In briefModeling information flows to assess the EU AI Act by role
SummaryThis paper explores using the Information Flow Methodology (IFM) to assess compliance with the EU Artificial Intelligence Act. IFM models decision systems, whether algorithmic, human, or hybrid, as networks of information sites and transformation channels to show how information, decisions, and responsibilities move across organizational and technical boundaries. The authors apply it to a recruitment support tool used by a public-sector organization, assessing both the Provider's system and the Deployer's process. They argue it supports transparency, identifies risk pathways, and compares intended with actual use.
In briefFOCUS: A Subspecialty AI Model for Intraoperative Brain and Spinal Cord Tumor Pathology
SummaryFOCUS is a pathology foundation model pretrained on 27,930 frozen sections enriched for central nervous system tissue, designed to support intraoperative diagnosis. It achieved mean AUCs of 0.969 on 11 diagnostic tasks and 0.877 on 12 biomarker-prediction tasks, maintaining strong performance across temporal and external cohorts. In reader studies with six neuropathologists, FOCUS assistance raised junior readers' lymphoma sensitivity from 81.6% to 94.7%.
WhySubspecialty model for intraoperative CNS frozen sections, with multi-task evaluation and six-reader study showing clinically meaningful gains.
In briefAutomated AI auditing: giving auditors extra context can change the results
SummaryThis paper uses the Petri tool to audit LLM coding agents and tests whether giving automated auditors an external codebase or benchmark transcripts improves their effectiveness, across four auditor-target pairs and three behaviours. The external codebase suppressed reward hacking in some pairs, while benchmark transcripts raised reward hacking scores in all four pairs. No consistent effect appeared for other behaviours. The authors stress that measured target behaviour rates are highly sensitive to configuration, so careful piloting is important.
In briefConDenseDrive: A Knowledge-Dense, Token-Sparse Vision Encoder for End-to-End Driving
SummaryThis paper presents ConDenseDrive, a visual encoder for end-to-end autonomous driving. It distills visual, semantic, and geometric knowledge from three vision foundation models, then integrates multi-frame spatiotemporal context into register tokens for planning. Register-guided token compression reduces redundant computation, yielding a 2.0× speedup while improving planning performance. The authors report 94.5 PDMS on NAVSIM v1, 90.8 EPDMS on NAVSIM v2, and 37.4 HD-Score in zero-shot evaluation on HUGSIM.
In briefAnalyzing Where Camera and LiDAR Disagree in Self-Driving 3D Detection
SummaryThe paper labels each ground-truth object by which sensors detected it, mining a 0.4M-object database across nuScenes, KITTI, and Waymo. It introduces Spatially Stratified Feature Attribution (SSFA) to show that modality reliance shifts with distance and object type. The authors build a Modality Reliability Map and inject it into four pretrained fusion detectors, improving mAP by up to 1.3 while fine-tuning under 4% of parameters.
In briefNeural Bayesian Networks: modular, scalable models for decisions under uncertainty
SummaryThis paper introduces Neural Bayesian Networks (NBNs), which keep the graph and factorization of classical Bayesian Networks but replace each conditional with a differentiable, tensor-valued mechanism. The authors say this lets NBNs represent any learnable conditional distribution and run natively on GPUs. They also release a PyTorch library and a public benchmark. The paper reports better expressiveness and, on large networks and query batches, speedups of about two orders of magnitude.
Giovanni Briglia, Marco Lippi, Stefano Mariani et al.
In briefObserve, Share, or Predict? How Partner Information Affects Embodied Robot-Human Teamwork
SummaryThis paper studies whether LLM-based planners can support collaboration among embodied agents without dialogue, using the Habitat-based PARTNR benchmark. Agents perceive only individually, and the authors test three additions: observing the partner's state, exchanging object knowledge, and predicting the partner's next action. Observing the partner's state mainly improves task quality, object knowledge exchange mainly improves efficiency, and next-action prediction usually lowers success rates.
Adri\'an Szlatincs\'an, Filippo Ziliotto, Bruno Lepri et al.
In briefBarycentric Guidance: Making Foundation Image Editors Control Emotions Continuously
SummaryThis paper introduces Barycentric Guidance, a training-free method that turns pretrained foundation image editors into continuous controllers over the valence-arousal emotion space, without finetuning, extra data, or model changes. Any target valence-arousal point is mapped to simplex-based guidance weights. On three face datasets, the method improves VA target accuracy, VA-plane coverage, and expression diversity over the strongest baseline, and it extends beyond human faces to animals, artworks, and complex scenes.
Harvey Mannering, Zhiwu Huang, Adam Pr\"ugel-Bennett
In briefUsing AI to Find Tumor Growth Rules from Sparse Data: Refitting and LLM-Based Ranking
SummaryThis paper studies an AI-assisted workflow that uses symbolic regression, mathematical validity checks, coefficient refitting, and large language model ranking to infer tumor growth laws from sparse, noisy data. On synthetic Gompertz data, dense noiseless sampling recovers the true growth law, but sparse noisy sampling often leads to different forms. Physics-informed neural network refitting often increased growth-law error, while classical nonlinear least squares usually reduced it, and LLM rankings showed no consistent gain over selection by training error and contained mathematical reasoning errors. The authors conclude that plausible saturation, accurate trajectory fits, and convincing explanations can coexist with incorrect growth laws.
In briefFORGE: Generating Expert Weights On-the-Fly to Speed Up Sparse MoE Inference
SummaryThe paper introduces FORGE, which regenerates Sparse Mixture-of-Experts expert weights on-chip from compact rotor parameters instead of loading them from HBM. Replacing 30% of MoE FFN layers across five models is estimated to raise system throughput by 1.3-2.8x and lower per-user decode latency by 1.10-1.15x at short context. Downstream accuracy falls below fine-tuned models without replacement, and the replacement ratio sets this tradeoff.
Chanyoung Moon, Travis Pence, Daisuke Yamada et al.
In briefTransparency as Answerability: Role-Dependent Information Flow Modeling
SummaryThe paper proposes defining transparency as a relation: a system is transparent to a stakeholder when that stakeholder's legitimate questions can be answered from the system's structure. It formalizes this with information flow networks and analyzes structural, compositional across organizational boundaries, and role-dependent answerability. The framework is demonstrated on a pre-integration assessment of an AI recruitment tool in a public sector municipality, using participatory design and interviews to model the decision system.
In briefLearning Operators for Symmetry Discovery: From Lie Groups to Local Symmetry Geometry
SummaryThis paper asks whether the map from a function to its symmetry structure can be learned across a family of functions from finite observations, rather than estimated separately for each function. For global Lie-group symmetries, it gives a representation-theoretic characterization and shows the inverse symmetry operator is equivariant under coordinate changes. On SO(5) benchmarks, amortized equivariant estimation improves symmetry recovery from sparse observations, and the approach extends to local symmetries with integral submanifolds guaranteed by Frobenius' theorem.
In briefAcceptable Misalignment: Shifting from Maximizing Alignment to Defining Tolerable Misalignment Risk
SummaryThis paper examines the value alignment problem for autonomous robots capable of open-ended learning. The authors argue that full alignment is both theoretically and practically impossible, so misalignment incidents will continue to occur. They introduce the concept of acceptable misalignment and formulate normative requirements for alignment mechanisms that bring misalignment to an acceptable level. The paper concludes with a methodological framework built from alignment mechanisms that meet these requirements.
Maud van Lier, Ludovica Marinucci, Vieri Giuliano Santucci et al.
In briefEviStress: Learning to Stress Visual Evidence for Face Forgery Detection with Text Explanations
SummaryThis paper addresses the difficulty of detecting face forgeries spreading on the Web, which often involve unseen manipulations and visual changes. It proposes EviStress, a two-stage framework that perturbs highly relied-upon image patches during training and then tunes a multimodal model (LLaVA) to predict authenticity and explain its judgment in text. The reported experiments show improved cross-dataset, cross-method, and corruption robustness.
In briefStudy of how pretraining sequence composition affects EHR foundation models
SummaryThis paper examines how arranging patient event trajectories into pretraining sequences affects clinical prediction in electronic health record foundation models. It compares alternative composition methods against the common global-stream baseline. The best method improved macro AUROC, AUPRC, F1, and balanced accuracy on four MIMIC-IV v2.2 tasks, with a 9.3% relative F1 gain and 2.7% relative AUPRC gain. Improvements were also reported on four additional evaluation settings, including INSPIRE and eICU-CRD.
Joshua Placidi, Inigo Stibbe, Alfred J. Balston et al.
In briefCLIVO: A method for choosing and mixing training instructions for vision-language models using clusters and learned validation goals
SummaryThis paper presents CLIVO, a framework for selecting instruction data to train vision-language models. It groups data into semantic clusters whose training shares can be adjusted, and uses merged cluster experts as proxies for evaluating mixtures. Validation objectives are learned by correlating proxy losses with real probe performance, and mixture weights are found through differentiable search. Selecting 100K instructions from an 8M pool, the authors report CLIVO reaches 97.2% of the full-data four-target average score using 1.25% of the data, outperforming the strongest evaluated baseline.
In briefSpecGuardBench: Can AI coding agents write reliable tests from natural language specs?
SummaryThis paper introduces SpecGuardBench, a benchmark of 291 tasks from 13 real-world repositories covering 8 types of verification logic. Agents must build a validation harness of executable checks and test scenarios from a natural language specification, repository context, and a reference implementation. None of four tested coding agents produced a harness rated faithful for more than 63% of tasks. The main weakness is writing checks that faithfully reflect the specification, as agents tend to mirror the reference implementation and wrongly reject implementations the specification allows.
In briefFrom Parallel to Sequential: Small-Batch Noise Restores Muon's Simplicity Bias
SummaryThe paper finds that Muon, a spectral-descent optimizer, learns features in parallel in the full-batch setting rather than one at a time as gradient flow does. It shows that stochastic gradient noise restores sequential feature learning and derives the effective continuous dynamics of noisy spectral descent for deep linear models. Experiments on low-rank matrix reconstruction, vision transformers, and large language models support these findings.
In briefDavinci-Tool: A tool-use model trained on verified executable environments for AI agents
SummaryThe paper introduces Davinci-Tool, a 30B tool-use model, and Davinci-ToolData, a dataset grounded in executable environments with verifiable outcomes. Two pipelines produce 30,395 verified trajectories covering 2,260 tool names. The model is trained with supervised fine-tuning and Direct Preference Optimization. It reports gains on τ²-bench, BFCL v4 multi-turn, and MCP-Atlas, outperforming Qwen3-Coder-480B-A35B-Instruct with 1/16 of its parameters.
WhyCombines verifiable executable environments with outcome-grounded preference optimization, substantially improving three tool-use benchmarks and beating a 16x larger model.
In briefPivotRL: Dynamically Shifting GPU Resources for Efficient RL Post-Training with Generative Rewards
SummaryThis paper presents PivotRL, a reinforcement learning post-training system for large language models that dynamically shifts GPU resources among policy rollout, generative reward evaluation, and policy training based on runtime demands. It uses a Global Parameter Storage-based runtime, a shift-aware scheduler for sequence migration, and a planner that weighs throughput gains against shift overhead. On a 128-GPU cluster, the authors report 1.81-2.05 times higher end-to-end throughput than colocated deployment and 1.50-1.95 times higher than disaggregated deployment, while preserving RL convergence.
In briefIs human-machine teaming better together? A performance theory of skill policies under varying difficulty
SummaryThe paper presents a performance theory for tasks completed by a human and a machine together, separating a systematic part set by each skill from a residual that only paired observation can reveal. It analyzes how collaboration methods such as routing, weighting, and synergy can capture this value. In a public chest-radiograph study with 227 radiologists, the radiologists were near parity with the model; assisted reading returned less than averaging both responses, while recalibrating the model alone beat every deployable arrangement of raw responses.
In briefFrom Volpiano to Pixels: A 30,000-Line Ground-Truth Dataset for Medieval Chant Music Recognition
SummaryThe paper builds 30,567 image-target pairs for medieval chant manuscripts by linking existing Volpiano transcriptions from the Cantus Database to page images, drawn from ten sources across eleven exports. A CRNN trained from scratch reaches a 9.6% symbol error rate on 3,086 test lines, compared with 45.3% for an earlier model. Leave-one-manuscript-out experiments show error rates of 21.5–31.0% on the excluded manuscripts.
In briefGpFlash: A Gaussian process library for fast, memory-efficient training and prediction
SummaryThis paper presents GpFlash, a Gaussian process library designed for fast training and posterior evaluation, low memory use, and scaling to high-dimensional problems. It was benchmarked against scikit-learn, GPy, GPflow, GPyTorch, and SMT, and used less memory and ran faster in all cases considered. With 10,000 training points in 10 dimensions, it needed 9.6 to 35 times less memory and was 2.6 to 54 times faster for marginal log likelihood and gradient evaluation.
In briefVision models capture complementary child and adult brain representations
SummaryResearchers compared EEG brain data from 8-year-olds, 12-year-olds, and adults with image-text aligned (CLIP) and vision-only AI models. CLIP aligned significantly with children's brain activity, while vision-only models aligned more strongly later in development. Residual analyses showed the two model types capture complementary structure. Much of the alignment reflected semantic features such as category and animacy, while low-level image properties stayed weakly aligned across ages.
Domenic Bersch, Timothy Schauml\"offel, Michela Proietti et al.
In briefANCHOR: A Benchmark for Whether Coding Agents Follow Engineering Constraints
SummaryThis paper introduces ANCHOR, a benchmark that tests whether repository-based coding agents follow developer constraints on modification scope, workflow order, validation, tool use, and reporting, not just functional correctness. It contains 358 multilingual tasks and evaluates both code diffs and execution trajectories. The strongest measured system reaches 70.1% atomic and 63.7% relation compliance, and only 19.0% of tasks pass every checklist item under the strict criterion.
In briefUsing LLMs to Cut the Cost of Building Video Datasets
SummaryThe paper proposes using large language models to pre-screen videos based on their metadata before downloading them, reducing storage, time, and annotation costs. The LLM is evaluated against a majority vote from four independent annotators, with inter-annotator agreement checked using Gwet's AC1. It is also compared with two keyword-based baselines that do not use LLMs, and cost comparisons with existing approaches are reported.
Mohammed Shabot, Jonathan Adams, Jan Van den Stock et al.
In briefRadial Tail Repair: a post-training fix for diffusion and flow models' extreme-value errors
SummaryDiffusion and flow-matching models can struggle to represent distribution tails, especially for heavy-tailed data. This paper proposes a lightweight post-training correction that leaves the model unretrained and its sampling unchanged, using a radial-angular decomposition to keep the learned directions and fix only the radial tail. Experiments on synthetic heavy-tailed distributions and real weather and wildfire datasets report substantially improved extreme-tail coverage with only a small number of fitting samples.
Hamza Cherkaoui, H\'el\`ene Halconruy, Antonio Ocello
In briefA multi-task deep learning method that recognizes faces and locates facial landmarks at the same time, even when the head is turned
SummaryThis paper proposes a multi-task deep learning framework that performs face identity recognition and facial landmark localization together in one architecture, aiming to handle large pose variations. The authors adapted the CPLFW dataset by adding five-point landmark annotations alongside identity labels. On the prepared dataset, the model reports a validation recognition accuracy of 81.58% and a landmark localization mean squared error of 96.38.
In briefFeature Normalization Under Finite Training Budgets: When to Scale Rare Features
SummaryThis paper studies how standardizing a rare binary feature versus leaving it unscaled affects the lowest risk achievable on a rare subgroup when stochastic gradient descent has a fixed observation budget. For two independent binary features, the authors derive formulas for both optimized risks as prevalence approaches zero, and use finite-prevalence bounds to compare the two approaches. A controlled study finds that both orderings persist under observation reuse, estimated preprocessing, and intercept fitting, while selecting a predictor after training can reverse the ordering.
In briefSelecting a Few Arms to Cover Many Objectives in Bandits
SummaryThis paper studies how to pick a small set of options (arms) that serves many objectives in a multi-objective bandit, learning their performance from noisy feedback. It introduces few-for-many best-arm identification, which seeks arms minimizing the worst weighted degradation from each objective's optimum. The authors propose GFM-Elim and AFM-Elim algorithms with instance-dependent sample-complexity guarantees, and derive a lower bound for Gaussian rewards in the latent-group setting. Experiments on synthetic and real-world tasks show that sharing observations across objectives can substantially reduce sampling cost.
In briefLabel collapse in in-context learning of diffusion language models
SummaryThis paper studies a failure mode in in-context learning where models generate query-independent labels, which the authors call label collapse. Across five 8-9B models spanning autoregressive and discrete diffusion families, collapse appears even under class-balanced alternating arrangements. The authors show accuracy alone cannot detect it and propose separation, the log-odds of the two label words at the query position, as a complementary diagnostic.
In briefTesting whether AI safety guardrails correctly honor user permissions
SummaryAs language-model agents perform actions such as writing code, moving money and changing records, deciding whether each action is authorized becomes a key safety issue. The paper introduces \bench{}, a benchmark that compares guardrail decisions for the same operation with and without valid authorization, using expert-annotated scenarios, executed and verified operations, and externally sourced native-policy requests. Experiments show that many blocked operations result from failed reviews rather than correct denials. Fixing the output format restores usable decisions but changes which authorized requests are allowed, and adding a second reviewer removes some errors while also rejecting valid authorization.
In briefLarge-scale study testing whether syntax-aware language models still help as they grow
SummaryThis paper studies syntactic language models, which generate sentences together with their syntactic tree structures. The authors compare 12 controlled configurations, including one baseline and 11 variants, at scales up to 1B parameters and 100B training tokens. They evaluate them on 11 downstream tasks and two syntactic benchmarks. Models with standard causal attention match or exceed structural-mask variants, and at 1B parameters the best variant no longer clearly outperforms a compute-matched non-structural baseline.
In briefBell boundary: when shared representations can no longer jointly support multiple tasks
SummaryThis paper uses Bell's theorem to study when a shared representation becomes insufficient for multiple tasks in multi-task learning. It shows that if four binary task outcomes can be jointly represented by a common representation, their correlations must satisfy the CHSH bound |S|≤2. The authors introduce NCnet, a controlled model, and find that limited capacity can produce |S|>2, with these exceedances disappearing as capacity grows. Similar transitions are reported across CNNs, PLMs, and LLMs, which the authors call the Bell Boundary, a measurable diagnostic for representation bottlenecks.
In briefWhy average accuracy is a poor basis for deciding whether to release a model
SummaryThe paper argues that release decisions depend on the catastrophic failure risk of the next rerun of a training recipe, whereas standard benchmarks report only mean accuracy over past seeds. Auditing 2,124 fine-tuned models across eleven language models, the authors find that shipping based on mean accuracy violates a 10% deployment failure budget 48.0% of the time. They introduce recipe cards, which certify rerun tail failure rates using exact finite-sample binomial bounds and require at least 22 samples.
In briefDTA-GT: A Direction- and Topology-Aware Graph Transformer for Predicting Neural Network Performance
SummaryThis paper addresses performance prediction for neural architecture search, where candidate networks are operation-labeled directed acyclic graphs and edge direction encodes computational flow. The authors propose DTA-GT, a graph Transformer that maintains directional consistency across node initialization, spectral encoding, structural attention, and feature update. On NAS-Bench-101 and NAS-Bench-201, it outperforms representative sequence, GNN, Transformer, and hybrid predictors for accuracy and latency under limited-budget settings.
In briefExtrapolating simulation-based inference to larger sample sizes via statistical scaling
SummaryThe paper addresses a problem in simulation-based inference where neural posterior estimators trained on limited sample sizes become inaccurate on larger datasets. It proposes CE-NPE, which pairs a statistical center with an explicit contraction rate and learns a local center correction and residual densities. On six regular benchmarks, it reports the best agreement with reference posteriors among tested amortized estimators at all out-of-range sizes, up to five times the training maximum, without retraining.
In briefCalibrated α-barycenter method for multiobjective f-divergence optimization
SummaryThe paper addresses matching one distribution to many targets without repeatedly evaluating every target. It proposes CAL-α-MODM, which aggregates a nonlinear predictor into one barycenter, calibrates its residual derivatives, and checks steps against the original f-divergence objective. Under smoothness assumptions, the authors establish stationarity and local convergence guarantees with exact oracles. On 300-target Jensen–Shannon problems, CAL-p2 shows median paired speedups of 3.15× over L-BFGS and 1.10× over second-order Direct-TR.
In briefGrowing Vision Transformers Step by Step to Cut Pretraining Compute
SummaryThis paper studies progressive depth growth, which starts training with shallow self-supervised vision transformers and adds layers over time, in MAE and DINOv3 models. It examines effects on downstream global and dense tasks, intermediate-layer representations, and internal attention patterns. The authors report that growing reduces training FLOPs by 13.9% to 24.8% while improving downstream performance, with ImageNet linear-probe accuracy gains of up to 3.36 percentage points.
Amir Mohammad Karimi Mamaghan, Ferdinand Kapl, Andrea Dittadi et al.
In briefStatistical Downscaling of Daily ERA5 Precipitation Fields to IMERG Resolution over West Africa
SummaryThis paper uses U-Net based neural networks to downscale daily ERA5 atmospheric fields to finer-resolution IMERG satellite precipitation over West Africa for June to October seasons from 2000 to 2024. Inputs include geopotential, temperature, humidity and wind at three pressure levels, plus elevation and land fraction. The deterministic networks were found to be too dry and to lose most observed variance below a few hundred kilometers, so a GAN-based alternative was used, and permutation importance ranked 500 hPa specific humidity first.
In briefCalibration of Crowdsourcing Aggregators: Are Their Probability Outputs Reliable?
SummaryThis paper provides the first theoretical and empirical study of calibration in crowdsourcing label aggregation. It finds Majority Vote tends to be underconfident while model-based aggregators such as Dawid--Skene, GLAD, and CrowdFM tend to be overconfident, across 39 human datasets and LLM-based crowds. The authors prove the Bayesian aggregator is perfectly calibrated under its data-generating model. They also propose a post-hoc method that combines predictions from complementary aggregators before temperature scaling, improving calibration especially with many annotations per item.
WhyFirst systematic study of calibration in crowdsourcing aggregation, proving Bayesian aggregator calibration and validating across 39 human datasets and LLM crowds.
In briefCarve VFM: Compressing Vision Foundation Models into Task-Specific Compact Models
SummaryCarve VFM is a structured pruning method that compresses vision foundation models into smaller, task-specific encoders. It jointly searches for per-layer capacity allocations and restores removed behavior through iterative closed-form low-rank corrections, without retraining the backbone. The authors report the highest mean accuracy among evaluated dense pruning methods in most configurations, along with 1.3 to 2.0 times higher GPU throughput than the uncompressed model.
Lotfi Abdelkrim Mecharbat, Abdelrahman Elsayed, Ibrahim Almakky et al.
In briefDASS: A Solver-Adaptable Dynamic Auxiliary Search Strategy for Symbolic Regression
SummaryThis paper proposes DASS, a dynamic auxiliary search strategy for symbolic regression, which finds interpretable math formulas from data. DASS builds and refines a subspace of high-potential basis function terms to guide various downstream solvers. Solver outputs feed back to update the basis terms in a closed loop. Experiments on LLM-SRBench report significantly improved accuracy with a manageable runtime increase.
In briefX-Ragent: Tool-Augmented Reinforcement Learning for Evidence-Grounded Chest X-Ray Reasoning
SummaryThis paper proposes X-Ragent, a framework that trains a model to read chest X-rays by learning when to seek auxiliary evidence from external tools and admitting only valid quantitative measurements into its diagnostic reasoning. The authors build Rad-CoT, a 7.7K-sample corpus of tool-augmented trajectories and report-style reasoning, and optimize the policy with reinforcement learning using an accuracy-gated reward. Across three chest X-ray benchmarks, X-Ragent reports the strongest results, including an 8.6-point gain on the fully held-out ChestAgentBench.
In briefPredicting Which Audio Clips Get Popular: Comparing Metadata and Sound Features
SummaryThis paper studies how to predict the popularity of music and sound effects on an online sharing platform. Popularity is treated as a classification task based on download counts and average user ratings. The authors compare Random Forest, XGBoost, and neural network models using metadata, platform descriptors, and eGeMAPS acoustic features. Metadata gave the most informative predictions, while acoustic features alone performed significantly above chance.
Ikram El Hadi, Nihad Elbari, Kawtar Tlem\c{c}ani et al.
In briefWhen repeated measurements hurt tabular foundation models
SummaryThis paper studies how tabular foundation models (TFMs) respond when the same inputs are measured repeatedly. Across six TFMs in synthetic and semi-synthetic experiments, extra measurements can improve predictions at measured inputs but degrade them elsewhere, with harm most frequent at low noise. The authors attribute part of the harm to row-count dependence in TabPFN-v3's attention scaling, and they compare strategies such as averaging measurements and resetting the count.
WhyShows repeated measurements can harm tabular foundation models, explained by protocol mismatch and row-count attention scaling, validated across six models and datasets.
Yixiang Liu, Ruizhe Deng, Bibhas Chakraborty et al.
In briefOne-layer Transformers with chain-of-thought can provably learn in-context K-nearest neighbor prediction
SummaryThis paper studies how chain-of-thought (CoT) lets a one-layer transformer learn in-context K-nearest-neighbor (K-NN) prediction. Prior work handled only 1-NN, and the authors show that softmax attention struggles to attend to the 2nd through K-th nearest neighbors, with a rigorous negative result for odd K>1. They give an explicit one-layer construction that solves K-NN via CoT and prove that gradient descent can recover this construction under a stylized training setup.
In briefTPPO: Tracking-Value PPO for Language Model Reasoning
SummaryThis paper addresses the problem that standard value critics struggle to track how value evolves along a reasoning trajectory when rewards arrive only at the end. It proposes Tracking-Value PPO (TPPO), which models successive prefix values as additive updates in log-odds space and builds the value from summed position-wise outputs. The authors report gains over PPO on mathematical reasoning benchmarks and on several other models and tasks.
WhyProposes a cumulative log-odds value parameterization improving PPO value tracking; multi-model, multi-task experiments support gains useful to RLVR research.
In briefViewStruct: Predicting how object parts connect and move from a single view
SummaryViewStruct is a framework that jointly predicts how rigid parts of objects and scenes are grouped, attached, and articulated, using an RGB image with part segmentation and depth. It uses a pretrained vision-language model to predict structure in seconds, without observing motion or per-instance optimisation. Trained only on synthetic data, it outperforms evaluated baselines and transfers to unseen categories and real-world human-interaction videos. Simulated experiments also show the predicted joints can generate robot motion trajectories.
In briefFaster preconditioned optimization using frequent-directions matrix sketching
SummaryAdaptive optimization methods perform well but require maintaining a preconditioning matrix, which is costly in large-scale settings. This paper accelerates them with frequent directions (FD), a matrix sketching technique, introducing FTSL with a tighter regret bound than prior work. It extends the approach to matrix-valued variables (FTSL-Shampoo) and one-sided variants to reduce computation. The methods are also extended to offline stochastic convex optimization and validated empirically across a range of tasks.
In briefInfoVis: A unified model for generating and editing presentation slide images
SummaryThe paper defines four slide-centric tasks and builds a data engine with over 13M continual-pretraining and 1.9M fine-tuning records. It introduces VisSlideBench, a benchmark of 1,527 screened instances with multidimensional evaluation whose rankings agree with human judgments. The authors train InfoVis, a 9B unified slide generation and editing model, which leads all evaluated open-source models on all eleven metrics and scores 7.94 overall versus 7.99 for Nano-Banana-2.
WhyFormulates four slide tasks with large-scale data and benchmark, unifying generation and editing; validated across eleven metrics with human agreement checks.
In briefLIBERO-Industry: A benchmark for testing robot manipulation under industrial process limits
SummaryThis paper presents LIBERO-Industry, a robot manipulation benchmark built on LIBERO that turns industrial constraints such as product variants, moving material flows, limited execution windows, and tight tolerances into controllable tasks with measurable success criteria. It includes 216 tasks across three scenarios: static-to-dynamic loading, dynamic-to-static unloading, and dynamic-to-dynamic transfer. Experiments on five generalist robot policies show performance degrades progressively as these constraints are introduced and combined.
Ming Shin Cindy Sung, Duan Leijiahao, Haoxiang Wang et al.
In briefA Bethe-ansatz source of early token correlations for parallel proposals
SummaryThe paper builds probability laws from the periodic XXZ spin chain using coordinate Bethe ansatz and compares them with maximum-entropy pairwise projections matching the same one- and two-site moments. In an eight-site, three-excitation sector the full law difference is one-dimensional, set by a three-site moment δ, and vanishes at Δ=(3+√3)/2. Two readouts give distances 2|δ| and 4|δ|, and a strict speculative accepted-prefix advantage is proven. The authors stress these are finite-law existence results, not measured language-model or wall-clock gains.
In briefGame-Creation-Agents: Autonomous Game Production
SummaryThis paper presents Game-Creation-Agents (GCA), a system that turns a natural-language request into a runnable Phaser 4 browser game. It first builds a multimodal game design document, then uses eight specialized agents to implement scenes and produce assets before merging them into one build. Build-linked UI checks and a code-blind critic verify the result, and blockers are sent back for repair. The authors report that GCA outperforms a comparable open pipeline on requirement satisfaction, mechanics, content depth and visuals, and was also preferred in a human side-by-side study.
Tencent Multimodal RL Team: Xiangxin Zhou, Jiaqi Tang, Ke Ma et al.
In briefUsing geospatial vector geometry as a built-in spatial input for vision-language models
SummaryVision-language models struggle with accurate spatial reasoning at large geographic scales, because limited visual views make spatial references and relations hard to recover. The paper introduces GeoVec-VLM, which builds a structured vector geometry representation as an explicit spatial basis and lets it interact with visual and language inputs. Across five spatial reasoning tasks, it improves mean accuracy by about 11% over the strongest baseline, keeps its gains under geographic transfer, and shows better geometric consistency.
In briefRelational Learning as World Modeling: Inferring How Relational Systems Change From Executions
SummaryThe paper introduces TabWM (Tabular World Model) and WorldRel, a generator of executable relational worlds for pretraining. Given observed executions from an unseen world as context, TabWM predicts sparse structured edits caused by a new action without parameter updates. Across 512 held-out worlds, complete-edit F1 rises from 0.016 with no observed executions to 0.685 with 64. Because predictions are explicit relational edits, they can be applied recursively to imagine candidate futures and compare actions.
In briefCombining Several Keypoint Detectors to Find Complementary Points for Image Matching
SummaryThe paper proposes Mixture of Detectors (MoD), a framework for dense matching that combines different keypoint detectors. It first distills heterogeneous detectors into one unified model, then uses a lightweight router to pick the best detector for each image patch. The router is trained with rewards based on downstream matching quality. Experiments on MegaDepth, ScanNet, and HPatches report consistent gains over strong baselines across various keypoint budgets.
In briefJudging Harm From Partial LLM Answers, Not Just Prompts
SummaryThe paper argues that most LLM safety guardrails judge only the prompt and reject it before any answer is generated, which blocks many requests that would have received safe responses. The authors built a dataset of 41,944 prompt-response pairs from seven open-weight models across the Llama, Gemma, and Qwen families. They found prompt and response labels agree on only 50.46% of items (Cohen's κ = 0.06). Fine-tuned classifiers predicted response-level harm well from partial outputs, with best AUPRC 0.885 and F1 0.807.
Ram Vikas Mishra, Shashank Aggarwal, Jahnavi Kumar et al.
In briefImage tokenizers should be trained between images, not only on them
SummaryLatent diffusion models generate images by decoding codes from an image tokenizer, but generated codes often fall between training codes, where the decoder can produce flawed images. The paper proposes InterpTok, which fine-tunes MAETok so its decoder turns mixtures of two image codes into natural images. A discriminator checks that each mixture looks as natural as its sources, and a mixing ratio loss keeps the mixture faithful to the two codes. The authors report that InterpTok keeps reconstruction quality while improving interpolation and generation FID.
Ole Bei{\ss}wenger, Jan-Niklas Dihlmann, Hendrik Lensch
In briefDoes the Platform Affect the Measurement? A Protocol for Testing Cross-Platform Equivalence of Machine-Labelled Policy Sentiment
SummaryThe paper examines whether sentiment scores from a machine classifier can be compared across social media platforms, an assumption it argues is rarely tested. Adapting psychometric equivalence methods, the authors propose a six-step protocol and run a simulation with known ground truth. In that simulation, unadjusted pooled indices gave the wrong aggregate sentiment sign on nearly a quarter of days, while adjusted indices did so on under one day in sixty on average. A case study of UK economic-policy comments from four platforms motivates the design, and the full study is left for follow-up work.
In briefMultilingual Artificial Hivemind: Do All LLMs Think Alike in Every Language?
SummaryThis paper examines whether the limited output diversity seen in English text from large language models also appears in other languages. Across 23 open-weight models and 13 languages, the authors find the homogenization effect is widespread but uneven. The within-model effect holds across three embedding spaces and a lexical metric that uses no embedding model. The size of the gap tracks each model's language competence on a held-out translation benchmark, suggesting a performance-diversity tradeoff.
Xaver Raphael Davey, Shubhra Mishra, Anton Adel\"ow et al.
In briefNFMap: A Reversible Method for Drawing Data Maps in Two Dimensions
SummaryThis paper presents NFMap, a dimensionality reduction method for 2D visualisation that extends Generative Topographic Mapping (GTM) by replacing its RBF mapping with an invertible normalizing flow. Any point on the map can be mapped back to data space, and heavy-tailed Student-t mixture components are supported to reduce crowding on large datasets. The authors state that the residual scale is the main control between local and global distance fidelity.
In briefThe Midas Touch for Metric Depth: A Interpretable Method for Recovering Metric Depth from Monocular Estimates
SummaryFoundation models have greatly improved monocular relative depth estimation, but converting these estimates to accurate metric depth remains hard due to scale ambiguity and local scale inconsistency. This paper extends the authors' previous MTD work to MTDv2, a hybrid framework combining explicit depth metricization, which applies segment-wise affine transformations and pixel-wise refinement, with a learnable multi-path data-driven refinement. Experiments on 16 public datasets show robust and generalizable depth scale recovery across different sparse observation sources and depth estimation paradigms, as well as applicability to downstream 3D perception tasks.
In briefHidden-state geometry before attention sinks form, and how it can be changed
SummaryThis paper studies whether measurable changes in hidden-state geometry appear before attention sinks form in language models. Comparing prompts within the same model and layer, the authors find that simple spectral summaries of hidden states add predictive information about sink timing, most clearly about two to three layers before formation. Modifying only the singular values of post-layer hidden states changes later sink behavior in Qwen2.5, GLM-Z1, and GPT-OSS relative to norm-matched random perturbations.
In briefFoveaTree: Query-guided pruning of visual tokens to make video AI models cheaper to run
SummaryVideo large language models convert each frame into hundreds of visual tokens, which makes video processing expensive. FoveaTree is a training-free, two-stage pruning method that first builds a quadtree-based density map from merge-gain estimates to decide which tokens to drop, merge, or keep. It then refines text-to-vision attention in intermediate layers using the surviving tokens and the inherited tree structure. Tested on InternVL3.5-8B and Qwen3-VL-8B across five video question-answering benchmarks and one captioning benchmark at 50%, 30%, and 10% token retention, it reports overall state-of-the-art performance.
In briefCopyQuad: Transferring Quadruped Motion from One Side-View Video Without a Depth Model
SummaryCopyQuad converts a single uncalibrated side-view video into a quadruped robot motion reference without a monocular depth model or per-clip 3D reconstruction. A frozen 2D keypoint detector and fixed analytic retargeting produce reachable foot targets, and one phase-indexed reward structure learns dog and horse gaits, which are deployed zero-shot on hardware. The paper reports that the lateral axis it omits is smaller than the pipeline's existing foot error in motion-capture measurements.
In briefAmortised adaptive importance sampling: a trained lightweight sampler that adapts to new targets
SummaryThis paper proposes an amortised adaptive importance sampling algorithm, in which a lightweight neural network adapts the proposal distribution iteratively using information gathered during sampling. By training across many targets from the same distribution family, the sampler can adapt to a specific target without prior access to its parameters. The authors report that it samples efficiently from targets within the training family, outperforming target-specific amortised samplers and training-free adaptive importance sampling. They also explore generalisation to targets outside that family.
Arran Carter, V\'ictor Elvira, Esmeralda S. Whitammer
In briefDetecting Rooftop Solar Systems on Industrial Buildings from Satellite Images with Machine Learning
SummaryThis paper uses free, open Sentinel-2 multispectral satellite imagery and OpenStreetMap building outlines to train a Random Forest model that recognizes the spectral signatures of solar PV modules on industrial rooftops. The authors also build an interactive web prototype that lets users retrieve recent satellite observations and check for PV systems. The tool includes quality metrics indicating whether a prediction has enough training data support or needs manual verification. The authors conclude that open medium-resolution data can help close the monitoring gap.
Frederick Bruch, Christian Knoth, Florian Detsch et al.
In briefVAttnPO: Visual Attention-Guided Policy Optimization for Multimodal Reasoning
SummaryThis paper addresses multimodal reinforcement learning, where a single reward is applied to a whole response and token-level changes in visual-evidence use are left implicit. The authors propose VAttnPO, which compares each generated token's visual attention with a reference built from a frozen policy given the correct answer, using Jensen–Shannon divergence, and uses this to adjust the GRPO sequence advantage. On Qwen3-VL-8B-Instruct, six-task image AVE rises from 70.97 with GRPO to 73.50, and gains also appear on Qwen2.5-VL-7B-Instruct and InternVL3-8B-Instruct.
In briefPathPilot: A Question-Adaptive Multi-Magnification Agent for Navigating Whole-Slide Pathology Images
SummaryThis paper presents PathPilot, an AI agent that navigates whole-slide pathology images across multiple magnification levels, mirroring how pathologists move between scales. It includes CrossMagLoc, which identifies the most diagnostically relevant region to start from, and EviRegionSearch, which gathers question-relevant evidence and adaptively chooses the next navigation step. The authors report that PathPilot outperforms competing state-of-the-art methods on question-answering evaluations across five cohorts.
In briefQueue-aware token communication linking AI representations to wireless service
SummaryThis paper proposes Queue-Aware TokenCom, which adds token queues to the AI service pipeline to handle mismatches between token generation and time-varying wireless capacity. It presents an end-to-end architecture covering AI representation generation, token workload management, wireless transmission, and downstream task inference. The authors examine token budget strategies, both fixed and adaptive, to characterize the tradeoff between AI performance and communication efficiency in dynamic wireless environments.
In briefTrace-to-Architecture: diagnosing which agent harness parts to fix from execution traces.
SummaryThis paper introduces Trace-to-Architecture (T2A), a framework that uses execution trajectories from successful and failed tasks to decide what to change in an LLM agent's harness and when those changes apply. A development agent diagnoses failures, revises the relevant components, and checks whether target tasks improve and successful tasks stay successful. T2A achieves the highest aggregate score on ALFWorld, ScienceWorld, and AppWorld across all three task agents. Integrated into Meta-Harness, it raises mean ScienceWorld score by 21.6 points.
In brief"Lost in Contrast": Decoding That Preserves Instructions for Reliable Visual QA
SummaryThe paper observes that visual contrastive decoding can turn correct abstentions into wrong answers during answer correction. The authors find abstentions are lost at two stages: candidate filtering can drop the abstention token before comparison, and subtracting the reference prediction removes its response to the abstention instruction. They propose instruction-preserving decoding, yielding IP-VCD and IP-M3ID, which outperform all compared non-IP baselines on decision utility across all nine models on VizWiz. On Food-101, IP-VCD also leads on eight of nine models.
In briefLearn from Move: A Next Step for GUI Agents
SummaryThe paper introduces LatentGUIWorld, a six-task benchmark in three matched pairs that tests whether GUI agents can explore unfamiliar interfaces and learn their dynamics. It proposes MOVE, which separates movement from button events, shows intermediate visual feedback, and uses iterative exploration demonstrations. After SFT, LatentLearner-9B reaches 50.0% overall success, and joint reinforcement learning raises two egocentric tasks to 92.0% and 75.3%.
In briefWhere Is the Self of an LLM? From Self-Index to Self-Reference
SummaryThe paper distinguishes two levels of self, a self-index that separates a model's own material from external material and shapes processing, and self-reference, the ability to recognize what it wrote. The authors test eight open-weight models from four families by placing identical text in Assistant, User, or System messages. They find the two levels are nested and do not track model size, and argue that self-reference is built on top of the self-index.<|im_end|>
WhySeparates self-index from self-reference and tests them with interventions, but evidence rests on a limited set of models.
In briefChecking Teacher Reasoning Before Distilling It into Small Medical Models
SummaryThe paper studies when teacher-generated rationales help a smaller model answer Brazilian Portuguese health professions exam questions, and when they reinforce errors. Based on these findings, the authors train only on rationales attached to correct teacher answers and propose Distill-SFT, which merges the answer and rationale into a single target. Distill-SFT showed a modest advantage over step-by-step distillation on previously missed questions and, with rationale-based training, reduced catastrophic forgetting on the out-of-domain ENEM dataset compared with answer-only training.
Tiago Martinelli, Jo\~ao Paulo Papa, Adriano Jos\'e Pereira
In briefUsing attention-graph curvature to verify mathematical reasoning
SummaryLanguage models' natural-language proofs cannot yet be checked by proof assistants, and an early error can invalidate all later steps, so detecting errors early matters. The paper studies whether the directed Forman curvature of a model's attention graph, a local value computed from nearby attention weights, carries information about whether a reasoning step is correct. Four per-step verifiers were tested on ProcessBench with and without curvature added to their input. Adding curvature improved the overall mean performance of all four, with the largest gain for ReProbe, which takes every token of a step as input. Computing curvature increased the language model's forward-pass time by 18% on average.
In briefUniversal distributed feature extraction that complements side information, using H-Score
SummaryThis paper studies a distributed estimation system in which edge nodes send averaged features to a center node for parameter estimation. The authors orthogonally decompose the joint function space and show, under weak dependence, that modal decomposition of the canonical dependence kernel's conditional-additive component yields task-independent optimal features. They build a distributed nested H-score to learn this decomposition. With a limited feature budget, the mean-squared error gap to the best task-specific features depends only on the task's alignment with the discarded least informative modes, and vanishes to leading order for square-integrable tasks once the budget is sufficient.
In brief3D-VITAL: Visibility-aware training that uses 3D mesh lifting to improve matching of objects seen from different viewpoints
SummaryThis paper tackles cross-view object re-identification, matching the same object across very different viewpoints such as aerial and ground cameras. It lifts source images into textured 3D meshes with single-view reconstruction and renders them from new viewpoints as extra training data. The authors note that reconstructions can be geometrically wrong and that rendered regions can contain hallucinated textures, so they add a consistency filter and a visibility-guided attention supervision. The method is reported to set new state-of-the-art results on three benchmarks.
William Grolleau, Astrid Sabourin, Guillaume Lapouge et al.
In briefEHR-Bench: A Chinese Multi-Task Benchmark for Evaluating Recruitment AI
SummaryThis paper introduces EHR-Bench, an open Chinese benchmark with twelve tasks across six capability families and five recruitment scenarios, including generation, extraction, classification, retrieval, semantic similarity, and reciprocal ranking. The authors built 8,336 task instances from public materials and controlled augmentation, with expert review. They evaluated twelve LLMs at task, capability, and aggregate levels, finding that closely ranked models have distinct strengths and that performance varies with input sources, reference evidence, and candidate pools.
In briefSpectral costs of multinomial and residual resampling in high-dimensional particle ensembles
SummaryThe paper studies how a single resampling step in particle filters changes the spectrum of the particle second-moment matrix in a high-dimensional regime where dimension over particle count converges to a constant. For any unbiased scheme, the resampled matrix dominates the original in conditional convex order while keeping the conditional mean, with excess second spectral moment equal to the Frobenius energy of the perturbation. It derives limiting count laws for multinomial and residual resampling and draws exact spectral consequences from them. The authors also show that the effective sample size cannot determine the limiting spectrum.
In briefEvoSpatial: A Spatial Reasoning Agent That Learns Reusable Skills and Tools From Experience
SummaryThis paper presents EvoSpatial, a framework in which spatial reasoning agents learn from execution experience by updating a linked library of skills and tools. Skills record procedures, geometric assumptions, and verification steps, while tools implement the required operations. Across 20 benchmarks, EvoSpatial averages 63.9, 4.0 percentage points above the average reported by SpatialClaw. The library is updated without ground-truth answers or benchmark scores.
In briefAmortized variational inference of latent intensities from repeated point patterns
SummaryThe paper proposes a neural-network amortized variational framework that learns a reusable rule mapping a variable number of replicated point patterns to a distribution over their shared latent intensity. It is trained from event times using the point process likelihood, without ground-truth intensity supervision, and performs inference in a single forward pass at test time. On synthetic data, more realizations improve intensity recovery and narrow credible intervals. On bicycle-sharing data, it achieves the highest mean held-out log likelihood among the evaluated methods with substantially less inference time than the Log-Gaussian Cox Process baseline.
Wenqing Liu, Xenia Miscouridou, and D\'eborah Sulem
In briefUsing LLM-Generated Prompts to Improve Open-Vocabulary Video Recognition
SummaryThis paper studies how to apply CLIP-style vision-language models to open-vocabulary video recognition, where limited text information can cause overfitting. The authors propose P2CLIP, which uses prompts at the category, video, and frame levels, including discriminative category prompts generated by LLMs to reduce semantic overgeneralization. It also introduces a knowledge-guided temporal aggregation mechanism to improve vision-text alignment. The authors report experimental results across zero-shot and few-shot benchmarks.
In briefMaking LoRA Identifiable: Orthogonal Alignment for Continual Learning
SummaryThis paper addresses using parameter-efficient fine-tuning (LoRA) for continual learning, where existing methods either grow memory with the number of tasks or align representations poorly within a single LoRA. The authors propose SOAR, a two-stage framework that constrains the A matrix to the Stiefel manifold and uses Orthogonal Procrustes alignment to estimate an optimal rotation for the B matrix, plus a per-rank retention merging strategy. Experiments across multiple models and benchmarks show that SOAR achieves strong continual learning performance without additional memory overhead.
Pratik Rakesh Singh, Mohammadi Zaki, Akash Saha et al.
In briefStructural baselines matter for detecting and attributing agent failures
SummaryThe paper studies LLM tool-using agents that fail silently, stopping early without raising an error. It compares two monitors: one that detects whether a run failed, and one that identifies the decisive step where failure became unavoidable. The authors show that model-free regularities, such as longer failing runs and the last step being labelled decisive, already yield high scores. They argue that results should be reported as gains over these baselines, which removes the advantage of several methods on some splits, while agent internal states outperform text baselines for detection.
In briefCompression Through Time: How Class Separability, Stationarity, and Capacity Shape Recurrent Networks
SummaryThis paper studies how information is compressed over time in recurrent LSTM networks, finding strong compression in the final timesteps across many datasets. The authors propose three interacting mechanisms: class separability, temporal stationarity, and limited network capacity. They test these through controlled synthetic interventions and real sequence tasks. The results suggest compression can act as a regularizer that helps when spare capacity exists, but can hurt generalization when it uses memory the task needs.
In briefRFWeave: A learned model of wireless propagation that generalizes across scenes and transceiver setups
SummaryThis paper introduces RFWeave, a learning-based wireless propagation model that generalizes across scenes, transceiver locations, and transceiver configurations. It represents individual antenna elements with variable-length tokens and encodes transmit-receive spatial relations, then uses a configuration-conditioned decoder to predict spatial spectrum and received power. Trained on a simulated dataset of 10,000 indoor scenes and over 5 million channel records, it reports improvements over the strongest baselines on unseen scenes, with SSIM gains up to 15.7% and power MAE reductions up to 35.4%.
In briefDoes a hosted AI decision model stay consistent when extra answer options are added? A pre-registered test of jev-1.13.0
SummaryThe study tests whether the hosted decision model jev-1.13.0 follows independence of irrelevant alternatives, using 299,463 calls on intent and topic classification tasks. Adding an unrelated option systematically shifts the odds between the correct label and its strongest rival, and the distortion grows with the number of options. The authors conclude that the model's probabilities are best interpreted relative to the specific option list used to compute them.
In briefGuidance Without Labels: Flow Matching from Coupled Gaussian Mixtures
SummaryThis paper improves how flow matching generative models are trained. Instead of a Gaussian source, it uses a fitted Gaussian mixture and couples each data point to a mixture component via entropic optimal transport. The authors show this reduces to the standard Gaussian-source problem plus a guidance-like correction, so a pretrained velocity field can be reused. On published CIFAR-10 and CelebA-HQ checkpoints, it improves FID at every step count and matches baseline quality on CelebA-HQ with about half the steps.
Stanislas Ducotterd, Bassam El Rawas, Michael Unser
In briefCounting with attention and LayerNorm: how norms and representation shape cost
SummaryThis paper studies how a single attention layer counts tokens up to a maximum length L, focusing on softmax attention and layer normalization. It defines B_min, the smallest norm bound on trainable vectors needed for exact counting after rounding, and derives its growth rates across normalization placements. Learned post-normalization counts at a smaller norm (about L^{3/4}) than softmax without normalization (about L^{3/2})."}
In briefIND-Eval: A benchmark aligning evaluation and simulation for industrial time series models
SummaryThe paper introduces IND-Eval, a public benchmark that groups industrial time series from different domains into source-specific forecasting tasks, with finer sampling and more multivariate forecasting than GIFT-Eval. Among Chronos-2, TiRex-2 and TimesFM-3, TimesFM-3 achieves the lowest source-balanced aggregate nMASE and nCRPS. A case study on two offshore well-monitoring regimes shows that an oscillator-augmented simulator lowers target-domain error but hurts performance on GIFT-Eval's univariate subset, while a 50/50 mixture of the two simulation recipes gives the lowest error and recovers most of that loss.
Vebj{\o}rn Haugland (1), Pavel Gulin (1), David Foss (1) et al.
In briefSIREN: A Framework for Checking Whether Crisis Translations Keep Emergency Instructions Actionable
SummaryThe paper introduces SIREN, a human evaluation framework that assesses whether machine translation preserves protective action, directive force, and urgency in emergency messages. At the error level on Italian-English translations, 31.63% of ChatGPT-4o errors and 48.73% of Google Translate errors compromised actionability, mainly affecting protective action. At the message level on English-French translations, actionability was preserved in 85.34% of TranslateGemma-4B translations and 92.67% of Google Translate translations. The authors also release an English-French crisis translation resource with actionability annotations and post-edited references.
Maria Carmen Staiano, Lifeng Han, Johanna Monti et al.
In briefLearning incompressible 3D flows from sparse samples with a set transformer for divergence-free kernel fitting
SummaryThe paper introduces DFK-Former, a transformer that takes sparse, noisy, irregularly sampled 3D velocity observations and reconstructs a divergence-free velocity field. It predicts divergence-free kernel parameters directly in one forward pass instead of running a separate optimization for each field. According to the paper, it reduces endpoint error by up to 4.6x versus per-field optimization and cuts reconstruction time from 9.8 s to 7.5 ms.
In briefLearning stable action-response shapes under time-varying baselines
SummaryThis paper studies how to learn the response of continuous decisions from observational time series when outcome levels drift over time, which can make pooled models confuse action effects with shifting baselines. The authors propose deployable causal action geometry, the action-response shape left after removing action-independent level shifts on a chosen link scale and profiling out amplitude variation, together with a two-stage orthogonal estimator. In a randomized Taobao A/B test, the deployed system improved GMV and purchase counts by 1.16% and 1.70%.
In briefSuPER-Mamba: Surprisal-Based Routing to Extend Context in Mamba Models
SummaryThis paper proposes SuPER-Mamba, a training-free method that uses surprisal, computed from next-token predictions, to route long-range states in pretrained Mamba models. Neighboring surprisal values are aggregated via local pooling and used to prioritize updates to global channels. On three state-space models evaluated on LongBench-E, the average score rose from 8.06 to 15.69.
In briefACED: A Framework Combining Natural-Language and Executable Evidence for Math Reasoning
SummaryThis paper proposes ACED, which treats the natural-language self-consistency answer as the default reference and executable program outputs as supporting evidence rather than equal votes. It fuses the two evidence sources with a weighted logarithmic pool and overrides the reference only when a calibrated relative-evidence margin is exceeded. Across five math word-problem benchmarks, ACED matches or outperforms the baselines. On GSM8K, ACED-ALL reaches 86.9% accuracy versus 82.6% for direct Program-of-Thought aggregation.
In briefLanguage2Drive: Do Driving VLA Models Actually Follow Instructions?
SummaryThis paper introduces Language2Drive, a benchmark that tests whether vision-language-action (VLA) driving models truly act on natural-language instructions. It keeps the driving scene fixed while changing only the instruction, using paired counterfactual units, and evaluates in CARLA simulation, reconstructed real-world scenes, and open-loop benchmarks. The results show that 93.2% of successful rollouts have no successful counterpart under the alternative instruction, often because models repeatedly pick one behavior. The authors argue that standard success and driving scores can overstate instruction-following ability.
WhyIntroduces paired counterfactual units to test whether driving VLAs truly follow instructions, validated across multiple simulation and open-loop settings; impact mostly within the subfield.
In briefREEL: Learning to Retrieve, Evaluate, and Evolve Visual Memory for Long Multi-Shot Video Generation
SummaryThis paper introduces REEL, a general-purpose visual-memory planner that learns through a unified tool interface to retrieve references, create missing visual evidence, and update persistent memory. Backend-specific adapters convert its plans into conditioning inputs so the same policy can guide different black-box video generators. The authors train REEL with supervised fine-tuning and two GRPO stages on two newly built datasets. On JoyAI-Echo~1.5 and MiniMax-H3, REEL improves cross-shot consistency on ST-Bench by 15.1% and 20.7% over a no-memory baseline.
In briefAeroChem: A Closed-loop Physics-Informed Model for Long-term Air Quality Forecasting
SummaryThis paper proposes AeroChem, a closed-loop physics-informed state space framework for forecasting air quality involving chemical reactions. It uses two branches: a UDE-based process branch offering an imperfect physical prior, and a Mamba-based branch that captures long-range context and produces probabilistic virtual observations, linked by Neural Kalman Fusion. The authors also introduce HanoiAir, a real-world dataset. Experiments reportedly show improved long-horizon stability and sudden-change forecasting.
Kien Le Trung, Tung Kieu, Dinh Duc Nha Nguyen et al.
