In briefVolCo: Volumetric contact representation for more precise, less-penetrating human grasp generation
SummaryThe paper introduces Volumetric Contact (VolCo), a representation that expands object-surface contact points into 3D volumetric grids to capture hand-object contact more precisely. Its framework, VolCoDiff, uses a 3D variational autoencoder for local contact details conditioned on object SDF, and a prior-guided diffusion model for global hand geometry. On two benchmark datasets, the method reports state-of-the-art penetration and stability results, generating tighter grasps with fewer severe penetrations.
In briefHuLiGen: Generating Human LiDAR Point Clouds from Body Models
SummaryReal LiDAR point clouds of humans are costly to collect and label, which limits research on human analysis with this sensor. This paper introduces HuLiGen, a generative model that creates human LiDAR point clouds from a parametric body model using a point transformer trained with flow matching. The authors report that the generated data is closer to the real capture distribution. Pretraining on synthetic data from HuLiGen improves LiDAR-based human pose estimation, reducing MPJPE by up to 50% in low-annotation and low-data settings.
In briefA general approximate MCMC method for switching SDEs using uniformization and neural likelihood estimation
SummaryThis paper proposes an approximate Markov chain Monte Carlo sampler for Bayesian inference in switching stochastic differential equations (SSDEs), whose parameters change with a latent continuous-time Markov chain regime. The method combines uniformization with factorized neural likelihood estimation (FNLE), using a time-conditioned FNLE model to approximate transition densities over arbitrary time intervals. In synthetic experiments it recovered regime paths and parameters for three SSDE models where prior methods were limited. It was also applied to a real dataset, where it detected a regime transition.
In briefUniversal local error and realized amplification in the first-order EDM sampler
SummaryThis paper analyzes the first-order deterministic diffusion sampler EDM from Karras et al. (2022) in 2-Wasserstein distance, separating local discretization error from its amplification by later learned steps. The authors prove that the one-step error is quadratic in step size, with a constant independent of the data distribution. Experiments on a one-dimensional Gaussian mixture show that measured amplification explains slower error decay on finite sampling grids.
In briefFaster Posterior Sampling for Image Reconstruction with Diffusion Priors
SummaryThis paper proposes RED-KLwSGS, a sampler for Bayesian imaging inverse problems that keeps an exact Gaussian update for the data variable and updates the auxiliary variable with kinetic Langevin diffusions driven by a one-shot denoising score. It runs at the same per-iteration cost as Langevin-within-SGS. The authors prove non-asymptotic Wasserstein-2 convergence for strongly log-concave priors and introduce Joint-RED-KLwSGS, which applies kinetic Langevin to both variables. Experiments using diffusion probabilistic models as priors on FFHQ and ImageNet show faster convergence and high-quality reconstructions.
In briefTuning Bayesian optimization settings by running Bayesian optimization on simulated data
SummaryThis paper proposes a framework for choosing the parameters of a Bayesian optimization algorithm before it runs. It draws sample paths from a Gaussian process fitted to the information available at the start, and measures each parameter setting by its cumulative regret on those paths. Because this parameter search is itself a black-box problem, the authors solve it with a second, outer Bayesian optimization, and experiments show it can select well-performing configurations.
Satoshi Katayama, Shoyo Hunt, Shintaro Masuda et al.
In briefAI Agents Help Build Production Scheduling Models with Constraint Programming
SummaryThis paper tests whether general-purpose large language models, used as agents without task-specific training, can turn natural-language production scheduling problems into constraint programming models. It compares single-agent, multi-agent, and direct LLM approaches on six industry-oriented problems, using a Model Context Protocol server to retrieve solver documentation and reduce hallucinations. Formulation is largely achievable with current LLMs, while implementation is the main barrier; the multi-agent workflow raised the share of correctly running scripts from 14.8% to 59.3%. Tightly coupled intralogistics models remain an open challenge.}
\'Angel S\'anchez-Fern\'andez, Javier Pernas-\'Alvarez, Diego Crespo-Pereira
In briefVideoEvolve: Co-Evolving Memory and Retrieval for Long Video Understanding
SummaryThis paper presents VideoEvolve, a framework for understanding long videos that updates both what is stored in memory and how it is retrieved. It starts from a coarse low-frame-rate overview, with a Memory Evolver adding selected content to memory and a Retrieval Evolver searching that memory. The two components are trained alternately through agentic reinforcement learning, guided by feedback that targets the current bottleneck and underdeveloped video capabilities. The authors report experiments on several long video understanding benchmarks showing its effectiveness.
In briefEEG and eye-tracking study on how AI labels change face judgments
SummaryThis study compares AI-generated and real human faces, pairing each with no label, a "Made with AI" label, or a "Made by a human" label. Brain recordings (ERPs) and gaze tracking show that the faces alone produced equivalent responses, but labels drew early attention (N2), and labels conflicting with the actual source prompted re-evaluation (P3). Labels also altered visual exploration. The authors provide a validated stimulus set.
Teodora Mitrevska, Luise Donat, Andreas Butz et al.
In briefFrom Log-Odds to Shapley Values: A Geometry for Explaining Weighted Naive Bayes
SummaryThe paper builds an explanatory space for a weighted naive Bayes classifier from the supervised representation the model induces. It reformulates the classical distance based on conditional log-likelihoods using log-odds, which relate more directly to classification decisions. The authors show this distance exactly equals the ℓ1 distance between analytical Shapley value vectors, and compare several supervised distances with a k-nearest neighbors classifier.
In briefArgos: Reusing Geometric Priors for Generalizable Online Scene Change Detection
SummaryThe paper presents Argos, which adapts features from Geometric Foundation Models to jointly perform scene change detection and 3D reconstruction, aiming to handle viewpoint changes, occlusions, and noise. The authors build a large benchmark with two synthetic datasets and one real-world dataset and train across them to improve cross-domain generalization. They also introduce Argos-SLAM, a real-time robotics system for online change detection and change-aware 4D mapping. The paper reports gains of up to 42.01% in change IoU and 27.91% in F1 over existing baselines.
In briefBeyond Outcome Rewards: Assigning Retrieval Credit to Search Agents
SummaryThis paper studies how intermediate supervision can improve reinforcement learning for search agents that iteratively retrieve information to answer multi-hop questions. The authors examine several reward-shaping and credit-assignment strategies that derive learning signals from intermediate retrieval steps. They propose a training framework combining these intermediate signals with final outcome rewards, and report improved aggregate performance across benchmarks under matched training conditions. The results indicate that both the choice of intermediate signal and where its credit is assigned affect training behaviour.
In briefFast Linear-Time Context Compression for LLMs Using State-Space Encoding
SummaryLong contexts make Transformer-based language models expensive in computation and memory. The authors propose Boa Compressor, a plug-and-play online compressor built on a Mamba state-space model that encodes context of any length in one linear-time pass. The compressor is trained by distillation so that a frozen target LLM produces the same next-token distribution from the compressed representation as from the full context. The paper reports that Boa matches or surpasses existing offline baselines at equal compression ratios, runs much faster than Transformer-based compressors, and remains effective up to 16x compression.
Marc Serra Ortega, Andrey Barsky, Carlos Boned Riera et al.
In briefA Unified Information-Theoretic Approach to Constrained Multi-Fidelity Multi-Objective Bayesian Optimization
SummaryThis paper addresses Bayesian optimization with multiple objectives, constraints, and fidelity levels, focusing on jointly choosing evaluation points and fidelities to find the highest-fidelity feasible Pareto frontier. It measures query utility by the information gain about this frontier and, since the mutual information is intractable, derives a variational lower bound using mixed under- and over-truncated approximations. Multi-fidelity surrogate models propagate information across fidelities to produce a cost-aware acquisition function. Experiments on synthetic, benchmark, and real-world problems are reported.
In briefDoes Document Structure Help Dense Retrieval? A Placebo-Controlled Test of Four Methods
SummaryThis paper tests four document-structure methods for retrieval-augmented generation: structure-aligned chunking, LLM-generated chunk context, heading-path metadata, and hierarchical two-stage retrieval. It runs them under one protocol with a placebo that shuffles heading paths across documents. Experiments on 200 Wikipedia Featured Articles and 1,585 QASPER papers show that structure-aligned chunks with real heading paths perform better, while naive two-stage hierarchical retrieval hurts. Effects are small but consistent across both corpora.
Andrey Kuehlkamp, Priscila Correa Saboia Moreira, Samuel Rund
In briefConformal prediction for spatial data via sequential whitening
SummaryThis paper improves conformal prediction intervals for spatially dependent data collected with nonrandom sampling. It sequentially conditions on calibration residuals to remove spatial variation that only they can predict, with nearest-neighbour approximations for large networks. Under a correct working covariance and elliptical residual law, the intervals have exact finite-sample coverage under any spatial design. In simulations and a national PM2.5 application, the intervals are narrower and more stable, and a diagnostic flags regions at risk of undercoverage.
In briefPolicy Learning with Weak Signals: Personalized Treatment Decisions in Noisy Experiments
SummaryThis paper studies policy learning in digital experiments with weak signal-to-noise ratios, rich covariates, and large data volumes. It models treatment-effect estimates from finer covariate partitions as Gaussian observations with bounded signal-to-noise ratios. The authors show the optimal treatment policy is generally not learnable here, but when effects vary smoothly, minimax-adaptive policies based on linear smoothers achieve vanishing welfare regret. Applied to large-scale Netflix experiments, personalized linear-smoothing policies outperform unpersonalized ones.
Benedikt Koch, Winston Chou, Aur\'elien Bibaut et al.
In briefUniSkill: Learning Actor-Aligned Skill Proposals for an Evolving Policy
SummaryThis paper presents UniSkill, a method in which an LLM agent's shared policy both acts in the environment and proposes edits (Add, Update, or No Edit) to its skill bank. Contrastive action feedback measures how swapping in a proposed skill changes the actor's log-likelihood gap between successful and failed trajectories, avoiding extra rollouts per proposal. The authors report 98.4% success on ALFWorld and 84.7% on WebShop, with stable joint training.
In briefBeyond Anonymous Captions: Identity-Grounded Video Captioning and QA
SummaryThe paper presents a framework for identity-aware video captioning and person-centric question answering that identifies characters, grounds them with bounding boxes, and adapts models via LoRA fine-tuning. It builds a benchmark of 750 captioned clips and 3,000 questions from LSMDC v2 movie clips. The authors release the benchmark, training data, code, and BAC checkpoints.
Anas Filali Razzouki, Killian Steunou, Khalil Guetari et al.
In briefCARing: Coverage-Aware Reasoning with Medical Tokens for Predicting Diagnoses
SummaryThis paper studies using large language models to predict a patient's diagnoses at the next visit, noting that rewarding only final-answer correctness fails to encourage covering several valid diagnoses. It proposes CARing, which encodes diagnoses as compact semantic IDs via residual quantization and optimizes multi-label coverage with a coverage reward and multi-positive supervision. On MIMIC-III and MIMIC-IV, the authors report higher weighted F1 than all EHR-trained baselines and the highest top-k recall at every cutoff.
WhyIntroduces coverage reward and semantic IDs for multi-label diagnosis, beating all EHR baselines on MIMIC-III/IV; useful for clinical prediction subfield.
In briefBagDINO: Visual Matching of Checked Bags Across Cameras Using DINOv3
SummaryThis paper addresses finding mishandled checked baggage by identifying bags visually, matching a query image against registered bag images captured by multiple cameras. It places a BNNeck re-identification head on a DINOv3 foundation-model backbone and adapts it efficiently with LoRA. Experiments on the MVB benchmark compare a frozen backbone with LoRA and fine-tuning. The authors report that this parameter-efficient approach is effective and stable with limited training data.
Vita Santa Barletta, Danilo Caivano, Rebecca Margiotta et al.
In briefHeiCo-FOCUS: A Clinical Dataset for Long-Video Understanding
SummaryThis paper introduces HeiCo-FOCUS, a dataset built from Heidelberg colorectal surgery videos to test whether vision-language models can track foreign objects over long procedures. It contains 30,000 visual question answering pairs across five capabilities and was annotated with help from 39 surgical experts. Experiments on ten frontier models show the task is far from solved, with temporal grounding being the hardest, at a mean accuracy of 19.7%.
In briefHySPE: A Position Encoding Method Using Symplectic Dual Shears
SummaryThis paper proposes HySPE (Hyperbolic Symplectic Positional Encoding), which replaces the rotations in Rotary Position Embedding (RoPE) with non-compact symplectic transformations built from damped dual shears. The authors diagonalize the operator in its invariant eigenbasis and use blockwise coordinate rebasing to avoid exponential representation drift, keeping numerical bounds independent of length. In experiments, HySPE keeps lower perplexity than RoPE when extrapolating to longer sequences, reducing tail perplexity by 83.9% on WikiText-103 at length 8192, while the authors note that scaling to large foundation models remains future work.
In briefHow many spectral directions a truncated diffusion sampler must keep, with matching bounds for power-law data
SummaryThis paper asks how many spectral directions a truncated diffusion sampler must retain when data covariance spectra follow a power law. For Gaussian data compared against a smoothed target, the authors prove matching bounds on the required number of retained directions when the ambient dimension is large enough. They show that truncation error depends on the combined Wiener gains of omitted directions, so keeping only directions above the output noise level can leave non-vanishing error.
In briefVisible reasoning isn't a universal fix: effects vary by persona and thinking setup in analytics code generation
SummaryThis paper examines whether visible chain-of-thought reasoning reliably improves analytics code generation in SQL and pandas. The authors build a query-matched SQL-pandas benchmark that varies persona phrasing, target language, visible-CoT format, control prefixes, direct generation, and internal-reasoning settings, evaluating outputs by execution. They find no universal accuracy gain from visible CoT and no consistent benefit from matching the reasoning language to the target, with effects depending on persona, target, model configuration, and internal reasoning.
Bhawani Shankar Leelar, Pawan Chorasiya, Davin Hill et al.
In briefInterView-C: A synchronized multimodal corpus of avatar-mediated VR survey interviews
SummaryInterView-C is a German multimodal corpus of 27 survey interviews conducted entirely in virtual reality, with both participants represented by avatars, covering 54 recordings. It aligns speech with synchronized gaze, head, body, facial, and hand-tracking data, and provides manually post-edited word-timed transcripts plus negation cue and scope annotations. The authors show that several open ASR systems disproportionately misrecognize short answers and number words, and that negation models trained on existing corpora perform lower and more variably on these transcribed interviews than on a model trained on the corpus's own annotations.
Patrick Schrottenbacher, Leon Hammerla, Lydia Kleine et al.
In briefAI Progress and Outlook in the ARPES Experimental Workflow
SummaryThis review surveys how machine learning is used across the full workflow of angle-resolved photoemission spectroscopy (ARPES), from sample preparation and real-time data acquisition to post-experiment analysis and comparison with theory. The authors note that open ARPES databases are small and fragmented compared with large shared datasets. They propose sharing pretrained models and building a larger, standardized open ARPES dataset repository. The paper closes by using a six-level laboratory automation framework to discuss the path toward a fully autonomous, self-driving ARPES lab.
Sandy Adhitia Ekahana, Aalok Tiwari, Pratik Saud et al.
In briefLLM4Impact: A Method That Combines Diverse Evidence to Predict a Paper's Scientific Impact
SummaryThe paper introduces LLM4Impact, a method for predicting the future citation impact of newly published papers by combining semantic, graph, LLM, and temporal information. An adaptive gating mechanism weights different evidence sources, and a calibration module adjusts for domain and publication-time differences in citation scales. The authors also built a benchmark of 2 million papers, and report a 10.13% reduction in year RMSE in-distribution and 6.87% out-of-domain compared with strong baselines.
In briefEstimating Attention Using Black-Box Vector Search
SummaryThis paper studies how to approximate attention in sparse attention mechanisms using only a small subset of keys. Assuming black-box access to a maximum inner product search (MIPS) oracle, it determines how many keys must be retrieved for an ε-accurate estimate. A single index needs Θ(√n/ε) keys, while Θ(log n) indices need only O(log n+1/ε²) keys, which the authors show is near-optimal. In LLM inference, the methods outperform top-k and sampling approaches.
WhyUnifies sparse attention and MIPS retrieval via priority sampling, giving tight key-count bounds and an augmentation-based single-index method validated on LLM inference.
Stepan Zharkov, Krish Singal, Ashwin Padaki et al.
In briefHarnessIR: Using Multimodal Foundation Models for General Real-World Image Restoration
SummaryThis paper proposes HarnessIR, an agentic framework for real-world image restoration that uses a multimodal foundation model as its executor. It runs five stages: perception and diagnosis, on-demand tool invocation, prompt composition, execution, and verification-driven refinement. The authors report state-of-the-art results on the MiO100 synthetic benchmark and better generalization on challenging real-world scenes than prior agentic systems.
In briefWhen Algorithmic Exploration Becomes Cheap: EDA Agentic Research Case Study
SummaryThe authors ran eight trials of agentic algorithm exploration in EDA, involving one faculty member and seven students. With limited intervention, agents built mathematical constructions, analyzed tools, and implemented improvements, though some efforts missed practical goals. They also used AI to analyze 8,420 papers from 2022 to 2026, classifying 97.7% of 2,380 core papers as computationally closed.
In briefAdversarial m-set bandits where only the winner's feedback is observed
SummaryThis paper gives upper and lower bounds on regret for m-set adversarial bandits, across different utilities (winner reward or sum of rewards) and feedback models (winner index, winner reward, sum of rewards, and combinations). Comparing with standard bounds for combinatorial and MNL bandits, it shows that subtle changes in the setting can strongly affect learning rates. The main technical contribution is information-theoretic lower bounds on regret, and experiments on synthetic data support the theory.
In briefDiagnosing multi-agent LLM failures using communication topology
SummaryThis paper studies how the communication topology of multi-agent LLM systems relates to failure types. It introduces MAScope, a two-stage framework in which a Trace Structural Extractor recovers topology from execution traces and a Topology-Conditioned Judge classifies failures. Experiments report a statistically significant association between topology and failure type, and show that topology context improves diagnosis while lowering cost under a fixed orchestration structure.
In briefD-SLR: A Disjoint Row-Sparse plus Low-Rank Matrix Decomposition
SummaryThis paper proposes D-SLR, a closed-form alternative to the truncated SVD for matrix compression, where each row is either stored exactly or approximated by a low-rank fit, never both. The authors state that under squared error this disjoint restriction loses nothing compared with the joint optimum, while using fewer parameters. They also derive an a-posteriori lower bound on the error for every rank and stored-row count. Experiments on synthetic data and real data, including LLM embedding tables, network traffic, and hyperspectral images, are reported to confirm the gains.
In briefLifelong navigation to small objects in rearranged layouts: benchmark and method
SummaryThis paper studies household robots that must repeatedly find small, portable objects such as tools and toys, even when people move them unobserved and the robot starts with empty scene memory. The authors define this as Lifelong Small-object Navigation in Changing Object Layouts (LiSoNav-COL) and introduce LiSoNav-Eval, a benchmark with 28 indoor scenes and 45 small-object categories. They also propose IVAM-Nav, a method that observes supporting surfaces from multiple viewpoints and anchors memory to those viewpoints. Experiments show IVAM-Nav performs favorably against representative methods, and smaller objects, larger environments, and longer relocation distances are harder.
In briefTraining with missed targets in generative recommendation, separating supervision from probability competition
SummaryGenerative recommenders return a limited candidate set and may omit observed targets, so some training strategies append these missed targets to reranker training lists while inference still ranks only original candidates. The paper builds three matched losses to separately test appended-target supervision and probability competition between groups. In four prespecified Amazon Video Games comparisons, removing the competition improved FT-NDCG by 7.8% to 22.2%. The authors conclude that candidate completion should be evaluated for each generator rather than applied automatically.
In briefWhat a Gaussian process design can test: judging before observing whether the chosen inputs can expose a wrong model
SummaryA Gaussian process model may fit data because its assumptions are right, or because the chosen inputs could never reveal that they are wrong. The paper shows that the design alone, before any responses are observed, determines what kind of contradiction it can produce. For GPs built from finitely many features these relations are the null space of the kernel matrix, interpreted geometrically via Gale duality; for other kernels they become soft, prior-based constraints. In simulations, choosing the next input by predicted power raised power against a localised discrepancy from 0.48 to 0.72.
In briefRewardWeaver: Self-evolving reward adaptation to help language agents learn long, multi-step interactions
SummaryThis paper proposes RewardWeaver, a framework that helps language agents learn better in long, multi-step interactions. After each training stage, it traces low-scoring interaction trajectories back to find recurring capability bottlenecks and dynamically selects matching process rewards for the next stage. The authors evaluate it on SOTOPIA, Amazon HistoryPrice, and a newly built Sales Benchmark, reporting new state-of-the-art results. Ablation studies indicate that dynamic reward allocation and failure-grounded attribution contribute to performance.
In briefEdgeRock: Multimodal Reasoning Models with Constant Memory Use
SummaryThis paper presents EdgeRock, a family of multimodal models that keep memory usage constant during reasoning by storing useful information in a fixed-size bank through a method called Rock Memory Anchors. The open-source EdgeRock-4B reaches 82.9% mean pass@1 on AIME 2024-2026, outperforming the bounded-state models evaluated. Compared with its source model, it uses up to 99.1% less KV cache and achieves up to 6.42× higher batch throughput.
In briefA Probabilistic View of Wasserstein-Based Evidential Uncertainty for Out-of-Distribution Segmentation
SummarySemantic segmentation networks only recognize a fixed set of classes, so they struggle with out-of-distribution objects, a serious limit for safety-critical uses such as autonomous driving. This paper replaces the Euclidean objectives of Evidential Deep Learning with Wasserstein-based objectives that respect the geometry of the probability simplex, aiming to preserve uncertainty on unfamiliar inputs in a single deterministic forward pass. The authors evaluate on DeepLabV3+ and SegFormer using the SegmentMeIfYouCan benchmark datasets, finding that the optimal Wasserstein order depends on the architecture. They report that the framework surpasses comparable baselines on most metrics.
Arnold Brosch, Abdelrahman Eldesokey, Michael Felsberg et al.
In briefMechanics of Long-Context Hybrid Models, Part 1.1: From Hybrid Attention to Hybrid Position
SummaryThis paper studies hybrid large language models that combine full attention with either sliding-window attention (SWA) or gated linear attention variants such as GLA and GDN. It reports a Seesaw Effect: linear-attention hybrids gain more from long-context continual pretraining, while SWA hybrids perform better at length extrapolation, which the authors attribute to differing positional inductive biases. They also propose Sliding-Window Linear Attention, which achieves 16× training-free length extrapolation and maintains 100% accuracy on NIAH-SK1 at 64k context length.
In briefCAFE+FNO: a neural operator that builds Fourier kernels by multiplying features
SummaryThis paper proposes CAFE+FNO, which adds the Content-Aware Frequency Encoding+ (CAFE+) module to Fourier Neural Operator kernel generation for solving partial differential equations. CAFE+ combines Fourier and Chebyshev features through parallel affine branches and a Hadamard product, and a kernel MLP maps each spectral coordinate to a complex channel-mixing matrix. The method is compared with existing FNO variants on five PDE benchmarks, with ablation studies on basis configuration, multiplicative composition and bandwidth learnability.
In briefThe Cost of Classical Multi-Agent Path Finding
SummaryThis paper examines how much solution quality the classical discrete-time, graph-based multi-agent path finding (MAPF) formulation gives up compared with continuous-time MAPF (MAPF_R). The authors compare the two across agent counts, map types, and resolutions. Continuous time improves solution quality by at least 5% on narrow and constrained maps and 17% on open maps, sometimes over 20%. Doubling map resolution in classical MAPF recovers less than 3%.
Alvin Combrink, Sabino Francesco Roselli, Martin Fabian
In briefTranslating short user commands into navigator-friendly instructions for frozen drone navigation agents
SummaryThis paper addresses the gap between short, intent-driven user instructions and the detailed commands drone navigation agents are trained on, which sharply lowers success rate on a frozen OpenFly navigator. It introduces TGIT, a front-end translator that keeps the navigator frozen and learns from trajectory outcomes to convert weak inputs into executable commands. The translator raises weak-input success rate to 37.93% and transfers zero-shot to real human instructions, improving from 11.33% to 32.51%.
In briefWhy NLP papers increasingly use 'rather than' contrast phrasing
SummaryThe paper studies the overuse of the 'rather than' construction in scientific writing, comparing ACL papers from 2019, ACL-style arXiv papers from 2026, and GPT-generated papers with the same titles and abstracts. The construction's rate in 2026 is seven times its 2019 rate, and is higher still in GPT papers. The authors conjecture the pattern is a side effect of post-training on pairwise preferences.
Olga Zamaraeva, Adri\'an Gude, Roi Santos-R\'ios et al.
In briefExperienceIndex: A Memory Layer for AI Agents Built on Document Experience
SummaryThis paper introduces ExperienceIndex, an experience layer that lets AI agents capture and reuse knowledge about artifacts such as court cases or scientific papers, drawn from their earlier reasoning. It stores single-artifact experiences and artifact-pair experiences, and works as lightweight middleware that helps agents find the full set of relevant artifacts for new tasks. The authors report answer-quality gains of up to 11.0 points and online cost reductions of up to 50.5% across corpora and search frameworks.
Peter Baile Chen, Geoffrey X. Yu, Xinming Liu et al.
In briefTemporal Residual Bottleneck for Robust Asynchronous Collaborative Perception
SummaryThis paper treats asynchronous collaborative perception, where shared vehicle features arrive late or incomplete, as temporal residual prediction. It keeps a pose-warped collaborator feature as an anchor and uses a Δt-conditioned xLSTM to extract residual temporal evidence, applying only gated, regularized corrections before fusion. Experiments on DAIR-V2X and OPV2V report particular effectiveness under severe fixed or irregular delays and packet drops.
Melih Yazgan, Ahmed Abouelazm, J. Marius Z\"ollner
In briefSkillSandbox: Verifying Agent Skills Through Dynamic Scenario Synthesis
SummaryThe paper addresses self-evolving agents whose distilled skills may encode incorrect or non-transferable procedures, and proposes SkillSandbox to check whether each skill stays useful in new tasks. It synthesizes a skill-relevant but novel task and environment per skill, then compares executions with and without the skill to issue a Keep or Reject verdict. Across ALFWorld and WebShop with three models, the authors report the strongest downstream performance and improved execution efficiency.
In briefMulti-agent coordination via support-preserving teacher-student distillation
SummaryOffline multi-agent reinforcement learning often distills a centralized teacher into decentralized one-step actors using generative policies. The authors find that standard flow-based teachers can route nearby noise samples to conflicting coordination modes, producing invalid samples between modes that propagate to the student. They propose Mode-Support Semi-Discrete Optimal Transport (MoSDOT), which assigns each noise sample to a single mode before teacher training. Experiments on controlled diagnostics and offline MARL benchmarks report improved endpoint quality and routing consistency, especially on multimodal datasets.
In briefActivation-Aware Weight Tensorization: A Calibration Step That Improves Tensor-Network Compression of LLMs
SummaryPost-training tensor-network compression replaces Transformer linear layers with Tensor Train or Tree Tensor Network operators, but standard methods minimize weight-space error rather than error under the layer's activation distribution. The paper proposes AWT, a training-free calibration wrapper that rescales each weight matrix using a diagonal activation-derived scale before the unchanged decomposition solver. Across Llama 3.1 8B, Ministral 8B, and Qwen2.5 7B at 2-6 times compression, AWT consistently improves vanilla TT/TTN, closing 12-60% of the WikiText perplexity gap to the dense baseline, with gains also seen on HellaSwag and ARC-Challenge.
Alessandro Beatini, Marco Maronese, Emanuele Rodol\`a
In briefPrecise statistical theory explaining how well a common machine-learning model's settings can be estimated from data
SummaryThis paper studies Gaussian process models that use the widely used RBF (squared exponential) kernel, and how their parameters are estimated by maximum likelihood. Under fixed-domain asymptotics, it provides the first complete characterization of the joint MLE of the spatial variance, lengthscale, and nugget variance. The authors prove consistency, convergence rates for all three parameters, and joint asymptotic normality, and show these rates are minimax optimal.
WhyGives the first complete fixed-domain asymptotic characterization of joint MLE for RBF-kernel GPs, proving consistency, rates, normality, and minimax optimality; substantial for the subfield.
In briefOutcome-conditional recalibration for better-calibrated forecasts of key events
SummaryThis paper proposes outcome-conditional recalibration, a post-hoc method that makes probabilistic forecasts well calibrated within user-defined regions of the outcome space. It applies the quantile recalibration of Kuleshov et al. (2018) to conditional distributions, then rescales them so forecast event probabilities match observed frequencies. Across regression benchmarks, it improves calibration on particular outcomes over existing methods while staying competitive overall. In day-ahead electricity price forecasting, it substantially improves calibration for negative prices with negligible loss in accuracy.
In briefHGP: An on-device personalized agent memory using hybrid graph storage
SummaryThe paper argues that LLM-based agents struggle in personalized interactive tasks because long-term traces are heterogeneous, multi-typed, and implicitly constrained, and existing memory methods handle routing and retrieval poorly. It proposes HGP, a hybrid graph memory framework with a lightweight self-enhancement classifier for personalized memory routing, which builds episodic, semantic, and procedural memories as graphs and extracts working memory as a state trajectory. On the PAL-Set solution selection task, HGP reaches an S-score of 35.58, nearly 7 points above the strongest baseline.
In briefFrom Pixel to Coding: Testing AI Models on Recreating Figures as Code
SummaryThis paper introduces FigCodeBench, a framework for testing multimodal large language models on reproducing figures as executable code. It includes 6,194 instances spanning 7 functional categories and 4 programming languages, split into three difficulty tiers. Evaluating 24 proprietary and open-source models, the authors report a non-linear performance drop across languages and difficulty levels.
In briefComprehension Audits: A Safeguard Against Risks from Automated AI Research
SummaryThis paper notes that AI already writes most code at frontier AI labs, which creates safety risks when human oversight is insufficient. The authors propose comprehension audits, where responsible people explain their R&D contributions to independent auditors; failure to demonstrate understanding halts that development until remediated, with escalating consequences for repeated failures. Their analysis of leading open-source AI projects finds rising code output alongside falling human review comment rates per line, with automated fleet accounts far lower than human accounts.
Ronald J. Bodkin, Bahrad A. Sokhansanj, Gillian K. Hadfield
In briefLoud vs. Quiet Failures: How Tool-Using Agents Detect and Recover from Faults
SummaryThe authors inject four types of faults into the executable environments of a function-calling benchmark and test whether tool-using language model agents notice them, change plan, or recover the task. Across six models and 1,920 trials, agents flagged a problem in 91.3% of trials when a tool returned an explicit error but only 58.8% when it returned a plausible wrong value. The study concludes that agents respond to the error channel rather than the content of results, so failures that stay within the expected format often go undetected.
In briefTRACK: A Telemetry-Based Tool for Analyzing and Coaching Driving in Sim Racing Games
SummaryThis paper presents TRACK, a framework that analyzes sim racing telemetry to profile individual drivers' behavior. Each session is represented in a four-dimensional space (speed, braking, strategy, consistency) and grouped using unsupervised clustering, with each result checked against a null baseline. Some groupings do not separate from chance, and the authors say so. The sample is small, so the findings are preliminary.
In briefCaching Encoder Outputs for Compact, Reusable LLM Memory
SummaryThe paper introduces EncBank, which treats the lower layers of a pretrained LLM as a reusable document encoder and stores their outputs compactly for an adapted upper-layer reader. Across five benchmark suites on three Qwen models, 4-bit storage stays within one point of native-precision EncBank on each aggregate score. In a fixed Qwen3-8B workload it keeps 28.1% of the native-precision persistent GPU store. Results depend on workload, preparation cost, and reuse frequency.
In briefWxFM-XL: Adapting Univariate Time Series Models to Multi-Station Weather Forecasting
SummaryThis paper addresses extending univariate time series foundation models to forecast weather across multiple stations, a setting where existing multivariate extensions overlook station spatial information and station-specific error priors relative to the foundation model. The authors propose WxFM-XL, which introduces a cross-station error correlation prior graph to capture each station's error prior and a dynamic fusion mechanism that adaptively combines it with a spatial correlation graph. According to the paper, experiments on multiple datasets show the model outperforms state-of-the-art baselines.
In briefTransition path sampling via Koopman operators and exit-time optimal control
SummaryThis paper proposes a new method for sampling rare transitions between metastable states in dynamical systems such as molecular dynamics. It uses the leading eigenfunctions of a Koopman operator to identify metastable sets and estimate the committor function without transition path data. Transition path sampling is cast as an optimal control problem up to the first hitting time of the target set, solved in a reproducing kernel Hilbert space via a linear KKT system. On the double well and alanine dipeptide, the fraction of trajectories reaching the target rises from 0% to 99.8% within 1000 steps and to 93% within 1 ps.
In briefHas LLM Screening Performance Stalled in Software Engineering Reviews?
SummaryThis paper asks whether large language models have improved at screening studies for systematic reviews in software engineering. The authors tested eight newer LLMs against seven older ones on an existing benchmark and a smaller sampled dataset, also comparing prompts, model agreement, and refined inclusion criteria. Newer models performed only marginally better, with average MCC rising from 0.347 to 0.365. The authors conclude LLMs are not yet ready to replace human screeners, and the benefits of costlier new models appear very limited.
Aleksi Huotala, Miikka Kuutila, and Mika M\"antyl\"a
In briefWhether extra reasoning time reduces cognitive bias in large language models
SummaryThis study tests whether reasoning models become less biased when given larger thinking budgets, across six classic decision biases. The authors ran about 12,350 API calls over four model families, using the reasoning tokens actually consumed as the measure of deliberation. Reasoning models were not reliably less biased than matched non-reasoning siblings, and more realized deliberation did not reliably reduce bias. Only anchoring appeared in the human direction.
In briefAdSpark: A dataset and benchmark for generating product ads from images and prompts.
SummaryThis paper introduces AdSpark, a large-scale dataset and benchmark for generating product-centric advertisement videos, built from data on a major e-commerce platform. The AdSpark-300K dataset contains about 300K image-prompt-video triplets with annotations such as product identity, selling points, creative plans, and audio scripts. The AdSpark-Bench evaluates generated ads across six dimensions, revealing challenges in product preservation, multi-shot storytelling, and selling-point visualization.
In briefCANDO: A cooperative multi-agent network for optimizing layout design
SummaryThe paper addresses layout generation for real-world facilities, which involves irregular boundaries and access and motion-planning constraints. It introduces ALPS-Bench, a benchmark of 1,000 professionally annotated facility layouts, with an instance-specific scoring method based on a structured design manual. It also proposes CANDO, a training-free multi-agent framework that iteratively refines layouts through a verification-based loop, and reports that it outperforms trained and LLM-based baselines on PubLayNet, RICO, and PKU-PosterLayout.
Athanasios Masouris, Zheng Jing, Benjamin Sam Chandler et al.
In briefLearning to Accumulate Knowledge with Mutual Information
SummaryThe paper introduces Knowledge Weaver, a reinforcement learning framework that trains a language model to curate reusable knowledge from agent trajectories. It combines token-wise mutual information feedback with marginal success rewards so that new entries add information beyond what the knowledge bank already holds. On ALFWorld and WebShop with k=10 retrieved entries, it reaches mean success rates of 54.0% and 42.0%, exceeding GRPO by 16.9 and 18.7 percentage points.
In briefEvolving Models on a Host, Predicting Stock Returns on an Edge Device
SummaryThis paper applies ONE-NAS, an online neuroevolutionary architecture search, to daily cross-sectional stock return prediction for US mid-cap equities. A host runs the search and sends champion genomes over TCP/IP to a Raspberry Pi 4B, which makes predictions online. On four panels from 2022 to 2024, an ensemble of island champions returned +27.5% net of transaction costs, above online LSTM, GRU and monthly-retrained LSTM baselines and the single best genome from prior ONE-NAS work at +4.5%.
In briefOpenESM: Energy-Steered Models as Scalable Foundation Learners
SummaryThis paper proposes energy-steered models (ESMs), a language modeling architecture that first refines a latent state over an energy landscape and then uses it to generate tokens. The authors pre-train ESMs on three large-scale corpora and evaluate them on 26 downstream tasks. They report promising scaling with model size and training data, and better results than GPT models on most tasks. The model weights and training infrastructure are publicly released.
In briefGaussian equivalence for multi-head self-attention, a random-matrix result on how attention outputs behave
SummaryThis paper uses random matrix theory to show Gaussian equivalence for multi-head self-attention. Replacing softmax attention with rescaled scores plus Gaussian noise preserves the limiting spectral law of the centered output. The equivalence also covers key-dependent value and output projections. The results separate the effects of head allocation and projection widths, and distinguish spectrum-preserving across-head sharing from within-head key-value dependence.
WhyExtends Gaussian equivalence to multi-head self-attention with limiting spectral laws separating head allocation and projection width effects.
In briefStudy showing fruit fly brain wiring alone supports efficient vision
SummaryResearchers linked the proofread adult fruit fly connectome to an anatomically faithful eye model, learning only synaptic gains and neuronal thresholds. The model performs multiple visual tasks, including color discrimination, shape classification, and numerical discrimination. At matched wiring cost, the biological network achieves higher accuracy than randomized rewired versions.
Eudald Correig-Fraga, Roger Guimer\`a, Marta Sales-Pardo
In briefControlling Dependence in Generative Models with Spread Mutual Information
SummaryThis paper proposes Spread Mutual Information (SMI), a weighted integral of mutual information across noise levels, obtained by applying a common spreading kernel to the generated variable. Gaussian spreading makes the conditional and marginal densities smooth and strictly positive, extending gradient estimation to distributions that may originally be singular. Across various experiments, SMI consistently achieves effective dependence control among mutual-information-based methods and remains competitive with established task-specific approaches.
Jiahao Yu, Song Liu, Jos\'{e} Miguel Hern\'{a}ndez-Lobato et al.
In briefWave-Based Graph Learning over Sheaves for Long-Range Communication
SummaryThe paper observes that simply increasing propagation depth in graph neural networks does not guarantee effective influence between distant nodes. It proposes ONDA, which evolves stalk-valued representations through second-order dynamics driven by learned sheaf transport operators. Experiments on long-range propagation, severe bottlenecks, graph transfer, and heterophilic benchmarks report gains over scalar wave propagation, diffusive sheaf baselines, and state-of-the-art models.
Jan-Willem Van Looy, Alessandro Trenta, Alessio Gravina et al.
In briefPhase-HDC: a training method that replaces optimizer history with gradient thresholds for compact discrete-phase models
SummaryThe paper asks whether a compact hyperdimensional classifier with low-bit angle parameters can be trained while storing only the model, without optimizer state. The proposed Phase-HDC method moves each angle by at most one step, against the gradient sign, only when the gradient is large enough. Across eleven datasets it stores 16 to 23 times less than float32 Adam and 4 to 6 times less than 8-bit Adam, with about five points lower accuracy than float32 Adam and higher accuracy than 8-bit Adam on six datasets. The storage savings are described as logical state rather than measured hardware memory.
In briefWhat the Sleeve Feels: Explainable ML for Pressure-Based Posture Screening
SummaryThis paper studies a knitted piezoresistive sleeve worn on the forearm that uses pressure-distribution signals to screen posture. Everyday activities are grouped into three screening categories, and 29 interpretable pressure features are engineered. Under a strict subject-wise split, a tuned XGBoost classifier reaches 0.818 accuracy and 0.801 macro F1 on unseen test subjects. The authors also use SHAP explanations, feature-group ablation, and a simulated donning-rotation test to examine what the model relies on and where it degrades.
In briefCan a fruit fly's whole brain wiring learn human-designed tasks?
SummaryThe paper uses the publicly released adult male Drosophila connectome (MaleCNS v1.0) as the fixed recurrent topology of an artificial network, learning one scalar weight per anatomical edge. Separate models are trained on bounded addition and on a grounded relational language task built from a 100-word lexicon. The anatomical graph reaches 92.77% mean accuracy on held-out addition, versus 67.93% for degree-preserving rewires, and 61.59% on the language task, versus 44.17% for matched rewires.
In briefScalable Patch-Level Self-Supervised Learning: A Principled, Stable Method for Visual Representations
SummaryThis paper introduces JEM, a self-supervised visual learning method derived from an information-theoretic objective built on the multi-view assumption. It learns by aligning corresponding patch representations across views, with information and structure preservation losses as regularizers. JEM trains stably from 300M to 7B parameters and consistently surpasses DINOv2 on segmentation benchmarks. At 7B parameters, it exceeds DINOv3 on panoptic segmentation while using 12 times less training data and no refinement stages.
WhyDerives a principled patch-level SSL objective, first shown at 7B scale, beating DINOv2 and DINOv3 on segmentation with substantial experiments.
Maximilian Seitzer, Gabriele Trivigno, Anton\'in Vobeck\'y et al.
In briefKruskal-based algorithms for flat and hierarchical watershed image segmentation
SummaryThis paper studies watershed segmentation on edge-weighted graphs and links it to optimization problems such as the Minimum Spanning Tree, enabling efficient algorithms for computing flat and hierarchical watershed segmentations. It reviews related literature and presents an end-to-end pipeline, from graph-based image representations to connected components of the final segmentation, covering supervised and unsupervised variants and different seed computation methods. The authors aim to offer a compact reference for readers who want to use or reimplement the watershed framework.<br>
Jean Cousty (LIGM), Laurent Najman (KUSTAR, LIGM) et al.
In briefASTRA: a perception-aligned benchmark and scorer for evaluating style transfer
SummaryStyle transfer lacks a reliable evaluation standard, and existing automatic metrics often disagree with human preference. The paper introduces ASTRA, comprising ASTRA-Data, a benchmark with generated results and two-stage pairwise human judgments, and ASTRA-Score, a learned evaluator that predicts preference-aligned scores from content, style, and stylized image triplets. Experiments show ASTRA-Score correlates substantially better with human rankings than prior metrics.
Yang Deng, Eleftherios Ioannou, David Mould et al.
In briefAttacks on bandits that boost a target arm, not just suppress rivals
SummaryThis paper studies offline attacks on warm-start bandits, where an attacker can inject only valid action-reward pairs into the history before deployment. It shows that when the target arm's reward is near the lower bound, cost-optimal attacks against UCB must spend a nonvanishing share of cost promoting the target. The authors design an attack reaching optimal sublinear cost and extend it to Thompson Sampling, ε-greedy and other algorithms. Experiments on real and synthetic data validate the attacks.
In briefTemporal predictive multiplicity: equally accurate forecasting models produce different trajectories
SummaryThis paper studies predictive multiplicity in time-series forecasting, where models with near-identical performance give substantially different predictions. The authors introduce a framework called temporal predictive multiplicity that measures disagreement over complete forecast trajectories. Experiments with 19 neural forecasting architectures on 11 datasets show that near-optimal models can produce highly variable trajectories, and that trajectory-level disagreement is largely unrelated to horizon-wise disagreement.
Emanuele Albini, Francesca Toni, Saumitra Mishra et al.
In briefHybrid patch-based and diffusion anomaly detection for wafer bin map open-set detection
SummaryThis paper proposes a hybrid one-class framework for detecting anomalies in semiconductor wafer bin maps, combining a patch-based student-teacher detector (EfficientAD) with a denoising diffusion model (DDPM). Their percentile-calibrated scores are fused through a fixed convex combination. Trained on only 700 normal wafers and tested on 18,658 held-out wafers, the fused detector achieves an AUROC of 0.9985 and reduces misclassifications to 618. The authors also note that fixed-weight fusion recovers only 40 to 70 percent of the available correction.
In briefJointly optimizing prices and LLM-generated ads for revenue
SummaryThe paper studies a sequential problem where a seller jointly sets a price and an LLM-generated advertisement, observing only whether each offer results in a purchase. The authors propose an online actor-critic algorithm that combines LoRA-adapted pretrained LLM with a fitted demand model. Evaluated on three synthetic demand models and a simulator built from real marketplace data, it reports expected revenue gains over a reference policy of 5.69%, 5.18%, 55.96%, and 5.81%.
Alessandro Barro, Francesco Bacchiocchi, Francesco Emanuele Stradi et al.
In briefExtreme Binary Classification: Using Extreme Value Theory to Tightly Limit False Negatives
SummaryThis paper introduces Extreme Binary Classification, which aims to learn a classifier whose false negative rate shrinks toward zero as the number of positive training examples grows. The authors propose a threshold adaptation method grounded in Extreme Value Theory, along with a feature selection procedure based on a permutation test on sample maxima. Experiments on four real-world datasets of varying sizes show favorable comparison with state-of-the-art methods, and an application to a cancer screening dataset illustrates its interpretability.
In briefEASE: Entropy-based adjustment to make AI text evade detectors
SummaryThis paper shows that detectors of AI-generated text are sensitive to how the source language model decodes its output, since perturbing next-token logits or changing sampling temperature lowers detection performance. The authors propose EASE, a training-free, detector-agnostic method that uses the source model's predictive entropy to adapt logit perturbation and sampling temperature, without detector feedback or fine-tuning. Experiments with three source LLMs and several detectors report consistent drops in detection performance, with negligible loss in text quality and negligible inference overhead.
In briefProbabilistic Reformulation Linking Expected Harmfulness to Likelihood for Jailbreaking LLM Agents
SummaryThis paper studies jailbreaking LLM agents when output harmfulness can be quantified, where a natural objective is maximizing expected harmfulness over admissible input changes. The authors show through a probabilistic reformulation that this objective is precisely connected to methods that construct harmful target outputs and increase their likelihood. Building on this, they propose OPUR, a sampling distribution for generating highly harmful target outputs, and use those samples to guide likelihood-based input optimization. Experiments demonstrate the method's effectiveness in jailbreaking LLM agents.
In briefTR-PTQ: Integer-only Transformer quantization via Taylor-region reformulation
SummaryThis paper proposes TR-PTQ, a post-training quantization method for Transformers that uses shared Taylor Region exponential and logarithm primitives so that division and square roots run in the log domain with integer arithmetic. The authors find that accuracy loss mainly comes from LayerNorm scale parameters and compounded GELU approximation errors, while SoftMax is relatively robust. The approach removes the need for floating-point units for nonlinearities and reports less than 1.5% absolute accuracy degradation on vision and language benchmarks.
In briefSuccessive Training Stages and LLM Persuasion: Effects of Misalignment, Fine-Tuning, and Preference Optimization
SummaryThis study tested how three successive training stages affect how persuasive a large language model is, using 835 participants reading texts on 10 political issues. Persuasive supervised fine-tuning produced more attitude change than conspiracy training alone (d = 0.30). Adding identity preference optimization gave no further benefit, and GPT-4 did not differ from neutral text.
Antony Dalmiere (LAAS-TRUST), Pascal Marchand, Guillaume Auriol (LAAS-TRUST et al.
In briefConfining LLM agent self-evolution to the runtime harness in credit pipelines, with a measured admission gate
SummaryThe paper argues that self-evolving LLM agents are reviewable only when changes are confined to the runtime harness while model weights stay fixed, so each adaptation is a diff with a cause and test. It presents a dual-loop engine with an admission gate that writes a hash-chained record before deployment. The gate is measured in simulation using a simulated agent and seeded-search proposer rather than language models. Without the gate, the same loops admitted 309 harmful changes, while the gated loop admitted 144 candidate changes, none of which worsened held-out error.
In briefForce without transmission: a depth-induced collapse no added loss can fix
SummaryThe paper studies rank collapse in small transformers, where all token representations point in one direction and learning stalls. The authors induce collapse by weakening the skip connection and find that adding a corrective loss term does not repair it, because the task gradient no longer reaches the query and key weights. Restoring the skip connection reopens this path and the rank recovers, but the loss stays above that of a healthy network.
In briefPossibilistic radial transport for approximate inferential model inference
SummaryThe paper proposes a possibilistic radial transport that encodes a parameter's contour value in the radius of its source point, so that sampling a confidence cut reduces to truncating the radius. A deep learning algorithm learns the contour while enforcing a depth condition and maximizing within-shell entropy. In simulations, the learned contours match or improve on ellipsoidal approximations, with coverage and power tracking the exact reference. The method is also applied to synthetic AMH data to probe hypotheses about ovarian aging.
In briefMSU Team's Explainable Deepfake Detection Challenge 2026 Entry: Detection Grounded in Visible Artifact Evidence
SummaryThis paper presents a modular system for detecting deepfake images and explaining its decisions using visible forensic cues. The detector combines several DINOv3 models with Mesorch manipulation-localization features, while weak patch-level supervision is derived from Grounding-DINO pseudo-masks. Explanations are generated by Qwen3-VL models and simplified with a GRPO-optimized text model. On the full challenge test split, the method reports 0.9349 detection accuracy, 0.5571 explanation score, and a 0.7456 final score.
Artem Filippov, Aleksandr Gushchin, Kirill Koltsov et al.
In briefLearning Traffic Flow with Stochastic Physics-Informed Neural Cellular Automata
SummaryThis paper proposes a physics-informed neural cellular automaton (PI-NCA) that learns local transition rules for traffic flow from data. Its architecture matches the road topology and guarantees conservation of the total vehicle count, constraining learned rules to physically admissible dynamics. The authors extend it to probabilistic transition rules while keeping the same constraints. Tested on traffic scenarios generated by the Nagel-Schreckenberg and Kerner-Klenov-Wolf cellular automata, PI-NCA outperforms a standard NCA, and the stochastic version preserves the physical constraints.
In briefAgentTime: Testing whether AI agents can control and estimate their own runtime
SummaryThis paper introduces AgentTime, a benchmark of 222 tasks from 18 sources that tests whether AI agents can work for a requested duration, predict runtime, and estimate elapsed time. Results show large differences between agents in how far their runtimes deviate from requested durations, and some agents slept after apparently finishing their work. The authors conclude that completing a task does not guarantee control over time, so both task completion and time control should be evaluated for reliable long-horizon autonomy.
In briefMany Ways to Succeed: Diversity-Driven RL Fine-Tuning for Better Robot Generalization
SummaryThis paper examines how reinforcement learning fine-tuning changes exploration in vision-language-action robot policies, finding that successful trajectories become more diverse and cover more feasible solutions than supervised fine-tuning. Based on this, the authors propose DRIVE, which turns the diversity of successful behaviors into an explicit RL intrinsic reward. The method is evaluated on LIBERO-Plus, ManiSkill3, RoboTwin 2.0, and a dual-arm AgileX PiPER-X platform, where it improves out-of-domain average performance over vanilla RL fine-tuning.
In briefPurifying Backdoored Vision-Language Models by Removing Hijacked Directions
SummaryLarge vision-language models are vulnerable to backdoor attacks, and existing defenses are costly because they require retraining on clean data or per-query intervention. This paper proposes OrthoPurify, which cleans model weights with a single orthogonal projection that removes a small set of weight update directions hijacked to encode the backdoor. These directions are identified using a pseudo-benign model fine-tuned on only a few clean samples. The authors report that the attack success rate drops to near zero while original performance is preserved, without retraining or added inference overhead.
In briefJuno: Adapting predictive latents for vision-language-action models
SummaryThis paper introduces Juno, a framework built on an action-conditioned joint-embedding predictive architecture (JEPA) that supplies predictive latents to vision-language-action (VLA) models. It covers pretraining, policy learning, and deployment, and at deployment it adapts the world model using observed transitions, including failed rollouts, then re-aligns the policy with LoRA adapters and a trainable action head. On SimplerEnv, average success rises from 60.9% to 68.5% over the strongest baseline, and to 72.7% with test-time adaptation. On a real robot, it keeps 70%–75% success under background, height, and object shifts, where the base policy drops to 0%.
In briefFast holographic inversion of superconducting domes: a faster training method that recovers a scalar-mass-dependent mass function from superconducting phase diagrams
SummaryThe paper proposes a new training method for a holographic superconductor whose scalar mass depends on the gauge field strength, cutting the inversion time from days to about ten minutes to an hour. The critical-temperature condition and its gradient are obtained from two integrations without finite differences. The authors find that a dome constrains the mass function only where its horizons reach, so the scalar mass requires a second observable.
In briefCan a Decision Model Diagnose HVAC Faults? Reasoning, Input Representation, and Robustness
SummaryThis study tests whether a pretrained decision model, which outputs probabilities over allowed answers, can diagnose heating, ventilation, and air-conditioning faults. Using 128 fault days from four public datasets, it grades faults by the reasoning they require and compares raw data, physical features, and Brick topology as inputs. Given physical features, the model diagnosed faults supported by a single feature but not those needing operating context. Under shifts in season, control configuration, or building, it kept its accuracy and calibration, while a supervised model lost 0.33 macro-F1.
In briefItgan's parameter-efficient Whisper systems for three Arabic speech recognition tasks at NADI 2026
SummaryThis paper describes the Itgan team's systems for three speech recognition subtasks of NADI 2026: robust country-level ASR, mixed-dialect ASR, and Tunisian code-switched ASR. All systems adapt Whisper with LoRA on consumer GPUs, each adding a different component. The country-level system reached 57.1% country-average WER, the mixed-dialect system 46.7% WER, and the Tunisian code-switched system placed second at 14.49% WER.
In briefExpected sample complexity of decisions in multi-armed bandit problems
SummaryThis paper studies sample complexity in stochastic multi-armed bandit problems, measured as the number of suboptimal decisions an agent makes while interacting with an environment. It introduces an expected sample complexity measure analyzed in an "approximately correct in expectation" (ACE) framework. The authors show ACE guarantees imply almost sure convergence to the optimal expected reward and can be converted into explicit expected regret bounds. They also show deterministic algorithms cannot achieve favorable ACE bounds and analyze stochastic algorithms in settings where the suboptimality level is known and unknown.
In briefDo Generative Priors Align with Human Naturalness Perception?
SummaryThis paper examines whether the native priors of visual generative models reflect the regularities behind human perception of image naturalness. The authors measure native prediction errors across 25 open image and video generators, using content-preserving paired interventions that disrupt facial configurations or lighting consistency. The loss differences reproduce human-like sensitivities, including the Thatcher effect on faces, and track human naturalness judgments on individual stimulus pairs.
