In briefSharpness-aware Orthogonalized Momentum Algorithm (SOMA) for efficient LLM training
SummaryThis paper proposes SOMA, which combines sharpness-aware minimization's adversarial perturbations with the Muon optimizer, using stale stochastic gradient noise to keep extra computation low. The authors show theoretically that SOMA's sample complexity is invariant to the perturbation direction. They also study its robustness to heavy-tailed gradient noise. Experiments examine its effect on quality loss after post-training quantization and on catastrophic forgetting during LLM fine-tuning.
Geetika, Shivani Singh Chauhan, Gauranshi Gupta et al.
In briefReading Between the Layers: How Vision Transformers Align Text and Meaning
SummaryThis paper studies how transformer-based vision encoders process rendered text and its meaning, using cross-modal alignment, probing, and retrieval analyses. Layer-wise results show strong cross-modal mapping appears only in the deepest layers, after object and text recognition, as representations shift from surface-level to semantic. This hierarchy is strongest in contrastively trained models, while larger self-supervised models and text-rich training data partially reproduce similar behavior without explicit language supervision.
In briefDualGSL: Prior-guided graph structure learning with node-adaptive prediction fusion
SummaryDualGSL is a framework for node classification on graphs that learns a separate weight for each node to combine predictions from two propagation paths. One path uses a feature-based graph with virtual class nodes to incorporate training labels, while the other combines learned affinity graphs with that feature-graph prior. Across seven baselines on the AMG18, 20News and USPS benchmarks, DualGSL achieves the highest cross-dataset average accuracy and Macro-F1. Ablation studies on AMG18 show the benefit of fusing the two propagation paths through logits.
In briefMDMR-Bench: A multi-dimensional benchmark for generating images from multiple references
SummaryThis paper introduces MDMR-Bench, a benchmark for testing image generation models that take several reference images and text instructions. It has four evaluation subsets covering entity fusion and attribute transfer, multi-element extraction from one image, infographic manipulation, and visual reasoning-intensive generation. The authors also propose a checklist-based protocol that breaks quality into global and atomic-level questions, and use it to evaluate open-source and proprietary models. The dataset and code are publicly available.
In briefAutoMIL: Automated Multiple Instance Learning for Whole-Slide Image Classification
SummaryThis paper presents AutoMIL, an open-source framework that uses automated machine learning to handle preprocessing, training, and evaluation for classifying whole-slide images. With default settings, it reached ROC-AUC scores of 0.94 on UBC-OCEAN, 0.99 on PCam, and 0.93 to 0.96 on a TCGA-NSCLC subset, comparable to specialised pipelines. The authors report that ensembling mainly improved results on imbalanced data, so an automated pipeline must adapt to dataset characteristics.
In briefDIGOM: a graph model that captures relationships across patients and data types for clinical outcome prediction.
SummaryThis paper proposes DIGOM, a graph-based framework that jointly models three interaction types among multimodal patient data: intra-patient cross-modal, cross-patient same-modal, and cross-patient cross-modal. It learns sparse, disentangled graph views via a differentiable sampling mechanism. The authors evaluate it on MIMIC-IV and NIH-TB across five clinical prediction tasks, reporting the highest AUROC on all five and greater robustness to missing modalities.
Hakima Laribi, Olivier Lefebvre, Dan Poenaru et al.
In briefPOSEIDON: Automatically Building Verifiable Training Environments for Science Agents from Code
SummaryThe paper introduces POSEIDON, a pipeline that generates diverse, verifiable training environments from scientific code repositories. It decomposes scientific workflows into four task types (Reproduction, Subflow, Completion, Correction) and uses a recursive improvement loop to recover failed candidates. The resulting Pearl collection contains 97,659 environments from 7,110 repositories across 10 disciplines. Qwen3.5 models trained on Pearl show consistent gains on six scientific-agent benchmarks.
In briefSeparating Patient Personas from Clinical Workflows for Realistic Medical Dialogue Generation
SummaryThis paper introduces CoRPS-Dial, a controllable dual-agent framework that synthesizes pre-consultation medical dialogues. A persona-driven patient agent and a process-guided doctor agent interact to generate dialogues, followed by quality control for role, workflow, and textual consistency. Experiments on coronary heart disease show closer alignment with real dialogues in information coverage, clinical compliance, and naturalness, and downstream fine-tuning confirms the synthetic data's utility.
In briefFine-grained image classification that uses text only during training, then distills it into image-only models
SummaryThis paper proposes FG-LUPI, a framework for fine-grained visual classification that tells apart very similar categories. During training it uses textual attribute descriptions as privileged information, fusing them with image patches in a multimodal teacher via gated cross-attention. The knowledge is distilled at prediction, patch, and attention levels into an image-only student. Experiments on multiple benchmarks report higher accuracy than strong ViT baselines and lower Expected Calibration Error, while inference requires only images.
Aishwarya Venkataramanan, Christian Ickler, Joachim Denzler
In briefSCI-AGENT-EXAM: A Benchmark Testing AI Agents on Science Research Tasks
SummaryThis paper introduces SCI-AGENT-EXAM, a benchmark of 120 expert-written tasks covering 11 scientific fields and 7 research task types. Tasks require both executable outputs, such as code, and textual outputs, such as reports, and are scored with programmatic checks plus an expert-validated LLM judge. Tests on 10 frontier LLM agents show large performance gaps, with the top overall score at 80.9 while only 26.0% of that agent's runs exceed the 0.95 success threshold.
In briefPROVE: Probabilistic Visual Grounding via Patch-Level Evidence for Vision-Language Models
SummaryThis paper introduces PROVE, a post-hoc adapter that sits on top of frozen CLIP models and maps each image patch to a Beta distribution, giving both a localisation map and an uncertainty estimate. It learns per-patch evidence for whether a patch belongs to the referent, and pulls absent-referent cases toward an uninformative prior so that uncertainty signals absence. The authors report that localisation is competitive with existing CLIP grounding methods, uncertainty tracks localisation error, and the image-level score detects when the referent is missing.
In briefData quality audit and temperature-humidity analysis of a long-term LoRaWAN network
SummaryThis paper processes a long-term environmental LoRaWAN dataset from Tour Perret in Paris, containing about 422,000 frames from 2021 to 2023, using a C++17 pipeline. A coverage audit showed CO2 was entirely missing and pressure came from only one base station, so the analysis focused on temperature and humidity. Methods include anomaly detection, principal component analysis, clustering, and a neural network; the temperature-humidity correlation is -0.662 and the cross-validated R² is 0.672.
In briefSpectrally Tuned Implicit Neural Representations for Intracranial 4D Flow MRI Reconstruction
SummaryThis paper proposes an unsupervised reconstruction framework for intracranial 4D Flow MRI that represents the velocity field with an implicit neural representation using tunable spectral bias. It was tested on nine cerebrovascular cases, including healthy controls and moderate-to-severe intracranial atherosclerotic disease. Cohort-optimal spectral settings generalized across subjects with small error increases, and the data-driven variant reached near-comparable accuracy within minutes. The method stayed robust to realistic noise and moderate VENC-dependent aliasing.
Oliver Welin Odeback, Nerea Gonz\'alez-Aranceta, Sebasti\'an Jofre et al.
In briefProfile-Driven User Simulation with Hierarchical Reinforcement Learning Rewards for Task-Oriented Dialogue
SummaryThis paper presents ProfileSim, a framework that simulates users for financial marketing outbound calls to address the scarcity of domain-specific task-oriented dialogue data. It jointly models user profiles (demographics and Big Five traits) and business intents reverse-engineered from real call logs, and refines the model with GRPO using hierarchical dialogue- and utterance-level rewards. The authors report that ProfileSim achieves the highest colloquial naturalness and profile adherence among the evaluated user generators.
In briefTempo-Zero: Self-Play in Executable Worlds for Video Reasoning
SummaryThis paper proposes Tempo-Zero, a self-improvement framework for video reasoning. A model generates questions and answers them inside video environments rendered by executable programs, and it is rewarded by execution-verified ground truth rather than majority voting. The method also evolves tasks, difficulty, and environment rules over rounds. Across twelve real-world video benchmarks, it outperforms base models and recent self-play baselines on average, with especially strong gains on temporal grounding tasks.
In briefWhen to Stop Penalizing Actions: Martingale GRPO for Robot Vision-Language-Action Learning
SummaryThe paper observes that in GRPO, a failed vision-language-action robot rollout penalizes every action equally, including helpful recovery attempts. The authors derive stopping conditions for failure feedback using success-probability and policy-score martingales. They propose MART-GRPO, which uses a fine-tuned vision-language model to localize the last visible error and keeps failure feedback only through that interval. Experiments with OpenVLA-OFT on RoboTwin 2.0 report higher success rates than full-trajectory GRPO across several tasks.
Zongyao Zhao, Cheng-Jie Zhao, Ruihang Jiang et al.
In briefAsync-RBench: Benchmarking How LLM Agents Handle Asynchronous Task Results
SummaryThis paper introduces Async-RBench, a benchmark with 200 executable tasks across eight scenarios for testing how LLM agents handle results from subagents that arrive asynchronously. It defines an Event Response Contract to specify what each response must change, preserve, avoid, and re-verify, and a Dynamic Replanning Score reported separately from task success. Evaluating nine agents, the authors find DRS scores of 19.2 to 45.1, and lower task success under asynchronous delivery than batched delivery (26.0% vs. 29.7% on average).
In briefCycle-of-Thought: Self-Checking Iterative Reasoning for Table Questions in Document Images
SummaryThis paper addresses the inconsistent reasoning of vision-language models over tables, which often suffers from spatial misalignment and logical hallucinations. It proposes Self-RefineVQA, a training-free, inference-time method in which one model iteratively acts as generator, critic, and refiner through prompt-guided self-feedback. Experiments on tabular VQA benchmarks, in monolingual and cross-lingual settings, report consistent accuracy gains, with particularly large improvements for small models.
In briefLimits of linear speedup for stochastic Push-Pull on time-varying directed networks
SummaryThis paper studies whether stochastic Push-Pull, a distributed optimization method, keeps its linear speedup in the network size n when the communication network changes over time. The authors build a quadratic problem on a period-two network where the speedup fails. For periodic networks they define an effective node count n_eff≤n and, for smooth nonconvex objectives, prove matching upper and lower convergence bounds of order Θ(1/√(n_eff T)). Numerical experiments support these findings.
WhyConstructs a periodic-network counterexample with matching bounds showing lost linear speedup, backed by proofs and numerical experiments; influential within its subfield.
Feiming Wang, Yutong He, Sulaiman Alghunaim et al.
In briefCan LLM Annotators Match Human Consistency? Benchmarking Human Agreement and Optimizing Multi-Agent Annotation
SummaryThis paper introduces ANT-A, a benchmark of 7 cross-domain datasets in which each instance has five independent human labels, to test how closely LLM annotations match the spread of human judgments. It uses Fleiss kappa to measure alignment with human label distributions rather than a single gold label. The authors also propose ANT-OR, a multi-agent framework that refines prompts by distilling contrastive decision rules from feedback. They report that ANT-OR improves agreement accuracy over the strongest baseline and achieves the highest Fleiss kappa alignment with human annotators.
In briefWhen Dynamics Randomization Fails to Test Context Inference
SummaryThis paper asks whether randomized-dynamics environments actually require explicit context inference in world models. It defines the remaining one-step information as predictive headroom and analyzes discrete and continuous environments. Headroom is low when transition kernels barely differ or when visited states already reveal the context, and resetting state components increases it. The authors propose reporting headroom alongside the randomized parameters.
Oleg Smirnov, Sofiane Ennadir, John Pertoft et al.
In briefAlphaAnything: Extending Pretrained RGB Image Generators to RGBA (Images with Transparency)
SummaryThis paper presents AlphaAnything, a lightweight framework that extends pretrained RGB image generators to produce RGBA images with transparency, without an extra RGBA reconstruction pretraining stage. It inserts LoRA modules into the generative transformer and adds an alpha branch to the frozen RGB VAE decoder, keeping RGB decoding unchanged for any fixed latent. The authors also introduce Endpoint-Aware Alpha Loss (EA Loss), and report competitive results on text-to-RGBA and image-to-RGBA tasks.
In briefLeafSkill: Efficient Leaf-Level Action Grounding for Recursive LLM Agents
SummaryThe paper notes that recursive LLM agents often re-invoke a large model to turn familiar atomic subtask instructions into executable actions. It introduces LeafSkill, an external layer that retrieves related grounding experience, binds action templates to current objects, and uses a 1.7B selector to choose a bound action or defer to the original agent. On an independent set of 200 Robotouille tasks, LeafSkill completes 137 tasks versus 121 for ReCAP, and on the 106 tasks both completed it reduces Qwen3.5-27B token use and calls.
In briefA broader look at retrieval attention heads in multimodal LLMs
SummaryThis paper studies how multimodal large language models find relevant evidence in long inputs, using a sparse set of retrieval attention heads. Adding audio-visual input, the authors find these heads remain sparse and are largely shared across modalities. Ablating them removes nearly all audio-visual retrieval, and training them with a ranking objective improves document retrieval.
In briefTIE: Training-free CLIP embedding guidance to fix missing or mixed objects in text-to-image generation
SummaryText-to-image diffusion models often miss objects or assign attributes to the wrong object, which the authors link to CLIP's limited fine-grained text understanding. The paper analyzes text embeddings and proposes TIE, a training-free method that makes a single embedding update at inference time. Experiments report improved semantic fidelity in multi-concept prompts with minimal extra sampling cost.
In briefError analysis of a shuffled-window activity recognition system from the 3rd WEAR Challenge
SummaryThis paper empirically analyses a system built for the 3rd WEAR Dataset Challenge, which asks for activity labels on shuffled one-second windows from four unseen participants. The approach stacks window classifiers with video-neighbour features and uses a mean-field decoder with a soft per-activity count constraint. In simulation, stacking added 0.051 macro F1 and the count constraint about 0.011, while seam matching had essentially no effect. The authors note the protocol is transductive rather than fully inductive, and the selected submission scored 0.9231 on the private leaderboard.
In briefInteractive image diagnosis via test-time backpropagation from auxiliary tasks
SummaryThis paper proposes multi-task gradient assist (MGA), a method that adds interactivity to multi-task computer-aided diagnosis models without additional training. Radiologists can give feedback on auxiliary-task predictions and request further diagnostic suggestions from the system. Tested on mammography and CT datasets, MGA improved diagnostic accuracy after feedback on various auxiliary tasks, performed more robustly than an existing forward-propagation-only method, and could be combined with it for further gains.
In briefA Measure-Geometric Theory of Robustness in Machine Learning
SummaryThis paper proposes a foundational measure-geometric theory of robustness in machine learning. Robustness is framed as escape from a retained task state, defined by a reference measure, a perturbation gauge, and an admissible realization mechanism. The theory separates the structural possibility of escape on the reference manifold from its realization under observable readouts, and introduces a mesoscopic organization of escape geometry with explicit laws for related but distinct phenomena.
In briefPULSES: A multi-agent AI system that automates building and running excited-state molecular models
SummaryThe paper introduces PULSES, a multi-agent system that automates the workflow for building machine-learning models of molecular excited states, covering conformational sampling, quantum-chemical labelling, training, validation and dynamics. It operates within boundaries set by a human expert and planner. The authors evaluate it on 15 computational chemistry tasks and show end-to-end applications to perylene exciton Hamiltonians, solvent-dependent biphenyl relaxation, and excited-state proton transfer in 10-hydroxybenzo[h]quinoline. The authors emphasize that scientific validation is required rather than assumed from successful execution.
In briefPASTA: A training-free diffusion method that keeps designated blank areas clean while generating images
SummaryDiffusion models generate high-quality images, but visual design often requires keeping areas clear for headlines, logos, or UI elements. Existing layout-aware methods struggle to keep these reserved regions strictly free of unwanted content. PASTA is a training-free framework that applies semantic guidance only to regions where objects are allowed, using a shared latent trajectory guided by masks. Experiments show it improves reserved-region fidelity while preserving overall realism and perceptual consistency.
In briefKnitTrans: A system that coordinates trajectory transfer to speed up LLM reinforcement learning
SummaryThis paper presents KnitTrans, a system that coordinates how trajectory data moves from rollout generation to training in large language model reinforcement learning. It uses constraint-preserving mini-batch admission, on-demand trajectory data access, and multimodal encoding deduplication. The authors report up to 93% lower architectural staleness compared with train-batch admission, and on multimodal workloads, transfer latency reduced by two orders of magnitude with up to 5.8x speedup in transfer plus training.
In briefComparing Audio and Transcript Signals for Native Language Identification of Learners
SummaryThis paper studies identifying the native language of English learners from their speech and transcripts. It compares a WavLM-based audio classifier with a zero-shot GPT-4o classifier on transcripts, analysing results across CEFR proficiency levels. With full transcripts the audio classifier performs better overall, though relative strengths vary by proficiency. Transcript accuracy rises with proficiency, while audio performance stays comparatively stable.
Ahmet Yavuz Uluslu, Mengjie Qian, Kate Knill et al.
In briefUniWorld: A Unified World Model for Real-Time Visual World Simulation
SummaryThis paper presents UniWorld, a framework that combines depth trajectories and reference appearance within one shared video generation model to simulate visual worlds continuously. It uses noise-aligned overlapping continuation and a rolling depth latent state to keep generated history intact while limiting the active depth state to the window size. The authors report that the steady-state generation path reaches real-time throughput and that cached and incremental execution produce bitwise-identical outputs.
In briefDealsFlow: An evidence-gated AI agent system for screening venture capital deals
SummaryThis paper presents DealsFlow, a human-in-the-loop AI system deployed at a UK early-stage fund to triage startup pitch decks. An LLM reads each deck, cites evidence and rates categories, while code combines and caps the ratings so they can only be lowered. On 142 investments at least three years old, the full pipeline's scores correlated somewhat better with total-value multiple and identified write-offs more accurately than an unconstrained LLM. The authors note retrieval was not date-restricted, so screening skill may be overstated.
In briefSam-Correction: a fully adaptive method for reducing label noise in medical image segmentation
SummaryMedical image segmentation ground-truth masks often contain noise because consistent annotation is difficult. The paper proposes SAM-Correction, which uses the Segment Anything Model (SAM) to progressively correct noisy masks, with prompts supplied by another segmentation network trained alongside it. This removes the need for manually prepared well-curated masks. Experiments with label noise derived from real human annotations show precise noise reduction and well-performing segmentation models, even when clean labels are entirely unavailable.
In briefEigenGradCAM with AugSmooth: More Explainable and Precise Multi-Stage Alzheimer's Detection
SummaryThis paper presents an explainable deep learning method that classifies Alzheimer's disease into four stages from 2D axial MRI images. It combines eigenvector-based Grad-CAM with gradient smoothing and uses WGAN-GP to generate synthetic samples that balance the Moderate Dementia class. On the Best Alzheimer MRI Dataset, the model reaches 97.2% test accuracy and a Matthews correlation coefficient of 0.944, with heatmaps highlighting the hippocampus and entorhinal cortex. The authors note that further clinical validation is needed.
In briefPast as State, Present as Attention: Long-Horizon Speech-Driven Motion Generation
SummaryThis paper addresses long-horizon co-speech motion generation, which produces human body motion synchronized with speech. The proposed PASPA method summarizes past motion style with an inter-block persistent state and models the current segment with intra-block bidirectional self-attention. It is built on a blockwise flow matching framework and reports state-of-the-art FGD on public benchmarks while preserving motion style over long sequences.
In briefAdaptive control of scientific instruments whose responses drift with operating history
SummaryThis paper studies sequential tuning of scientific instruments whose responses change with operational history, using shot-to-shot tuning of neutral beam injection ion sources in fusion devices as a representative case. The authors propose a multimodal agent with auditable components: deterministic extraction of physical symbols and visual evidence, a two-layer memory feeding a multimodal large language model that proposes actions, and an independent feasibility check before execution. Under supervised operation on five ion sources, the system achieved an 86.80% beam-extraction rate across 409 executed shots, which the authors present as evidence of engineering feasibility.
In briefRegularity Condition Ensuring CPAB Transformations Are Diffeomorphisms
SummaryThis note examines CPAB (Continuous Piecewise-Affine Based) transformations, flows obtained by integrating continuous piecewise-affine velocity fields. It points out an overlooked point in the original papers: CPAB transformations are always bi-Lipschitz homeomorphisms but not necessarily diffeomorphisms. The authors identify sufficient regularity conditions that guarantee a CPAB transformation is a diffeomorphism. They show these conditions are easy to enforce in 1D and 2D and hold with probability one for common tessellations under standard sampling assumptions, with analogous behavior expected in 3D.
In briefFinding Hidden Subgroup Performance Gaps in Clinical AI: Testing whether algorithmically discovered patient groups reveal unequal model performance.
SummaryClinical machine learning models often perform unevenly across patient groups, but traditional subgroup analysis needs rich metadata that is rarely available. This study evaluates algorithmic subgroup discovery for performance monitoring, using cardiomegaly classification on 223,228 chest X-rays. The discovered subgroups showed larger performance gaps than metadata-based analysis and were often interpretable through patient, acquisition, and model metadata. As a proof of concept, annotations from the learned subgroups were used to help mitigate the biases they exposed.
Alceu Bissoto, Trung-Dung Hoang, Tim Fl\"uhmann et al.
In briefFOCUS: A Lightweight Mamba-2 Model for EEG-Based Cognitive Workload Classification
SummaryThis paper introduces FOCUS, a compact Mamba-2-based model that classifies cognitive workload from EEG signals under limited data. It uses a fixed FFT decomposition covering the five canonical EEG bands and has about 0.04 million parameters. In a pilot study with 11 participants, it reached 74.75% accuracy on the full dataset, offering a favorable trade-off between performance and computational cost in this controlled setting.
Ioannis Kyprakis, Stefanos Gkikas, Chariklia Chatzaki et al.
In briefSearch-based tool testing which input type (image or text) vision-language models are most sensitive to
SummaryThe paper presents MoST, a black-box testing framework that compares image-only, text-only, and joint perturbations to characterize how sensitive vision-language models are to each modality. It uses multi-objective search to select, combine, and adjust the severity of perturbations, aiming to degrade predictions while keeping input changes small. Evaluated on four VLMs performing object localization on ImageNet and driving images, the models show distinct sensitivities, and in several configurations joint perturbations match or exceed single-modality effectiveness.
Oliver Wei{\ss}l, Eduardo Breitembach, Andrea Stocco
In briefWhy full-duplex speech models break down in long conversations
SummaryFull-duplex speech models can listen and speak simultaneously, but because they generate a token every frame, they exhaust their context windows within minutes. The authors introduce FD-LHB, a long-horizon benchmark, and find that all five evaluated models eventually degrade or run out of context. For sliding-window models, they attribute gibberish and silence to evicting the first KV entry, and repetition to the model's own recent outputs. Two training-free fixes, Initialization-Anchor Pinning and Duplex Repetition Penalty, substantially reduce conversational collapse in replays of up to an hour.
In briefOptimal multiclass online regret for arbitrary label spaces
SummaryThe paper determines the optimal regret for agnostic multiclass online learning over arbitrary, possibly infinite, label spaces. For hypothesis classes with finite multiclass Littlestone dimension d, it proves an upper bound of O(min{T, √(dT)}) over T rounds, matching a known lower bound. This answers a question posed by Hanneke et al. (2023) and extends the optimal binary regret bound of Alon et al. (2021).
WhyResolves an open question by extending the optimal binary regret bound to arbitrary multiclass label spaces, matching the known lower bound via a proof.
In briefTENET: A Time-point Encoding Network for Spotting Anomalies in Multivariate Time Series
SummaryThis paper addresses unsupervised anomaly detection in multivariate time series, where local variation can distort reconstruction and forecasting errors. The authors reframe one-class detection as estimating whether the current time point's system state is compatible with normal behavior, and propose TENET, which builds a latent representation centered on the current time point and scores it with a fixed standard-normal negative log-likelihood. The authors report state-of-the-art results on benchmark datasets.
In briefBeyond Final Outcomes: Measuring How Claims Shift Across Multi-Agent AI Conversations
SummaryThis paper studies how information is shared, expanded, and forgotten when several LLM-powered agents interact, rather than judging only final task results. It introduces RIVER, which splits agent outputs into atomic claims and scores convergence, exploration, and retention. Across five interaction strategies tested on 40 controlled news-generation cases, repeatedly pooled interaction increased convergence but narrowed the information space, while distributed relay produced broader, less shared information. The authors argue that collective information dynamics are better explained by underlying information-flow mechanisms than by downstream performance alone.
In brief4DMemory: Testing long-term spatiotemporal memory in AI agents with added tools
SummaryThe paper introduces LongVideo, a dataset of about 900 hours of video with over 5,000 manually reviewed questions, and 4DMemoryBench for testing long-horizon spatiotemporal reasoning. It also presents the 4DMemory Harness, which combines four complementary skills, 22+ specialized tool models, and a model-specific retrieval-augmented generation system. Tests on 13 leading models show standalone foundation models struggle on long-horizon tasks, while the harness improves them.
In briefWhen to Trust Memory: Retrieval-guided probabilistic spatiotemporal forecasting under distribution shift
SummaryThis paper proposes REMIND, a framework for probabilistic spatiotemporal forecasting when data distributions shift between training and testing. It uses retrieval to estimate how well current observations are supported by historical patterns, with a memory bank and a dual-branch design. A shift-adaptive gate balances the two branches based on retrieval similarity and memory dispersion. Experiments on real-world datasets report better accuracy and probabilistic prediction quality than state-of-the-art methods.
In briefTGR-EvoDAG: A DAG-based multi-agent framework that automatically evolves code for generative recommendation
SummaryThis paper proposes TGR-EvoDAG, a multi-agent framework in which LLM agents automatically evolve code for generative recommendation pipelines that run in stages (tokenization, pre-training, alignment). It models the evolution as a directed acyclic graph to maintain multiple branches and merge complementary improvements, and adds stage-aware cost planning and downstream co-adaptation. On public Amazon datasets, offline metrics improve by up to 37.4%. In deployment on Tencent's commercial media platform, offline Hit@1 rises 12.16% relative and online click-through rate rises a statistically significant 0.40%.
WhyNovel DAG-based multi-branch code evolution with stage-aware cost planning for generative recommendation; validated on Amazon data and a large industrial deployment with significant online CTR gains.
In briefChoreoStudio: Planning Formation, Timing, and Dancers for Music-Driven Group Dance
SummaryThis paper presents ChoreoStudio, a system for generating music-driven group dance that explicitly plans the formation shape, when formations change, and which dancers execute each transition. Audio is decomposed into rhythmic and structural features, and a Director module outputs formation point sets, a transition gate, and a mobility allocation that condition a diffusion generator. The authors report state-of-the-art results on AIOZ-GDance in motion quality, group organization, and formation dynamics.
In briefFrom Visual Affinity to Learning Credit: Outcome-Conditioned Visual Credit for Multimodal Reinforcement Learning
SummaryThe paper addresses multimodal reasoning under reinforcement learning with verifiable rewards (RLVR), where a single response-level advantage is uniformly assigned to all tokens, ignoring the different roles of visual observations and later reasoning. The authors propose Outcome-Conditioned Visual Credit (OVC), a plug-in framework that separates visual-association measurement from token-level credit assignment, without auxiliary reward models or process annotations. Instantiated through GSPO-token, OVC improves six-benchmark mean accuracy over GSPO across three models, with gains of 1.86 to 2.39 percentage points.
In briefBeyond Gain: Measuring Whether Agents Still Need a Skill
SummaryThe paper addresses how public skill hubs for LLM agents can decide which skills a given model has internalized and can retire, noting that existing rules based on performance gain are unreliable. The authors withhold the skill, score the model on a task synthesized from it, and trust a high unaided score only when another model benefits from the skill. In experiments with four models, the published gain rule wrongly removed over half of uncovered skills. The authors release 1,251 skills with tasks and item scores.
In briefDeformStruct: Deformable Structure Modeling for Learning Action Timing from Single Timestamps
SummaryThis paper proposes DeformStruct, a method for localizing actions in videos when each action has only one annotated timestamp. It builds each action's structure in three steps: finding structural anchors, aligning latent parts to those anchors with deformation-aware matching, and modeling temporal relations among the parts. Experiments on THUMOS14, ActivityNet-v1.3, GTEA, and BEOID report consistent improvements in localization.
In briefSystem One Decision Models Are Not Better Rerankers
SummaryThis paper tests whether decision fine-tuning of System One (decision) models yields a better reranker than standard reranker fine-tuning. The authors compare six Qwen models from 0.6B to 27B parameters across ten TREC DL and BEIR collections. Reranker fine-tuning outperforms decision fine-tuning from 0.6B to 9B, while the two tie at 27B.
Michael Dinzinger, Kanishka Ghosh Dastidar, Michael Granitzer
In briefOrganizing video evidence in hyperbolic space to locate query-relevant moments
SummaryThis paper addresses video moment retrieval, which localizes the query-relevant time segment in an untrimmed video. The authors propose HTEO, which maps temporal representations into a Poincaré ball and organizes query-aware evidence using radial and orthogonal angular refinement. The organized representation is projected back to Euclidean space and injected into standard Transformer decoders through a lightweight residual update. The paper reports improved temporal localization across multiple benchmarks along with an interpretable geometric organization of evidence.
In briefSKG in practice: out-of-distribution generalization case studies across four domains
SummaryThis paper describes engineering implementations of the SKG inference framework and reports case studies in four domains: a 2D dissipative mechanical system, sim-to-real target detection, sine-to-square sequence modeling, and molecular dynamics on MD17. Each case is compared against a baseline. The authors also report run-to-run variance, the limits of single-seed results, and negative updates caused by parameterization mismatch.
In briefGeoLEON-CLIP: Graph-anchored contrastive learning linking maps, satellite images, and text for geographic regions
SummaryThis paper presents GeoLEON-CLIP, which aligns three views of each H3 region: an OpenStreetMap graph embedding, a satellite image crop, and a text description. Frozen LEON graph embeddings serve as a semantic anchor between vision and language. The authors propose chained cross-modal recall, and at R@1@1, routing retrieval through the graph beats direct image-text retrieval by 4.8 and 4.4 points. Downstream gains are largest on Airbnb and Beijing housing tasks.
Radoslaw Malek, Szymon Soltysiak, Aleksandra Kawala-Sterniuk et al.
In briefSupervised Separation of EV Interior Noise with a Hybrid Transformer
SummaryAn electric vehicle's cabin sound mixes wind, tire, powertrain, and auxiliary-drive noise. This paper frames decomposing that sound as a supervised binaural source-separation task with four target sources, adapting a hybrid time-spectrogram transformer originally built for music demixing. Operating-state information such as drive speed, load, vehicle speed, and auxiliary speed conditions the model. Training uses a physics-based parametric generator that synthesizes labeled sounds, and validation covers both synthetic and measured cabin recordings.
In briefMany Perturbations, Shared Failure Dynamics: Direction-Sensitive Influence in Language Models
SummaryThe paper examines why task-preserving rewrites can flip a correct language-model answer to a failure. The authors propose a direction-sensitive, readout-relative account and use a compact reserve-and-transmission model to predict failures in Qwen3-32B and Gemma 4 31B. They report shared predictive regularities, but find that directional causal effects depend on context and do not persist across option orders.
In briefSolvent: Long-Horizon Reasoning Model for Multi-Step Retrosynthesis
SummarySolvent is an 8-billion-parameter language model that extends long-horizon reasoning to multi-step retrosynthesis, the task of breaking a target molecule into purchasable building blocks through successive reactions. Trained through a two-round post-training curriculum, it interleaves iterative synthesis planning, external availability verification, and adaptive route revision. After reinforcement learning, which gave a 2.1× performance boost, it scores 61.86% across four benchmarks. The paper reports that this surpasses mid-scale models and prior search-based systems and rivals frontier trillion-parameter LLMs.
In briefDMAIC-Bench: A benchmark testing whether AI agents stop or redirect based on intermediate evidence
SummaryThe paper introduces DMAIC-Bench, which evaluates whether LLM agents in multi-stage analytical tasks can restrain themselves, meaning they change their next action based on intermediate evidence rather than only critiquing output. It uses 97 seeded tasks from 15 industrial cases built on Lean Six Sigma statistical prerequisites. Across 13 models, routing accuracy averaged 92% on GO decisions but 19% on NO-GO decisions, and blinded expert review estimated invalid evidence was acted on at 47% of reached NO-GO decisions. The framework was also applied to 16 life-sciences tasks.</summary>
WhyDefines self-restraint with paired counterfactual tasks, validated across 13 models and blinded expert review, showing only 19% NO-GO accuracy.
In briefCoPiT: A conditional Transformer for predicting airflow fields on irregular meshes
SummaryThe paper proposes CoPiT, a Transformer-based surrogate model for predicting aerodynamic flow fields on unstructured meshes. It builds on the Position-induced Transformer and injects global operating and design conditions through multi-stage feature-wise linear modulation. The model is tested on two-dimensional AirfRANS airfoil data and three-dimensional BlendedNet++ blended-wing-body surfaces. On the AirfRANS full-data test set, the authors report surface and volume loss reductions of about 54% and 24% relative to Transolver.
In briefMemCryst: Lifelong Agent Learning by Linking Memory and Reusable Skills
SummaryThis paper introduces MemCryst, a framework that lets a frozen language model agent keep learning over time within a limited context window. Stored memories support the creation and revision of reusable skills, and interactions guided by those skills generate new memories. Retrieved memories and skills together form the agent's context, and task outcomes adjust how much the served records influence future retrieval without rewriting them. The method leads seven of the reported comparisons across ten settings, and on ALFWorld joint serving reaches 63.17% success versus 59.83% for memory-only and 50.50% for skill-only.
In briefOlmoEarth v1.3 Embeddings: A New Model for Remote Sensing Data
SummaryThe authors train OlmoEarth v1.3, a model that condenses multimodal, multitemporal remote sensing data into a single 64- or 128-dimensional vector. They report state-of-the-art embeddings and competitive performance as an encoder backbone relative to v1 and v1.2, while needing 18% and 53% of their GPU hours to train. Using the model, they compute and release global, annualized 10-meter embeddings under the OlmoEarth artifact license.
Gabriel Tseng, Favyen Bastani, Patrick Alan Johnson et al.
In briefAdding Image Structure to DINO Features for Object Part Discovery
SummaryThis paper introduces Part-DINO, a training-free method that splits a foreground object into semantic parts. The authors find that DINO's fixed patch grid blurs part boundaries, so they use boundaries from low-level image segmentation to divide the image into regions. DINO features are pooled within each region and then grouped to discover parts. The method outperforms training-free baselines on four benchmarks and is competitive with weakly-supervised and unsupervised methods.
Samyak Rawlekar, Nikhil Chowdary Paleti, Amey Gupta et al.
In briefReLEAP: Reusing vision datasets by unifying tasks and labels with semi-supervised annotation
SummaryThe paper introduces ReLEAP, a scalable Python framework that unifies heterogeneous image datasets through data-source abstraction, annotation reuse across tasks, and mapping labels to a shared hierarchical taxonomy. The authors demonstrate it by integrating fifteen datasets for precision agriculture covering many plant species and annotation types. Cross-dataset and fine-tuning experiments examine how models trained on unified data and foundation models can generate higher-level annotations under unsupervised, weakly supervised and semi-supervised settings. The results indicate that unified access and annotation upcycling expand reusable training data and annotations.
In briefTrial by Trace: A Benchmark for AI Agent Harness Reliability and Failure Diagnosis
SummaryThis paper introduces TraceTrial-Bench, a benchmark for evaluating the reliability of harnesses that support LLM agents, organized into seven components such as context, execution, and verification. It includes 95 expert-level tasks with executable or state-based ground truth, along with modules for trajectory tracing, multi-source evidence gathering, and run-level failure attribution. Across 4 harnesses, 9 models, and 3,420 runs, macro-averaged harness scores range from 59.8% to 72.0%. The authors also report that 31.1% of runs succeeded at the task despite failing at least one harness criterion.
In briefPolyp-VSSD: A Transformer-Based Method for More Accurate Colon Polyp Segmentation in Endoscopy
SummaryThis paper introduces Polyp-VSSD, a segmentation framework built on a pre-trained Transformer backbone that jointly decodes global and local context to outline colorectal polyps in endoscopic images. It combines the feature strength of large-scale visual pre-training with global semantic modeling and fine-grained local reconstruction. The authors report that it outperforms state-of-the-art methods in Dice and IoU on multiple public endoscopic datasets, reaching 0.903 mIoU and 0.946 mDice on EndoScene.
In briefShared Negation Representations in Multilingual LLMs: A Causal Analysis
SummaryThis paper studies how multilingual large language models represent negation across languages. Using causal interventions on six open models across English, French, Vietnamese, and Chinese, the authors find that negation representations are controllable and largely shared. A single sparse autoencoder feature per model can control negation polarity across languages.
WhyCausal interventions across six models and languages show shared, controllable negation directions and a single sparse feature controlling polarity, with solid evidence.
In briefSimulating Driver Interactions with Data-Calibrated Differential Games
SummaryThis paper proposes a general-sum differential game model of vehicle interactions at urban intersections and calibrates it to recorded driving data from four intersections in Washington, D.C. On 96 vehicles not used for calibration, the model achieved the lowest mean position error of 7.42 m among the compared models. It also outperformed calibrated SUMO and inverse reinforcement learning baselines in cross-validation.
WhyCalibrated differential game beats SUMO and inverse RL on held-out vehicles with significance tests and verified approximate equilibria, a strong empirical contribution.
Amirmohammad Khakpour, Nachuan Li, Pedram Beigi et al.
In briefRootDown: Edge-Device Detection of Partially Occluded Sugar Beets
SummaryThis paper presents an image-based pipeline for detecting partially occluded sugar beets on loader conveyors, aiming to support automated post-harvest quality inspection. The authors introduce a benchmark and use the SAM 3 foundation model to generate bounding-box labels automatically, comparing models trained on these labels against those trained on manual annotations. They also benchmark inference configurations on an edge device, measuring precision, inference time and power consumption. The results indicate that foundation-model labels can replace a substantial share of manual annotation at a modest accuracy cost, and a Raspberry Pi 5 with a Hailo accelerator is presented as a practical low-power option for outdoor use.
Gerardus Croonen, Andreas Trondl, Julia Simon et al.
In briefRVG-Bench: A Benchmark for Reasoning-Driven Audio-Visual Temporal Grounding
SummaryExisting video grounding benchmarks mostly locate segments using descriptive queries of visual content. RVG-Bench targets real-world queries that require understanding the user's intent before localization, with each target identifiable only through such reasoning. It contains 1,255 videos and 1,890 query-interval pairs across eight reasoning types. The strongest of 22 evaluated models reaches only 49.6 mIoU.
In briefHyperparameter-free boundary-interior t-vMF Dice loss for medical image segmentation
SummaryThis paper addresses class imbalance and errors near anatomical boundaries in medical image segmentation by proposing a self-tuned, boundary-aware t-vMF Dice loss (STB-tvMF Dice). It adapts the similarity function's shape separately for boundary and interior regions, setting each from the previous epoch's mean prediction-target cosine similarity, without tuning parameters on validation data. Experiments on three datasets and architectures show its effectiveness, including a mean DSC of 70.75% on Synapse with U-Net versus 63.03% for Adaptive t-vMF Dice Loss. The authors state that code will be released soon.
In briefEmpirical concentration bounds for chains with infinite memory
SummaryThe paper studies concentration inequalities for stationary time series with infinite memory, extending prior work limited to Markov chains. It first proves a bounded-differences inequality on bounded Polish metric spaces under a summable Wasserstein contraction condition. Dependence coefficients are then estimated from a single trajectory, using transition counts for finite state spaces or kernel conditional-density derivatives for continuous spaces. A fixed multiplicative safety factor converts these estimates into empirical confidence bounds.
In briefTokamak-SI: A labeled catalog for interpreting tokamak discharges, unifying earlier event labels from DIII-D experiments.
SummaryThe paper presents Tokamak-SI, a scalable, time-aligned, categorical dataset that merges labels and labeling methods from many earlier studies on the DIII-D tokamak. It covers confinement, detachment, Alfvén eigenmodes, edge localized modes, sawtooth, tearing modes, and resistive wall modes. Models trained on it can extend labels to new shots and match or beat earlier detectors, including for AE, ELM, and TM detection.
Nathaniel Chen, Alvin Garcia, Ricardo Shousha et al.
In briefLapaFM: a laparoscopic surgery foundation model built with multi-stage continual learning
SummaryThis paper presents LapaFM, a foundation model for laparoscopic surgery trained on over 20,000 minutes of surgical videos using a multi-stage continual learning strategy. It supports tasks such as smoke removal, surgical phase recognition, instrument and polyp segmentation, and lesion identification. The authors report that it outperformed task-specific and existing foundation models across retrospective and prospective cohorts, and also generalized to gastrointestinal endoscopy tasks.
In briefBeyond Pattern Matching in 3D MLLMs: Part-Level Functional Reasoning for Generalizable Recognition
SummaryThe paper argues that 3D multimodal LLMs rely on pattern matching, which fails to tell apart geometrically similar objects such as dressers and wardrobes. It proposes PFIR, which builds part-level functional reasoning data from PartNeXt and trains with dual-objective GRPO. Under a strict out-of-distribution protocol, zero-shot accuracy rises from 60.5% to 66.7%, with larger gains on geometrically ambiguous samples.
In briefHeterogeneous space data centers as a key design for future in-orbit services
SummaryThe paper proposes heterogeneous space-based data centers made of large and small nodes that differ in computing power, size, and power budget, serving ground, aerial, and in-orbit users with AI-centric services. It analyzes dawn-dusk low-Earth-orbit configurations, covering solar power availability, visibility windows, and latency-aware coverage. Latency and service reliability are evaluated using baseline workloads, and energy requirements for different roles and deployment scales are quantified. The authors also offer design guidance and future research directions.
In briefMedCaseAgent: Generating Rare Disease Case Reports from Multimodal Clinical Evidence
SummaryThis paper presents MedCaseAgent, an agent-based framework that generates evidence-grounded case reports for rare and atypical diseases from unstructured clinical notes. It iteratively reasons by querying medical literature and knowledge bases, and also assesses how novel a case is. The authors release the MedCase-150K corpus and the MedCase-Bench evaluation suite, and report that the system outperforms existing general and medical LLMs and agents in automated and blinded expert evaluations.
In briefLearning Image Representations with Composite Subspace Flow Matching
SummaryThis paper proposes composite subspace flow matching (CSFM), which frames representation learning as inverting a deterministic degradation curriculum across spectral, color, and illumination subspaces. Experiments on CIFAR-10, CIFAR-100, and STL-10 report that CSFM outperforms standard optimal transport flows on linear probing and few-shot classification. The method also retains coherent synthesis and trajectory interpolation.
In briefAnti-coherence Regularized Absolute Scaling: A Post-hoc Method for Out-of-Distribution Detection
SummaryThis paper addresses out-of-distribution (OOD) detection, a key AI safety problem, by improving the SCALE method. The authors introduce an Absolute Fix that computes scaling ratios over absolute activation values, and Anti-coherence regularization that suppresses sign-mismatched evidence for the predicted class. The method needs no training data or labels, and under the OpenOOD v1.5 protocol on CIFAR-10 and ImageNet it matches or outperforms activation-shaping methods that use training data in AUROC and FPR@95.
In briefRSIHub: A Composable Framework for Building and Comparing Agent-Evolution Methods
SummaryThe paper presents RSIHub, an open-source framework for building, comparing, and extending methods that use execution feedback to improve agent prompts, skills, tools, and orchestration. It separates method-specific logic from shared execution and evaluation infrastructure, and instantiates AHE, Hyperagents, A-Evolve, and GEPA without changing the framework core. Evaluated with MiniSWE and Codex on Terminal Bench 2 and τ³ Banking, the largest observed overall pass@1 gain is 13.5 percentage points, with method rankings varying across settings.[]
In briefEasy-3Dfy: Building 3D scenes from one image by composing object-level geometry
SummaryThe paper proposes Easy-3Dfy, a workflow that separates object generation from scene layout estimation so each part can use specialized models. An RPF-based aligner matches object-local meshes to partial scene pointmaps to recover object layout for composing the scene. The method outperforms SAM-3Dfy on several geometric metrics on Scene-Maker and 3D-FUTURE200 without requiring complete scene-level supervision.
In briefDo halting heads truly reduce compute in looped language models?
SummaryLooped language models pass each token through the same layers repeatedly, and a halting head decides when each token can stop early to save compute. Pretraining models up to 1.3B parameters, the authors find that heads trained on mixed iterations predict early stops, yet tokens still run almost every iteration under the standard stopping rule. Scoring every iteration works only with a strong penalty favoring early stops. In batched generation, a model without a head performs almost as well, so average iteration counts may overstate savings; the authors recommend reporting actual iterations, thresholds, and batch size.
In briefMinimum Description Length Separates Law-Grade from Statistical-Grade Knowledge in a Science Agent
SummaryThe paper uses the MDL compression ratio as an unsupervised measure to grade which discovered relations are genuine laws. Applied to a self-improving science agent's ledger, the top-ranked band recovers canonical laws. The authors also find that 86% of the agent's discoveries compress nothing and are merely statistical regularities.
In briefBiLoCo: A binary low-rank correction method for post-training quantization of language models
SummaryThe paper proposes BiLoCo, a low-rank correction method that offsets accuracy loss when large language models are quantized to FP4 after training. It stores the correction factors as binary sign vectors instead of BF16 numbers, using 16 times less storage per rank-one component, and runs them on CUDA cores alongside the FP4 tensor-core matrix multiplication. On 4B to 72B models, it matches or improves memory-matched SVDQuant, and its fused B300 kernel stays within the slowdown limits reported in the paper.
In briefOnline 3D bin packing method combining spatial heuristics with diffusion reinforcement learning
SummaryThis paper proposes a method for online 3D bin packing that combines classical packing heuristic rules with diffusion reinforcement learning. It compresses the action space through heuristic action pre-screening, uses a graph neural network to model spatial structure, and adds a flatness-based reward to reduce spatial fragmentation. The excerpt mentions experimental results but does not show specific figures.
In briefCorrect but unselectable: hidden interface loss in multi-candidate reasoning
SummaryThe paper finds that in generate-then-rerank reasoning, a correct answer can be among the candidates yet still go unselected because it is incomplete, unparseable, or in a mismatched answer format. The authors propose Compatibility-Aware Candidate Construction (CACC), an interface layer between the candidate source and the fixed selector that repairs fragments and aligns answer modes without retraining. CACC raises final accuracy on numeric reasoning and GPQA Diamond, but on MMLU-Pro it lowers final accuracy under the original selector.
In briefMS-TTRL: Resolving the Frequency Conflict in Label-Free Multiple-Solution Reinforcement Learning
SummaryThis paper addresses reasoning tasks with multiple valid solutions, where fully labeling all solutions is costly. It proposes MS-TTRL, a test-time reinforcement learning method that uses unlabeled test instances and separates the roles of generation frequency: correctness is estimated through outcome-level self-verification, while frequency is used only to preserve diversity. Experiments across two benchmarks, four multiple-solution tasks, and three backbone models report an average 19.13 percentage point gain in Recovery Rate over the best label-free baseline, approaching supervised IPS-GRPO.
In briefTileShare: Training-free speedup for overlapping tiled ViT decoding
SummaryThe paper addresses overlapping tiled Vision Transformer decoders, where overlap tokens are processed by several decoder instances. TileShare keeps attention per tile, computes other operations once per token, and blends per-tile results in overlap regions. Without training, keeping six layers per tile decodes 1.38 to 1.51 times faster than batched tiling on DAVIS videos while matching the released decoder's PSNR, LPIPS and DISTS.
In briefPhysical camouflage attacks that fool multi-view bird's-eye-view 3D perception in self-driving cars
SummaryThis paper proposes BEVCA, a framework for generating adversarial camouflage aimed at multi-view bird's-eye-view (BEV) 3D perception models used in autonomous driving. It combines a BEV-feature-based adversarial loss with a multi-view neural rendering module to make attacks effective and transferable across models and tasks. Experiments in digital and physical settings reportedly outperform state-of-the-art baselines and show strong robustness under real-world conditions.
WhyAdds BEV-feature adversarial loss with multi-view neural rendering for camouflage attacks, validated in digital and physical settings against baselines.
In briefTraining a mixture-of-experts language model with 4-bit MXFP4 arithmetic, closing most of the gap to BF16
SummaryThis paper presents ALTO, an open-source recipe for training the 20-billion-parameter GPT-OSS-20B mixture-of-experts model in the MXFP4 4-bit format. It combines existing techniques such as block quantization, randomized Hadamard transforms and stochastic rounding with a simplified weight de-oscillation scheme. On C4 validation, the method reaches a loss of 3.3350, only 0.007 above the BF16 baseline.
In briefRoboCode: Learning Reusable Motion Building Blocks for Robot Grasping and Manipulation
SummaryRoboCode learns shared, soft-combined motion primitives from action chunks pooled across several tasks, instead of predicting actions directly. Each action is represented as a weighted mix of these learned primitives, and task-specific state encoders map observations into this shared space. On five ManiSkill manipulation tasks, it reports an average reconstruction MSE of 0.0187 and a 75.6% average success rate, compared with 68.8% for ACT and 41.6% for Diffusion Policy. The authors conclude the representation is effective and interpretable.
In briefRight Answers, Still Blocked: SolveBind for GUI Agents Facing Verification
SummaryThis paper addresses CAPTCHAs and anti-automation mechanisms that judge behavioral signals such as interaction trajectories and timing, not just correct answers. It proposes SolveBind, a GUI-agent framework that pairs hierarchical solving with runtime supervision. Tested on nine CAPTCHA task families and nine backbone models, it achieves the highest success rate on eight backbones, outperforming the strongest baseline by 13.3 to 34.5 percentage points.
In briefACWM-RM: A reference-free reward model trained on complementary mismatch preferences
SummaryThe paper addresses reinforcement learning for action-conditioned world models by proposing ACWM-RM, a reward model that scores both video quality and action-following fidelity without ground-truth videos. The authors build AVM-Eval, a 67.2k-pair benchmark, and AVM-Pref, a 506.3k-pair preference dataset. Results show inverse dynamics models are unreliable on generated videos, and downstream RL experiments show ACWM-RM outperforms reward baselines on both criteria.
In briefRxnExBench: A Benchmark for Testing AI Models on Extracting Chemical Reaction Data from Papers
SummaryRxnExBench is a multimodal benchmark built from high-impact chemistry journals, containing 3,565 expert-annotated reactions and 2,361 molecular structures. It evaluates nine state-of-the-art vision-language models on extracting structured reaction data from figures such as screen tables and substrate scopes. Performance drops as schema complexity and reasoning demands rise, and accurate molecular recognition alone is insufficient, with multimodal integration and chemistry-aware reasoning remaining the main bottlenecks. The data and code will be publicly released.
In briefAngular Supervision for Locating Objects in Autoregressive Vision-Language Models
SummaryAutoregressive vision-language models often fail to link generated text to the right image regions, and existing explanation methods give weak spatial maps. The paper proposes Angular Relevance, a heatmap measure based on cosine similarity between visual token embeddings and unembedding rows, and an auxiliary Angular Object-Aware Loss that aligns visual tokens with their object concepts. The authors report sharper object localization and competitive referring-expression segmentation, without degrading general VQA performance.
In briefPEAR: adaptive reshaping of advantages to improve visual perception and exploration in vision-language models
SummaryThis paper proposes PEAR, a method for improving multimodal reasoning in large vision-language models by better balancing visual perception and exploration. It uses visual dependency as the base signal and adds an entropy-based exploration bonus only at high-entropy positions, giving larger bonuses where visual dependency is lower. The fused signal is converted into positive token-level weights that reshape sequence-level advantages. Experiments on Qwen2.5-VL-3B and 7B report the highest average performance among evaluated baselines on multimodal reasoning benchmarks, with leading results on textual reasoning and general multimodal benchmarks.
In briefDemocratizing LLM Training at the Edge: Challenges, Systems, and Future Directions
SummaryThis paper examines distributed training of large language models across geographically dispersed sites with heterogeneous hardware and unreliable connectivity, covering federated, split, and decentralized approaches. It proposes a unified system in which a cloud-hosted parameter server coordinates heterogeneous clients over outbound gRPC connections, sharing one communication, compression, and fault-tolerance stack. Experiments pretraining GPT-2 Medium and Llama3-1B across multiple university sites show stable convergence. The authors conclude by listing open problems, including frontier model support, elastic fault tolerance, privacy, and incentive design.
In briefDIAL-S2ST: Learning Latency Control for Real-Time Speech-to-Speech Translation
SummaryThis paper presents DIAL-S2ST, a post-training method that lets simultaneous speech-to-speech translation models adjust their latency at inference time. It reuses pretrained translation ability through self-distillation from lookahead predictions and limits the influence of large text distillation losses. A single LoRA adapter is trained jointly across scales so the timing can be tuned. Experiments on French-to-English translation with Hibiki show that only 10 hours of source speech suffice, giving flexible latency control with a favorable quality-latency trade-off.
