBoard, with lower per-run cost and latency.
WhyCompiles recurring computer workflows into executable neuro-symbolic policies; experiments on two benchmarks show large gains in reliability and cost.
Hyewon Suh, Thanh Minh Nguyen, Chih-Lun Lee et al.
In briefA lightweight adapter that corrects short-term solar power forecasts from frozen models
SummaryThis paper proposes State Transport Routing (STR), a lightweight adapter that refines predictions from a frozen photovoltaic power forecasting model. It blends the original forecast with two trajectories based on the latest measured power and its recent trend, using a horizon-conditioned router over the first 120 minutes. On four public PV datasets, STR outperforms a parameter-matched residual adapter, and on PVDAQ it reduces all-horizon normalized MAE for five neural backbones, while showing no reliable gain for LightGBM.
In briefLearning Humanoid Skills from a Single Failed Human Video
SummaryThis paper proposes TRACC, a method that lets humanoid robots learn skills from one human video of a failed attempt. It imitates the usable motion prefix before failure, then uses a task-completion reward based on the inferred intended outcome to learn the goal. Experiments on six in-the-wild failed tasks from the Oops! dataset show it works without any successful demonstration.
In briefPrecogUI: Proactive GUI Agents Using Pre-cognitive Simulation and Experience Retrieval
SummaryExisting reactive GUI agents often fail in long, dynamic tasks where unexpected disturbances cause cascading errors. The paper proposes PrecogUI, which uses a Proactive Experience Pool, a simulation executor that forecasts the next UI layout for candidate actions, and an execution controller that handles foreseen anomalies with closed-loop error correction. The authors also build InterfereBench, a benchmark for long-horizon tasks with strong disturbances, on which PrecogUI outperforms state-of-the-art methods while remaining competitive on public benchmarks.
In briefA benchmark comparing speech recognition tools for Iberian languages
SummaryThis paper benchmarks eleven automatic speech recognition systems, ten open-weight models and one commercial API, across five Iberian languages, with German and Turkish as controls. The evaluation uses an 85-hour dataset of read speech, broadcast media and audiobooks, measuring word error rate and real-time factors. It finds that no single system dominates, that accuracy, efficiency and language coverage involve clear trade-offs, that low-resource languages such as Basque perform significantly worse, and that most systems show sex-based disparities.
Fernando L\'opez, Pablo G\'omez, David Solans et al.
In briefSeg3DParts: segmentation-guided generation of editable multi-part 3D models from a single image
SummarySeg3DParts is a framework that generates part-level 3D models from a single image, using segmentation as an explicit signal to tie each part to its image region. Cross-part interaction lets components share global context during generation, improving coherence across parts without post-hoc alignment. The authors also introduce PartObjectNet, a dataset with over 200K objects and 1M annotated parts. Experiments reportedly show better geometry quality, cross-part coherence, and controllability than existing methods.
In briefToken Pruning for MLLMs Using Representation Change Signals
SummaryMultimodal large language models face high latency from long visual token sequences, and this paper studies how token representation changes across encoder depth relate to foreground saliency and semantic consistency. It finds that large update magnitudes concentrate on foreground regions in two depth intervals, and that update-direction similarity separates same-class from different-class tokens better than output features. Based on these findings, the authors propose MSDG-Prune, a training-free method that groups tokens by update direction and prunes using update-magnitude saliency. On LLaVA-NeXT it keeps 91.9% of uncompressed performance with 5.6% of visual tokens and a 7.8x prefilling speedup.
WhyAnalyzes layerwise token update magnitudes and directions, linking them to saliency and semantics, and proposes a training-free pruning method with four-MLLM experiments.
In briefAeroManip-VLA: Scalable Learning Benchmark for Aerial Manipulation Using RL-Generated Demonstrations
SummaryThe paper presents AeroManip-VLA, a scalable benchmark for generating data and evaluating policies for vision-language-action models in aerial manipulation. It offers a GPU-accelerated simulator with payload-aware flight and manipulation control, and uses reinforcement learning policies with expert task rules to automatically create demonstrations without teleoperation. Automated event labeling and trajectory categorization are introduced to filter demonstrations and analyze failures. Imitation learning and VLA baselines are evaluated across different task settings to reveal their performance and failure modes.
In briefBellman-certified rounding for sparse policy deployment in MDPs
SummaryThe paper studies how much discounted return is retained when continuous policy mixtures are rounded to sparse binary policies in finite MDPs. It derives reusable envelopes from 2d+2 Bellman solves to give guarantees before rounding, including candidate-specific bounds. On the structured suite, pre-rounding certification coverage rises from 48.2% to 74.1%.
In briefCan Language Models Learn to Forecast Stock Prices: A Paper on Post-Training for Financial Forecasting
SummaryThis paper studies whether post-training can improve a language model's ability to forecast stock prices. In a chronological stock-price sandbox, the model uses tools to gather price, volume, relative-performance, and market-context evidence before predicting a future return. The authors apply supervised fine-tuning on tool-use demonstrations followed by proximal policy optimization with a terminal reward to Qwen3-4B, producing AURA-4B, whose direction-magnitude score rises from 20.94 to 43.31, comparable to frontier models on this benchmark.
In briefBaLEEN: Context-aware speech recognition that biases toward key named entities without retraining the model
SummaryThis paper proposes BaLEEN, a lightweight hypernetwork-based framework that adapts a speech recognition model to context without fine-tuning it. It encodes contextual keywords with a frozen pretrained language model, compresses them via a Perceiver bottleneck, and injects bias vectors into the encoder. Tested on a CTC-based ASR model with a Wikipedia-derived corpus of annotated named entities and synthetic speech, it reports an 8.7% reduction in keyword miss rate, plus 21% and 28% improvements in word and character error rates relative to the unbiased baseline.
In briefVariational Mixtures and Multi-Marginal Flow Matching for Biological Statistical Inference
SummaryThis PhD thesis develops statistical inference methods for biological distributions that are multi-modal, geometrically structured, and sometimes defined only up to a normalizing constant. It progresses from variational inference to more expressive approximations, including ensembles with multiple importance sampling ELBO and variational mixtures that automate component cooperation, along with efficient mixture learning methods. It then extends the work to flow matching, proposing multi-marginal flow matching with mixtures of variational interpolants, and applies these methods to biological problems such as three-dimensional spatial transcriptomics.
In briefSafeVantage: Viewpoint-Aware Memory for Reliable Robot Decisions
SummarySafeVantage is a framework that stores, for each claim, the viewpoints, camera poses, and estimated target location supporting it, and predicts target visibility from reachable viewpoints to choose where to look next. It outputs Yes, No, or Abstain decisions. On 7,424 paired episodes across 232 unseen ProcTHOR houses, it achieved macro-F1 gains of 24.7% at eight actions and 12.0% at twelve actions over equal-budget baselines, with 31.7% less travel at eight actions.
Sean Hardesty Lewis, Zuyi Guo, Benwang Chen et al.
In briefMultiTalk: Extending Full-Duplex Speech Models to Long, Multi-Party, Bilingual Conversation
SummaryThe paper addresses the limits of current full-duplex speech models in long-context and multi-party conversation, extending the Moshi approach to English and Chinese. It releases 57.6k hours of synthetic training data (MultiTalkPT and MultiTalkFT) and MultiTalkBench, an evaluation set built from real human recordings averaging 32.6 minutes per conversation. The authors report that their bilingual model maintains coherent multi-party conversations and outperforms open-source baselines including Moshi, MiniCPM-o-4.5, and Qwen3-Omni-30B-A3B-Instruct on this benchmark.
In briefRAEGNet: A Relation-Aware Evidence Graph Network for Harm-Aware Fake News Detection
SummaryExisting multimodal fake news detectors often retrieve external evidence by entity, which can add event-irrelevant noise and ignore how harmful different fake news items are. This paper proposes ELERF, an event-level evidence retrieval framework, and RAEGNet, a relation-aware evidence graph network. RAEGNet builds a directed graph of news-evidence stance and evidence-evidence relations and adds a conditional-harm branch to jointly model authenticity and potential harm. The authors report that RAEGNet outperforms several baselines on all metrics across Weibo-21, Fakeddit, and their self-built SSS dataset.
In briefSTAR-GRPO: Reliability-first advantage estimation to curb reward hacking in policy optimization
SummaryThis paper studies reward hacking in group-relative policy optimization, where a model exploits a brittle reward interface or permissive proxy to raise training scores without better answers. The authors propose STAR-GRPO, which measures rollout reliability from paired assessments of the same output and uses it to fit a robust location-scale baseline and attenuate unreliable advantages before group normalization. Experiments cover token-interface exploitation and rubric-proxy overoptimization in medical reasoning, reporting better independent semantic evaluation and a smaller proxy-judge gap.
In briefHorizonFlow: Variable-Length Planning for Offline Goal-Conditioned RL
SummaryThis paper introduces HorizonFlow, a hierarchical planner that treats plan length as an output of generation rather than a fixed input. It combines insertion-based generation with flow matching to jointly produce plan content and length, and uses that length to select candidates and favor shorter plans without a separate learned value model. The method is evaluated on Maze2D, Multi2D, and OGBench navigation and visual manipulation benchmarks, where it achieves the highest average performance among the compared methods.
In briefDiffReID: a discriminative diffusion model for object re-identification that learns identity-aware features.
SummaryThis paper proposes DiffReID, a feature learning framework for object re-identification, which retrieves objects across non-overlapping cameras. It uses CLIP-based identity-aware text features and a diffusion model that gradually learns identity-aware distributions and generates identity-invariant features. Experiments on five object ReID benchmarks show better results than most state-of-the-art methods.
In briefMoCo Rubric: momentum-coupled rubric adaptation for detailed image captioning
SummaryThis paper proposes MoCo Rubric, a two-stage framework for detailed image captioning that uses one vision-language model to act as caption policy, rubric generator, and rubric judge. Rubrics are built online from the current policy's captions, reference captions, and image evidence, and only the policy is updated with GRPO. A momentum model, updated by an exponential moving average of policy parameters, is shared by the generator and judge. The authors report an average pairwise win rate of 72.83% across five captioning benchmarks.
In briefHarness Evolution as Learning: Limits of Self-Improving Personal Agents
SummaryThis paper studies how the surrounding harness of a personal agent built on a fixed large language model, which manages context, memory, tools, and execution, evolves to suit individual users. It examines three questions: harness architecture, harness scale, and self-evolution algorithms. The authors introduce a preference-oriented benchmark for empirical analysis and formulate harness evolution as a learning problem. They explain their observations through approximation, generalization, and optimization errors.
In briefProGuT: Label-efficient panoptic segmentation for forest scenes without per-image masks.
SummaryThis paper presents ProGuT (Prototype Guided Training), a method for panoptic segmentation in forest scenes. It needs only unlabeled images and a one-time cluster-to-class mapping, clustering CLIP patch features and recovering trunk instances with a multiscale geometric prior based on structure tensors. On the Our-forest dataset it reaches a Panoptic Quality of 65.2, and on Freiburg Forest 65.9 mIoU, outperforming unsupervised baselines such as PiCIE and STEGO.
In briefSINO: A Scale-Invariant Neural Operator for Learning Low-Rank Closure Terms in Turbulence Problems
SummaryThe paper proposes the Scale-Invariant Neural Operator (SINO), which learns on normalized physical scales using a dual-branch design in spectral and spatial domains. Bottleneck MLPs generate continuous convolution kernels that embed a low-rank inductive bias. The authors report 1.5 to 38 times error reduction and 2 to 23 times parameter efficiency over baselines across several turbulence closure benchmarks.
WhyDual spectral-spatial operator on normalized scales is compared against many baselines across several turbulence closure benchmarks; substantive but combinatorial novelty.
In briefRobust change-point detection for locating human-LLM authorship switches in mixed text
SummaryThis paper addresses locating authorship transitions in documents mixing human and LLM text, using detector scores with varying reliability. It proposes Robust Weighted Profile-Loss Change Point Detection (RWCP), combining capped reliability weights, Huber profile gains, and narrowest-over-threshold search. On five cached-score benchmark families, core RWCP reduces family-macro WindowDiff by 17.6% versus weighted change-point detection, with RWCP-R lowering it further. Both fixed configurations miss changes in collaborative and densely alternating text.
In briefWEFT: A Whole-System Method for Scaling Tool-Use Training in General AI Agents
SummaryThis paper argues that scaling executable environments alone does not reliably improve tool-use models, and proposes WEFT, a framework that coherently optimizes the whole agentic interaction system. WEFT combines scalable system construction, execution-driven self-evolution that revises components based on failure traces, and techniques for stable post-training. The authors report that WEFT-8B and WEFT-14B outperform matched-size environment-scaling baselines on BFCL V4, τ²-Bench, and Claw-Eval, and that WEFT-35B-A3B extends gains to long-horizon workflow benchmarks.
In briefRNA Design Using Conditioned Flow Matching and Reinforcement Learning
SummaryThis paper addresses RNA design, finding sequences that fold into specified secondary structures. The authors propose a two-stage framework: RNA-IFlow models coordinated sequence variation using structure-conditioned Dirichlet flow matching, and RNA-IFlow-RL converts the learned flow into a pairing-preserving finite policy refined with thermodynamic feedback. The method reports 85.19% Pass@1 on Rfam-27, along with analyses of thermodynamic gains, policy dynamics, and robustness.
In briefMatching Model Architecture to Sparse Priors in Tabular Foundation Models
SummaryThis paper studies how tabular foundation models handle irrelevant features. It finds that the row-token model TabDPT degrades much more when null features are added than the cell-token model TabPFN v2. The authors train streamlined row-token and alternating-axis transformers under identical sparse-to-dense linear priors and compare them with the Bayes-optimal predictor. They conclude that preserving an addressable feature axis provides a useful inductive bias for task-adaptive relevance inference.
In briefWeakly Supervised Localization of Diffusion-Edited Fake Regions: The ReGFLoW Method
SummaryThis paper addresses localizing regions edited by AI diffusion models in images, which is important for forensic analysis. The proposed ReGFLoW framework requires only image-level real/fake labels and uses diffusion reconstruction errors as spatial guidance, along with artifact-centric multiple instance learning, avoiding pixel-level masks or semantic pseudo-masks. Experiments report competitive cross-generator localization, and it outperforms all evaluated fully supervised baselines on tests including partially edited and fully synthetic images and on cross-dataset evaluation, without target-domain adaptation.
In briefHow the Observed View Shapes Causal Effect Identification: Evaluating Causal Models Across Observational Views
SummaryThis paper introduces CausalIDView, a benchmark that keeps the underlying causal model and target effect fixed while varying only the observational view available to the estimator, to compare causal foundation models (CFMs). No single CFM performs best across all views, and model rankings shift considerably. Under controlled structural changes, each CFM shows its own failures in keeping estimates stable or tracking real effect changes. A modular method pairing a predictive tabular foundation model with explicit identification procedures is competitive with CFMs and beats several of them.
In briefS4VY: A Segment Anything Model for 4D Instance Segmentation in Dynamic Scenes
SummaryS4VY is a segmentation model built on feed-forward 4D visual geometry that turns a set of RGB observations into class-agnostic 4D instance masks, keeping each object's identity consistent across all observations. It supports prompt-free segmentation as well as point- and box-based selection. The paper also adds an agentic framework for grounding natural-language queries in 4D scenes. The authors report state-of-the-art 4D instance segmentation and strong language-guided grounding across static and dynamic scenes.
In briefSeeing Time: Visual-temporal learning to make time series clusters easier to interpret
SummaryThis paper addresses clustering of multivariate time series and proposes WAVE, which treats each series and its deterministically rendered waveform plot as complementary views of the same observations. It combines fine-grained temporal variations with holistic visual patterns and links each cluster to its centroid-nearest real sample, so clusters can be traced to observable waveform features. Across 10 public real-world datasets, WAVE achieved the highest macro-averaged clustering performance and best average rank among the compared methods.
In briefSTRAT: Predicting Agents' State as Text to Aid Reinforcement Learning
SummaryThis paper proposes STRAT, an auxiliary task that trains reinforcement learning agents to predict a short text describing their own state, including position, inventory, goals, and progress. The text is generated online from environment rules without human labelling, and the method adds one auxiliary head to a standard policy. Across 60 sparse-reward XLand-MiniGrid tasks, the authors report that STRAT solves complex environments where standard RL fails and produces more compact state representations.
Muhammad U. Nasir, Alex Vogt, Steven D. James et al.
In briefS2T-Unet: A Structure-to-Style Framework for Translating Between MRI Modalities
SummaryThis paper proposes S2T-Unet, a framework that synthesizes missing MRI modalities from available scans to reduce extra scanning. It uses vector quantization with a learned discrete codebook at the lower-level bottleneck to encode modality-invariant structure, and a modality transformation module at higher levels to shift representations toward the target modality's intensity and contrast. Experiments on the IXI multi-contrast MRI dataset across four translation tasks show results comparable or superior to state-of-the-art methods.
In briefLocating Where Models Change Their Minds for Efficient RL Training
SummaryThis paper proposes Hindsight-Divergence Localization (HDL), a method for reinforcement learning with verifiable rewards that uses changes in token log-likelihoods induced by hindsight feedback to select branch points. HDL generates a few complete root trajectories and fills training groups with continuations from the selected positions, reusing each root prefix to reduce generation cost. Experiments across math, code, and agent tasks report up to 2.5x fewer generated tokens and 1.8x faster rollout wall-clock time than GRPO at matched settings, with performance gains in all three domains.
Fanchao Chen, Hengyu Fu, Shivaram Venkataraman et al.
In briefIronLLM: Compact Language Models Built for Real-Time On-Device Use
SummaryThe paper presents IronLLM-0.6B, a 654M-parameter language model designed for efficient on-device inference, with a decoding speedup of 1.48x from a shared-KV multi-token prediction design. It is pretrained on about 6.2 trillion tokens and post-trained with multi-domain on-policy distillation from domain-specialized teachers. The authors report competitive performance against larger models such as Qwen3.5-0.8B and MiniCPM5-1B, and also present IronLLM-0.6B-Light, which replaces RMSNorm with Dynamic Tanh to improve efficiency.
In briefMarkovian Nonconvex ADMM for RL: Bellman-Resolvent Stability Beyond Smooth Blocks
SummaryThis paper studies a structural mechanism for Markovian nonconvex ADMM in reinforcement learning, using finite discounted MDPs as a test case. It shows that the discounted Bellman resolvent can supply the multiplier stability that classical analyses get from a smooth block. The authors prove convergence under controlled Markov sampling and stochastic observations, and give iteration and sample complexity bounds along with global optimality results for covered KKT points.
In briefHow wrong upstream messages can override correct LLM answers in multi-agent systems
SummaryThis paper studies whether a wrong or useful message from one AI agent can make a downstream agent abandon a correct answer its own evidence supports. The authors ran controlled experiments on five benchmarks and five receivers, comparing no message, the original message, and a message with the opposite conclusion. They found messages sometimes help and sometimes hurt, and in 94% of audited harmful cases the downstream agent copied the upstream's wrong answer.
In briefSocial anchors for group-aware trajectory prediction
SummaryThis paper addresses trajectory prediction in dynamic scenes, noting that existing group modeling often uses fixed thresholds and relative positions within the observation window, ignoring that grouping should vary by agent and context. The authors propose Socialality, a human-inspired framework with interpretable social anchors and an extended grouping window that learns agent-specific grouping rules from historical observations and short-term future trajectory previews. Experiments on standard benchmarks show performance gains, and qualitative analyses and statistical studies of anchor distributions are used to verify interpretability and stability.
In briefRoXDrive: Closed-loop reinforcement learning for self-driving that keeps only action-consistent simulated futures
SummaryThis paper proposes RoXDrive, a closed-loop reinforcement learning framework for end-to-end autonomous driving policies. It uses video world models to generate future scene rollouts and an Action-Vision Faithfulness Evaluator to keep only rollouts whose visual dynamics match the ego actions, then applies safety-aware scoring for policy post-training. Experiments on nuScenes and an in-house dataset report safety violation reductions of 27.6% with DiffusionDrive and 33.7% with Qwen3-VL.
In briefRethinking fake news detection for the generative AI era
SummaryThe paper notes that generated content increasingly takes part in producing and spreading news, making fake news more complex. The authors build Weibo26, a multimodal fake news dataset for scenarios involving generated content, and propose the GAHR framework, which combines global judgment with local correction to incorporate generativity into veracity reasoning. Experiments on several benchmarks and Weibo26 report competitive veracity detection and effective identification of generative content.
In briefA Robust Framework for Wearable-Sensor Human Activity Recognition
SummaryThis paper proposes ALAE-TAE-CutMix+, a framework for recognizing human activities from wearable sensor data. It enhances latent features of each sensor channel and introduces a new augmentation strategy for multi-sensor data. An extended version, ALAE-CIE-TAE-CutMix+, models interactions between pairs of sensor channels and is reported to be more reliable and robust. The authors state both frameworks significantly outperform state-of-the-art methods on four HAR datasets from diverse domains.
Nafees Ahmad, Ho-fung Leung, Muhammad Adil Abid et al.
In briefAutomated Screw Planning for Reduced Pelvic Fractures Using Statistical Shape Models and Deep Learning
SummaryThis paper proposes a fully automated pipeline for preoperative iliosacral screw planning in pelvic fracture surgery. Using patient-specific three-dimensional anatomy, it identifies safe screw corridors and generates individualized insertion trajectories. Evaluated on 200 clinical pelvic fracture cases, it reportedly increased the safety margin by 2% across four screw types, cut mean planning time by over 90%, and achieved a 95% clinical acceptance rate.
In briefDSWM: A decomposed spatio-temporal world model for repositioning drone base stations based on demand
SummaryThis paper proposes DSWM, a model for repositioning uncrewed aerial vehicle base stations (UAV-BSs) to follow traffic demand that changes over space and time. It plans decisions in a latent space, using a rolling observation window, a recurrent state, and imagined rollouts that penalize uncertainty. Tested on three real datasets (Milan, Shanghai Telecom, YJMob100K) against 14 methods, DSWM achieves the highest weekday served ratio on each dataset. On Milan it beats the strongest non-learning baseline, Greedy, by 0.109.
Shengjie Zhong, Zhongliang Zhao, Jingxuan Chen et al.
In briefGlassFormer: Real-time glass segmentation by fusing radar and depth sensing
SummaryThis paper proposes a multimodal framework that combines millimetre-wave radar with RGB-D sensing to segment transparent surfaces in real time. Radar reflects strongly off glass, supplying a geometric cue where cameras and depth sensors often fail, and this cue is fed into a lightweight transformer network, GlassFormer, via cross-modal attention. The authors report 0.88 mIoU on a mixed-condition split and 0.59 mIoU on a low-light split, both improving on vision-only baselines. They also state the system runs in real time on resource-constrained platforms.
Suhani Grover, Astik Srivastava, Viswas Dinesh et al.
In briefRolloutFaith: Auditing Lasting Internal Edits in Visual World Models
SummaryThis paper proposes RolloutFaith, a framework that checks whether an internal correction to a visual world model still improves semantic accuracy in later autonomous predictions, not just the immediate one. Ten fitted editors are tested on three world models across Crafter, Cartpole, and CoinRun, with Reference Activation Patching used as a benchmark. Current fitted editors recover only limited and inconsistent long-term effects. The authors also introduce Delayed LoReFT, which optimizes the intervention through four frozen future transitions and improves sustained effects to some extent.
In briefDo VGGT-Family Models Need All Their Layers? A Study of Layer Redundancy and Pruning
SummaryThis study examines which layers of the feed-forward geometry models VGGT, π³, and VGGT-Ω are needed to preserve camera poses and dense 3D structure. It evaluates 3,018 pruned configurations on seven metrics across four indoor and outdoor datasets, finding removable layers concentrated in an early region and a narrower late region. A closed-form linear calibration recovers accuracy after pruning without end-to-end retraining. The resulting models reduce aggregator parameters by up to 44% while keeping accuracy comparable to the intact models.
WhySystematically studies layer redundancy in VGGT-family models, validated across 3,018 configurations and datasets, with practical closed-form calibration.
In briefOn-policy distillation that reflects on visual evidence to train visual agents
SummaryVisual agents alternate between reasoning and image operations, so early errors in acquiring, reading, or grounding visual evidence can propagate through the trajectory. The paper introduces ReVuE, which compares multiple student trajectories for the same query, summarizes observed visual evidence, and diagnoses the first failure across the Acquire, Read, and Ground stages. These reflections give the teacher training-time context, and token-level distillation losses are grouped and reweighted accordingly. Across 11 benchmarks with Qwen2.5-VL and InternVL3.5 models, ReVuE achieves higher weighted-average scores than all evaluated OPD baselines and reduces redundant reasoning and tool calls.
WhyDiagnosing stage-wise visual evidence failures and converting them into token-level distillation supervision is a meaningful advance; experiments across 11 benchmarks are fairly thorough.
In briefYou Cannot Recover What Was Never Measured: The Information Limit of Ultra-Low-Field MRI Super-Resolution
SummaryThe study measures how much subject-specific information real 64 mT portable MRI scans actually contain, using paired scans of the same subjects at 3T. It finds that whole-brain structure specific to each individual is preserved only down to about 3 to 4 mm in-plane, and coarser through-plane. The authors report that fine detail produced by trained super-resolution models is no more correlated with a subject's own 3T scan than with a stranger's, and that PSNR and SSIM cannot separate recovery from fabrication.
WhyQuantifies subject-specific information in real paired 64 mT and 3T scans, validated on known-answer tests, showing synthetic degradations overestimate recoverable detail across 24 trained models.
Prathamesh Pradeep Khole, Shreya Handa, Utkarsh Gupta et al.
In briefARC-KV: Amortizing Anchor Search for Efficient KV Cache Compression
SummaryLong-context language model inference is slowed by KV caches that grow with sequence length. The paper proposes ARC-KV, which trains a value-aware indexer to pick real-key anchors in a single scoring pass, then merges keys under convex-hull constraints and fits attention-mass biases and compact values. The compact cache is built once per context and reused for all later queries. On Llama-3.1-8B-Instruct, ARC-KV outperforms reported compaction methods in most settings across QuALITY, RULER, and LongBench; at 10% KV retention on QuALITY, accuracy rises from 0.6409 to 0.6474 while compaction time drops from 959.8 s to 37.3 s.
In briefErasing Motion Concepts from Video Diffusion Models (MUTE)
SummaryText-to-video diffusion models can realistically depict actions such as kicking or shooting, raising safety concerns that motivate targeted concept erasure. This paper systematically studies erasing motion concepts from video Diffusion Transformers and proposes MUTE, a training-free method that modifies no weights and corrects the velocity output at each denoising step. Experiments on 20 motion concepts show MUTE outperforms representative prompt-level, weight-level, and inference-time baselines on Wan2.1-T2V, and the formulation transfers to CogVideoX.
In briefWhere Staleness Builds Up: A Pool-Aware Way to Control Lag in Async RL for LLM Training
SummaryThe paper studies how policy lag accumulates in asynchronous reinforcement learning for large language model post-training, splitting trajectory staleness into Generation Staleness and Waiting Staleness. It proposes PACE, which uses an adaptive rejection budget and an effective-staleness score to rank completed trajectories. The authors report an 18.7% gain in six-benchmark average validation accuracy over unfiltered asynchronous RL on single-turn math reasoning at the same wall-clock budget, matching synchronous RL with 47.1% less GPU time.
