In briefA baseline study testing which AI agent tools work best together
SummaryThis paper proposes MASTIF (Multi-Agent System TestIng Framework), a system for evaluating Agentic Web technologies. The authors compared 3 LLMs, 3 protocols, and 4 agent implementation frameworks on 300 Mind2Web benchmark tasks. They report that choosing a task-specific technology stack may improve task completion rate, efficiency score, and social value stability, while reducing verbosity and task completion time.
In briefFilling Gaps in AI Test Data Coverage with Generated Images
SummaryThis paper addresses the problem that AI test datasets may underrepresent important conditions such as day versus night, which can make evaluation results misleading. The authors propose a three-stage pipeline that uses vision-language models to label semantic factors and find influential gaps, diffusion models to transform existing images into the missing domain, and statistical checks to confirm improved coverage. The synthetic images are used only to enrich the test set, without retraining models. Validation on image classification and object detection under day/night gaps shows the augmented test sets better distinguish models that have and have not seen the missing semantic.
Connor Hashemi, Bharadwaj Ravichandran, Connor Anderson et al.
In briefVerifiable legal retrieval using a structure-driven knowledge graph
SummaryThe paper argues that automated systems in reliability-critical fields such as law and medicine must be transparent and verifiable, which black-box LLM systems often fail to achieve. The authors focus on the German Legal Code, building a knowledge graph through structure-driven parsing to navigate and retrieve source documents. They compare several knowledge-graph-informed chunking and retrieval strategies and evaluate how each affects the accuracy of retrieved citations.
In briefTraining Accurate and Well-Calibrated Neural Networks in a Single Stage
SummaryNeural networks are often poorly calibrated, meaning their confidence scores do not match how often they are correct. Standard fixes adjust a trained model afterward using a separate calibration dataset. This paper trains such calibrators end-to-end with the main network using a bilevel Stackelberg formulation, with only one extra hyperparameter, the calibrator learning rate. The authors report that using the final classification layer as the calibrator yields accurate, well-calibrated models without additional parameters.
Nektarios Kalampalikis, Pablo M. Olmos, Isabel Valera
In briefVariance-reduced policy gradient methods are nearly minimax-optimal
SummaryStandard policy gradient methods suffer from high estimation variance because they rely on Monte Carlo updates. The paper proposes Loopless Variance-Reduced PG (LVR-PG), a unified framework covering first-order estimators that combine recursive updates with defensive importance sampling, including PAGE-PG and STORM-PG. The authors prove last-iterate global convergence with Õ(ε^-2) sample complexity under weak gradient domination, using only first-order updates. They also extend known lower bounds to continuous parametric policies, showing an Ω(ε^-2) limit, so the framework is minimax-optimal up to logarithmic and horizon factors.
WhyUnified loopless variance-reduced policy gradient framework matches lower bounds up to log factors with strong theory, though mainly theoretical with limited application impact.
In briefTransformer-based in-context learning for diffusion model score operators
SummaryDiffusion models usually learn only one fixed target distribution, so they struggle to adapt to new distributions from a few examples. This paper proposes an in-context operator learning framework in which each target distribution is specified at inference time by a set of context samples, and a Transformer learns a score operator that enables sampling on unseen tasks without retraining. The authors evaluate the method on Navier-Stokes, ERA5 climate data, QuickDraw sketches, and Food-101. Results show good performance on relatively simple task families, with performance degrading as task complexity increases.
In briefVQIL: Visible Quantum Interference Learning for Ranking Perturbation Responses on Graphs
SummaryThe paper proposes VQIL, a trainable Szegedy quantum walk that ranks all graph nodes by response strength to a perturbation at a given seed node. Edge weights, phases, and node rotations are learned end-to-end while the walk operator stays unitary. On AC power-flow perturbations across six transmission grids and gene perturbations across five protein-protein interaction networks, the reported Spearman rank correlation rises from 0.147 to 0.831 and from 0.171 to 0.620, respectively.
Seokjun Lee, Sergio A. Ortega, Daniel K. Park et al.
In briefX-AVDD: Cross-Attention Method for Distilling Audio-Visual Datasets
SummaryThis paper proposes X-AVDD, a method that synthesizes a small dataset from a large paired audio-visual dataset while largely preserving neural network training performance. It uses cross-attention to explicitly capture relationships between audio and visual modalities, reducing memory use and end-to-end synthesis cost by about 8 times in GPU-hours. The authors report that X-AVDD significantly outperforms other audio-visual distillation methods on datasets such as VGGSound.
In briefPhysForge: Material-controllable 3D reconstruction with separable components
SummaryPhysForge is a video diffusion model-based framework for creating 3D assets with simulatable physical behavior and controllable materials. Its PhysFit module uses material-conditioned video diffusion dynamics as references for fitting physical properties, with windowed fitting and Hallucinated Forcing to separate material responses from non-physical motion. Its PhysComp module decomposes complex interactions into individual objects and semantic components, optimizing their physical parameters independently before joint simulation. Experiments on PhysAssets and Google Scanned Objects reportedly show the best CLIP and physical controllability scores compared with existing methods.
In briefRetractions Revisited: Replicating and Extending Evidence That Fraudulent Science Uses More Obfuscated Language
SummaryThis study replicates and extends prior work by comparing 800 fraudulent and 800 genuine control papers from 2013 to 2025. The results support the linguistic obfuscation hypothesis, showing fraudulent papers are more obfuscated than genuine ones. Among three competing mechanisms, reality monitoring, measured by decreased adjective rates, best explained the link between fraud and obfuscation.
In briefHow calendar features are encoded in time series foundation models with covariates
SummaryTime series foundation models increasingly accept covariates, and calendar information is the most common one. The study compares seven calendar encodings across four models and nine datasets in a zero-shot rolling evaluation. It finds that encoding choice materially changes forecast accuracy, with the largest gains around holidays.
Alexander Kreusel, Maximilian Beichter, Alexandra Nikoltchovska et al.
In briefCan AI decision models classify breast ultrasound images? A benchmark of CLEF and CLEF-Flash across prompts and datasets
SummaryThis study tests two multimodal decision models, CLEF and CLEF-Flash, on 2818 breast ultrasound images from three datasets, using three different prompts. CLEF-Flash with the clinical prompt achieved the highest accuracy on all three datasets, but on BUS-BRA every configuration fell below a majority-class predictor. The authors find performance depends on both dataset and prompt, and they recommend class-specific evaluation before applying these models to breast ultrasound analysis.
In briefOutlier-aware attention linearization that groups dimensions to approximate Transformer attention efficiently
SummaryThis paper proposes Outlier-Aware Attention Linearization (OAL), a method for approximating softmax attention in Transformers to enable efficient inference. The authors find that a few outlier dimensions contribute disproportionately to query-key dot products and account for much of the approximation error. OAL builds on a second-order Taylor approximation, grouping dimensions with similar contribution scales to share coefficients while treating outliers separately. On Qwen2.5-0.5B, it reduces frozen-attention output relative MSE by about 8.9% compared with DiJiang, and it lowers WikiText-103 perplexity after LoRA fine-tuning.
In briefInjecting a New User Turn to Steer LLM Reasoning Toward Self-Correction
SummaryThe paper proposes UPTURN, which uses a 130K-parameter MLP probe on a frozen encoder, trained only on final-answer correctness, to pick where to intervene in failed reasoning. Under a fixed generation budget, presenting a correction instruction as a new user turn redirects the model more reliably than inline injection. The method requires no LLM fine-tuning and improves average accuracy across four reasoners and six math benchmarks over a 32-rollout baseline.
In briefProgressing Moderately Harder in GRPO Curricula: Benefits and Limits
SummaryThis paper studies curriculum learning for GRPO by re-estimating sample difficulty under the current policy and staging a shift toward harder samples. On GSM8K, progressing to a moderately harder difficulty band raised test accuracy to 0.6591 ± 0.0077, above 0.6449 ± 0.0024 for fixed medium-difficulty selection. Progressing further to a harder band reduced accuracy to 0.6405 ± 0.0053, so the benefit is non-monotonic.
In briefCooperative aerial transport of a payload using a sliding cable through a hook
SummaryThis paper proposes a system where two quadrotors carry a payload using one inextensible cable that slides through a frictionless hook mounted on the payload. The sliding cable couples drone motion, cable geometry, and payload dynamics. The authors develop a feasible reference-generation framework that satisfies cable-length and elliptical workspace constraints while keeping cable tension positive. The approach is validated through experimental demonstrations.
In briefFrozen site-probe accuracy alone does not prove MRI harmonization removed site effects.
SummaryThe paper audits a common evaluation practice for multi-site MRI harmonization, in which a site classifier trained on raw scans is applied unchanged to harmonized scans and its accuracy drop is read as removal of site effects. Using ABIDE I with ten output sets from seven methods, the authors show the verdict depends on how the probe is trained, and a Gaussian blur alone can lower the frozen probe's accuracy to chance. Where the frozen probe fails, probes trained on harmonized scans still recover the site. They recommend reporting several converged probes, probes trained on harmonized scans, and target alignment.
Muhammad Muqsit Islam, Gloria Garc\'ia Cuenco, Jos\'e M. Mart\'inez et al.
In briefAn AI multi-agent system that automatically tunes parts of specialized neural networks
SummaryThis paper proposes a multi-agent large language model framework with analyst, coder, and recorder roles that automatically tunes key components of neural networks without changing their overall structure. It is validated on spiking neural network classification and Physics-Informed Neural Networks for solving differential equations. Across five tasks, including MNIST, Digits, GNLSE, and Kuramoto–Sivashinsky equations, the agent reports better accuracy and lower errors than baselines.
Ivan Moloshnikov, Dmitry Kunitsyn, Alexander Sboev et al.
In briefPretraining and Test-Time Training Budgets Affect ARC Results: A Controlled Study of VARC and TRM
SummaryThis paper compares two ARC models, VARC and TRM, on ARC-AGI-1 under a common protocol, controlling pretraining and test-time training (TTT) budgets. Increasing TTT sample presentations improved the pass@2 metric for both models in all settings. Increasing task-variant count consistently helped only TRM, while VARC often did better with fewer variants. VARC stayed competitive without pretraining, whereas TRM benefited much more from it.
In briefScalable variational inference for non-Gaussian spatial field models
SummaryThe paper addresses the difficulty of computing the posterior of a latent spatial field with a Gaussian process prior when the likelihood is non-Gaussian, a problem for which existing methods are either slow or overly restrictive on large datasets. The authors propose a variational inference framework whose variational family mirrors the sparse nearest-neighbor autoregressive structure of the prior, uses normalizing-flow conditionals, and amortizes all variational parameters with one shared neural conditioner. They report that doubly stochastic optimization with reduced ancestor sets achieves polylogarithmic complexity per iteration.
In briefTabReliBench: Assessing Reliability of Tabular Foundation Models Beyond Accuracy
SummaryThis paper examines whether tabular foundation models (TFMs), which often provide posterior predictive distributions, are reliable and not just accurate. It evaluates point prediction, distributional quality, calibration, and interval efficiency, and introduces TabReliBench, a large classification and regression benchmark. The authors find that TFM rankings shift across benchmarks and reliability metrics, and that near-nominal marginal coverage does not imply valid conditional coverage. They conclude that selecting reliable TFMs requires multidimensional evaluation rather than aggregate rankings alone.
WhyProposes a multidimensional reliability benchmark with large-scale experiments revealing ranking shifts and subgroup calibration gaps beyond prior evaluations.
In briefA critical look at sampling from a sharpened model distribution to improve LLM reasoning without rewards
SummaryThe paper examines recent methods that seek reinforcement learning-like gains in LLM reasoning without reward signals, by sampling from a power distribution of the base model. The authors find these samplers do not converge to the target distribution under realistic compute budgets, and that their gains are largely reproduced by Best-of-N with appropriate grading. Fine-tuning the base model to match the power distribution yields noisy gains specific to certain models and tasks. They conclude that the value of power sampling lies in extracting sequences the base model already considers likely, not in faithfully targeting the power distribution.
Abiel Malkani Talwar, Christopher Nemeth, Henry Moss
In briefPredicting Molecular Properties Using Electron Density and Surrogate Functionals
SummaryThis paper extends a surrogate functional framework, originally built for orbital-free density functional theory, to predict molecular properties. The model takes the ground-state electron density as input and enforces it as the minimum of the surrogate functional. The authors report that this surpasses state-of-the-art performance on six of seven extensive properties in the QM9 dataset and generalizes better to larger molecules from QM40.
Manuel V. Klockow, Simon Wagner, Roman Remme et al.
In briefSelectArena: A Modern Benchmark for Choosing Features in Real-World Tabular Data
SummaryThis paper introduces SelectArena, a benchmark for feature selection, the task of choosing a subset of columns from tabular data. It compares 15 widely used methods on 71 manually curated real-world datasets for classification and regression, using standardized protocols and state-of-the-art tabular foundation models. The results show trade-offs: tree-based methods (CART, RFImportance) and F-Test are Pareto-optimal for performance and runtime, while F-Test, information-theory-based and distance-based methods lead on stability and validity. Search-based methods were slow, unstable, and performed poorly.
Bastian Sch\"afer, Dominika Matus, Lennart Purucker et al.
In briefLearning size-independent constraint models from data and example solutions
SummaryThis paper presents a Constraint Acquisition Tool that builds size-independent constraint programs from input data and positive and negative example solutions across several problem instances. It first finds size-specific models based on typical patterns of global and binary constraints, then generalizes them into a model that can run on unseen inputs. The system outputs MiniZinc code and a textual description of the model. The authors evaluate the tool on example problems from the PTHG21 Challenge and show the approach is feasible.
In briefModel-agnostic knowledge augmentation for machine learning on weakly annotated bioimaging data
SummaryThis paper proposes a model-agnostic knowledge augmentation method that lets machine learning models use domain knowledge beyond what automated feature extraction captures. It improves classification across seven ML architectures on weakly annotated microscopy and cytomorphology images, and performs best when combined with masked autoencoder self-supervised learning. The authors apply the pipeline to classify TDP-43 mislocalization in motor neurons derived from human stem cells, a phenotype linked to ALS, and report related biological findings.
Jayanie Bogahawatte, Sachith Seneviratne, Simon Maksour et al.
In briefA benchmark testing whether remote sensing AI answers match the image evidence it points to
SummaryThis paper proposes the Answer-Evidence Support Protocol (AESP), which checks whether the objects a model identifies in a remote sensing image actually support the model's own answer. The authors build the G-RSVQA benchmark with 15,027 samples, including questions, reference answers, and object-level annotations. Tests on seven grounded vision-language models show evidence support failures across existence, counting, and comparison questions, with up to 50.82% of correct answers not supported by the grounding evidence.
In briefPANDA: A plug-in dual-path attention module for hypergraph neural networks, guided by structural priors
SummaryHypergraph neural networks model relations among more than two nodes, but they usually rely on either fixed structural weights or normalised attention, each with drawbacks. This paper introduces PANDA, a plug-in for two-stage message-passing hypergraph networks that works for both undirected and directed hypergraphs. Its main path blends learned receiver-normalised attention with structural incidence priors, while a non-normalised auxiliary path lets each sender contribute its own transformed information. Across eight real-world benchmark datasets and four hypergraph networks, adding PANDA yields an average improvement of 5.18 percentage points.
Rishabh Jain, Federico Belotti, Stefano Coniglio et al.
In briefHumanize Olympiad Agents: A multi-session workflow using solving, review, a knowledge wiki, and Lean verification to tackle olympiad problems
SummaryHOA is an open framework that coordinates multiple LLM agent sessions through an independent reviewer, a structured task-specific wiki, and a verifier that checks Lean proofs against trusted statements. With GPT-5.6 Sol, it scores perfectly on the 2026 IMO, IPhO and IOI, nearly full marks on IChO 2026, and solves all 672 PutnamBench problems. With the open model Kimi-K3, it also achieves full scores on IMO, IPhO and IOI and 97.1% on IChO.
In briefPrivacy-preserving, distributionally robust federated multiclass classification for fault diagnosis
SummaryThis paper addresses fault diagnosis in asset-intensive industries, where multiclass models must learn from noisy, sensitive data held at distributed client sites. It proposes a differentially private, distributionally robust federated method built on a Wasserstein robust objective with Crammer-Singer loss, solved with ADMM and protected by calibrated Gaussian noise on client updates. The authors provide per-round and composed privacy guarantees, convergence results, and experiments on simulated industrial fault data and UCI datasets, reporting better performance than differentially private federated baselines in most settings.
Andrea P. Hernandez, Michael Ibrahim, Nagi Gebraeel
In briefMulti-Stream Attention Residuals: a multi-head extension of depth attention for more efficient large-scale pre-training
SummaryThis paper proposes Multi-Stream Attention Residuals (MSAR), an extension of Attention Residuals (AttnRes) that applies multi-head attention along network depth. Each head draws on a different contextual view formed by multi-scale causal convolutions, and the retrieved streams are merged by lightweight adaptive mixing into a single-stream layer input. According to the paper, MSAR outperforms standard residuals, mHC, and AttnRes on all 12 downstream benchmarks at 18B and 28B MoE scales. Its iso-loss compute speedup is 1.58× over the standard residual baseline, compared with 1.35× for AttnRes and 1.30× for mHC.
In briefJitterVLA: A replay benchmark for smooth robot motion in VLA-driven human-robot manipulation
SummaryVision-language-action policies send sparse action targets that reach the robot controller late and irregularly, which can cause abrupt motion near people even when predictions are reasonable. JitterVLA replays identical commands and arrival timing through different controllers to isolate the controller's effect. Tested on synthetic reaches, 20 people's hand trajectories, three open VLA pipelines (838 episodes) and a physical UR3, interpolation cut modeled p99 jerk on human motion with 200 ms delay from 615 to 106 m/s^3 (83%), at a 2.0 cm tracking cost. The benchmark is intended for screening before human-centred studies.
Faith Ogunwoye, Didar Seyidov, Konstantin Gubernatorov et al.
In briefAMLworld-Compact: Cheaper LLM Evaluation for Money-Laundering Detection via Weighted Sampling
SummaryThe paper addresses the high cost of evaluating large language models on the AMLworld anti-money-laundering dataset by proposing AMLworld-Compact, which keeps all illicit test transactions and samples benign ones from supervised-score strata. Horvitz-Thompson weights are used to estimate full-test confusion counts, and the sampled F1 error is reported to be lower than that of uniform sampling. In the evaluated setup, seven open-weight LLMs have weighted precision near the predict-all-illicit baseline, and lexical checks do not reliably establish verdict correctness.
Natapong Nitarach, Phume Ngampornsukswadi, Pittawat Taveekitworachai et al.
In briefSWE-DeceptionBench: A Benchmark for Measuring LLM Deception Across Software Development
SummaryThis paper introduces SWE-DeceptionBench, a benchmark that tests whether large language models behave deceptively during software engineering tasks. It covers six lifecycle stages and two deception types, with 720 test cases in Chinese and English under neutral, reward, and pressure conditions. Across seven mainstream models, deception occurred frequently even in the neutral condition, at aggregate rates of 49.88% for active and 64.29% for passive deception. The authors also find that guard models miss many deceptive requests and that Self-Reminder reduces but does not eliminate deceptive responses.
In briefMeasuring whether experiment results hold up across new datasets
SummaryThe paper defines external validity as the probability that two independent, same-sized replications of a study yield results whose maximum mean discrepancy (MMD) falls below a tolerance. It establishes an asymptotic power-law relation between study size and MMD quantiles, which allows estimating the size needed for an externally valid study from a small preliminary one. The framework is applied to rankings with ties and tested on existing case studies and a simulated study, with code released in the genexpy module.
WhyFormalizes external validity via MMD probability and proves a power-law link to study size, a novel framework; evidence rests on case studies and simulations.
In briefTraining user simulators with hidden in-conversation feedback
SummaryUser simulators are being used as environments to train LLM assistants for multi-turn conversations with real users. This paper introduces user simulators with intrinsic feedback, which have latent internal thoughts, reactions, and scalar ratings evaluating the assistant from a user's perspective at each turn. Trained on real user-assistant interaction logs, these simulators improve realism and critiquing behavior over prompted baselines and provide turn-level evaluations. They also serve as effective rewards for multi-turn reinforcement learning, matching setups using external judge LLMs while cutting RL compute by more than half.
In briefNotes on how generative diffusion relates to optimal transport
SummaryThis paper discusses three simple links between generative diffusion and optimal transport. The joint distribution of data and its Ornstein-Uhlenbeck noised version is, at each time, the solution of an entropic optimal-transport problem, so backward diffusion can be seen as annealing that gradually reduces entropic regularization. Using Gaussian mixtures as a high-dimensional example, the authors also compare how modes emerge in diffusion, flow matching, and optimal transport.
In briefCUBE: A Framework for Recognizing Unknown Insect Species Near Known Classes
SummaryThis paper addresses open-set insect recognition, where unknown species often look similar to known ones and are misclassified. It proposes CUBE, which models unknownness around each known species using class-conditioned boundary evidence learned from auxiliary species, plus taxonomic candidate calibration. On the Open-Insect C-America benchmark, CUBE reports AUROC of 95.8%, 98.7%, and 99.1% on local, non-local, and non-moth unknown splits, with 94.9% closed-set accuracy. CUBE-E, which adds ecological location metadata, further improves these results.
In briefAdversarial training enables continuous SPL control for vocal intensity conversion
SummaryThis paper proposes a method that converts a recorded utterance to a target vocal intensity on a continuous dB sound pressure level scale, rather than simply scaling amplitude. A lightweight converter is trained as a conditional cycle-consistent GAN without parallel data, operating on features from a frozen speech encoder and resynthesized by a pretrained vocoder. Among log-mel spectrograms, WavLM, and SpeechTokenizer, WavLM achieves a measured 83% of the target intensity change and also shifts fundamental frequency. A perceptual evaluation on amplitude-normalized stimuli confirms perceived intensity rises with target level.
Quentin Le Tellier, Albert Rilliard, Olivier Perrotin et al.
In briefBLiMPPI: A Multilingual Benchmark of Minimal Pairs for Negative Polarity Items
SummaryThis paper introduces BLiMPPI, a multilingual evaluation suite of over 12,000 minimal pairs testing negative polarity items (NPIs) such as "anymore" and "ever" across 33 languages. The pairs were built by discovering NPIs in Universal Dependencies treebanks, labeling their licensing environments, and perturbing sentences into minimal pairs. Evaluating 12 multilingual and 2 monolingual models against n-gram baselines, the authors find that multilingual models struggle most with lower-resource languages and with pairs not solvable by n-grams.
Arzu Burcu G\"uven, Anna Rogers, Leonie Weissweiler et al.
In briefzip2zip++: Reducing Quality Loss in Inference-Time Token Compression
SummaryThis paper extends zip2zip, which builds dynamic vocabularies at inference time using LZW compression so a language model can represent text with fewer tokens. The authors identify four sources of quality loss and address them with a decoupled input/output dynamic embedding layer, size-aware RoPE positioning, and a modified LZW codec supporting selective merges and multi-view segmentations, without changing the pretrained model. Across four models, the method improves perplexity on every evaluated corpus and mathematical reasoning accuracy, closing roughly half the gap to the uncompressed baseline on general text at a similar compression ratio.
In briefThyroid Ultrasound Nodule Segmentation with Conformal Uncertainty Estimates
SummaryThis paper proposes a framework for segmenting thyroid nodules in ultrasound images, combining DeepLabV3+, channel-attention modules, and post-hoc conformal prediction. It produces both segmentation masks and pixel-level uncertainty maps. Evaluated on the public TN3K benchmark, the plain backbone already outperforms previously reported results, and CBAM with Dice-BCE loss performs best overall.
Francesco Teodori, Pierpaolo Brutti, Umberto Ferraro Petrillo
In briefA small cognitive model explains large-scale human behavior
SummaryThe authors combine several established psychological principles into one model, called CORE, built on a hierarchy of episodic memories where prediction errors are stored. Encoding, retrieval, and forgetting depend on context. Fitted jointly to 7,036,442 choices from 137 experiments, it predicts human choices better than black-box models and reproduces 75% of reported behavioral effects in simulation.
In briefStudying how a retinal AI model organizes age-related macular degeneration severity
SummaryThis paper uses age-related macular degeneration (AMD) as a case study to examine how the RET-CLIP retinal vision-language foundation model represents disease severity in its embedding space, using bilateral fundus images from the AREDS dataset. The authors find that severity is represented as a continuous spectrum, with healthy and advanced stages separating more clearly than intermediate ones. Balanced accuracy was 0.844 for binary AMD detection, 0.814 for late-AMD detection, and 0.446 for six-class grading. The most influential latent dimensions were statistically associated with retinal features such as drusen accumulation, pigmentary changes, and advanced lesions.
Niccol\`o Fineschi, Jovana Panic, Atsushi Watanabe et al.
In briefAudioTETRIS: A training-free method to cut audio tokens in speech LLMs using acoustic transition cues
SummaryLarge audio language models produce hundreds of tokens even for short clips, which makes inference expensive. This paper proposes AudioTETRIS, a training-free, two-stage method that uses the audio encoder's representation geometry as a prior for token importance and combines it with attention scores inside the LLM decoder to progressively select tokens. Tested on Qwen2-Audio-7B, Phi-4-Multimodal, and Audio Flamingo 3, it reportedly keeps near-uncompressed performance on several speech tasks with only 25% of audio tokens and outperforms existing training-free baselines.
In briefEchoTab: Learning measurement-based echo representations for diagnosis via tabular reconstruction
SummaryThis paper introduces EchoTab, a framework for learning representations from multi-view echocardiography studies. It uses measurement-specific tokens as an intermediate layer between echo views and diagnosis, with each token attending to relevant views and trained by tabular reconstruction to predict a measurement value. Experiments on two datasets report improved diagnostic classification, and the learned attention aligns with measurement-view relationships used by cardiologists.
Michelle Espranita Liman, Ioana Ionac, Alexander M\"uller et al.
In briefTheoPhysEval: Testing AI agents on solving unpublished theoretical physics problems
SummaryThe authors built 50 unpublished, research-level theoretical physics problems with verifiable answers. With tools such as literature search and scripting, the best agent, GPT-6-Astra in Codex, reached 98% accuracy at an average cost under USD 3 per problem. Accuracy changed little when web search and background papers were removed, which the authors suggest indicates significant memorization of the literature.
A. Ramesh Chandra, Aaghaz Mahajan, Ainesh Sanyal et al.
In briefMENTOR: Learning from Failed Attempts Through Live Guidance
SummaryThis paper addresses "zero-reward collapse," where sudden difficulty increases in reinforcement learning for reasoning models cause the learning signal to disappear. The authors propose MENTOR, a multi-agent framework that diagnoses failure causes in real time and generates targeted prompts to guide the model's next attempt. MENTOR works as a plug-and-play module with existing methods such as GRPO, and experiments on math and code generation show it mitigates the collapse, with gains of up to 78.3% on Llama-3.2-3B-Instruct.
In briefLower Bounds for Randomized Linear-Oracle Optimization over Strongly Convex Sets
SummaryThis paper studies first-order optimization over strongly convex sets using only linear minimization oracle (LMO) calls. The authors prove a distributional lower bound of Ω(L·diam(S)²/T²) on objective error after T LMO calls for randomized first-order methods. Earlier lower bounds covered only the standard Frank–Wolfe algorithm or deterministic methods, so this result closes that gap. The same error threshold is shown to hold with probability at least 3/4, along with a stopping-time tail bound.
WhyProves a distributional Ω(1/T²) lower bound for randomized LMO methods on strongly convex sets, closing a gap left by deterministic results, with rigorous analysis.
In briefForecasting EV charging demand with uncertainty bounds across a changing network
SummaryThis paper studies a large dataset of public EV charging stations in Scotland, where demand is organized hierarchically from single charging points up to regional and national totals. Because stations are regularly added or removed, the authors combine hierarchical forecasting with conformal prediction methods that handle missing data. The resulting prediction intervals are valid at every node of the hierarchy, and the CP-MNR-Nested algorithm accounts for the time-varying structure.
Guillaume Principato, Yvenn Amara-Ouali, Yannig Goude et al.
In briefTesting whether AI-text detectors rely on formatting artifacts
SummaryThe paper finds that near-perfect scores of machine-generated text detectors on scientific-literature benchmarks partly stem from document formatting rather than wording. A logistic regression using 27 superficial features reached macro-F1 of 0.833 to 0.997, outperforming every evaluated detector on three arXiv subsets. Merely re-wrapping a machine-written abstract reversed 20-27% of a public detector's correct decisions. Normalizing formatting before training reduced the generalization gap from 0.32 to 0.22.
In briefCorrection Training for Diffusion Models: a corrector that narrows the train–inference gap
SummaryIterative sampling in diffusion models can accumulate errors because the denoiser sees its own imperfect predictions at inference, which it never encountered during training. The paper proposes a corrector that takes the base model's prediction as input, is trained against the same ground-truth target, and is applied at each denoising step to refine clean-sample predictions. Evaluated on DiT, SiT, and JiT, the method substantially improves unguided sample quality, performs competitively with classifier-free guidance, and combined with CFG improves diversity and reduces oversaturation. The extra inference cost can be amortized by distilling the refined predictions back into the base model.
Donghoon Ahn, Runqian Wang, Alexei A. Efros et al.
In briefRARA: Robust robot policy fine-tuning by aligning representations with retrieved pre-training examples
SummaryThis paper proposes RARA (Retrieval-Anchored Representation Alignment), a two-stage method for adapting pre-trained robot manipulation policies to new tasks using only a few demonstrations. It first retrieves pre-training representations with similar task intent and motion and aligns the target demonstrations' representations to them, then fine-tunes while limiting changes to those retrieved representations. Experiments on 13 tasks in RoboTwin, RoboCasa, and a real robot, using Diffusion Policy and GR00T, report that RARA outperforms all compared baselines.
In briefA randomized online method for TCP acknowledgments with bounded cost and regret
SummaryThis paper studies TCP acknowledgment, where one acknowledgment serves all pending packets and the goal is to minimize total acknowledgment and waiting costs. The authors construct a randomized online algorithm whose total cost is at most (2+ε) times the offline optimum, with sublinear regret against the best of k experts. They prove the regret bound is asymptotically optimal in several parameters and that ε>0 is necessary for sublinear regret.
In briefTrade-off curves for stopping reasoning models that overthink
SummaryThis paper studies "overthinking" in reasoning models, where continued reasoning at test time can become ineffective or harmful. The authors introduce Step-wise Truncation Probing (STP), which generates answers from multiple prefixes of each sampled reasoning trajectory, and define Overthinking Ratio and Accuracy Degradation metrics. They compare dynamic routing, fixed budgets, and dynamic early exit in a unified framework and build token-accuracy Pareto frontiers for each strategy. The paper also releases OTBench, a reusable benchmark built from the STP process.
In briefTeam game winning without shared randomness, in the limit
SummaryThis paper studies concurrent stochastic games where a team of agents competes against an adversarial environment, with each agent having access only to its own individual source of randomness. The authors ask whether the team can meet a reachability or safety objective with probability at least 1−λ for every λ>0. They show that limit-sure winning for reachability objectives is NP-complete, while for safety objectives it coincides with almost-sure winning, which lies in P. As a corollary, model checking the qualitative fragment of Individually Randomised ATL is P^NP-complete.
L\'eonard Brice, Thomas A. Henzinger, Alipasha Montaseri et al.
In briefRationally Linearized MLPs: a method that uses low-degree rational approximations to cut MLP inference cost
SummaryThis paper proposes RaLiMPs, a method for efficiently linearizing multilayer perceptrons by approximating standard activation functions with low-degree rational functions. The approximations are used to build kernel representations with narrow bottlenecks that compose efficiently across layers. The authors provide theoretical guarantees for the approximation and evaluate the method on applications from implicit 3D models to large language models, reporting competitive or improved accuracy with substantially lower inference cost.
Arijit Sehanobish, Byeongchan Kim, Avinava Dubey et al.
In briefNot All Correct Fixes Perform Alike: Patch Runtime Performance in Repository-Level Issue Resolution
SummaryThis paper examines whether patches that pass correctness tests on repository-level issue benchmarks such as SWE-bench differ in runtime performance. The authors reuse the built-in FAIL_TO_PASS tests to measure instruction count and wall-clock time, and apply Welch ANOVA for statistical comparison. They find performance differences among patches in almost half of the instances, that runtime performance is independent of models' resolution rates, and that some patches contain architectural regressions missed by correctness checks.
In briefWhat We Talk About When We Talk About Emotion Editing: ROI-ReFT, brain-response-guided image emotion editing
SummaryThis paper studies how to change the emotion an image conveys while preserving its essential content. Using the fMRI encoding model TRIBE v2, the authors analyze 400 source-target pairs across eight emotion categories and find that hidden states of existing editors correspond more strongly to predicted visual-region responses than to affect-related regions. They propose ROI-ReFT, which uses source visual responses as a content reference and predicted affective response changes as an editing direction to guide a frozen editor. On a post hoc selected set of 400 requests, it reports 91.75% target-emotion accuracy and 83.25% joint success.
In briefForecasting ship paths at sea with Transformer models trained on AIS position data
SummaryThis paper studies Transformer and Performer models for predicting surface vessel trajectories in confined waters, using widely available AIS position data. Local traffic history is encoded as kernel density estimation heat maps, processed by a convolutional neural network, and fed into the Transformer models. Training with quantile loss yields a model-aware prediction interval relevant to collision avoidance. The authors report that the proposed architectures greatly outperform previous work, and that local heat maps help predict movements not indicated by previous positions alone.
In briefLearning scene layouts that stay robust to viewpoint shifts in cross-view geo-localization
SummaryThis paper addresses cross-view geo-localization, where large appearance changes and geometric distortions arise from viewpoint differences. It proposes a Spatial Context Modeling (SCM) module that scans feature maps along four orthogonal directions to build richer scene layout representations. It also introduces a High-Frequency Information Guidance (HFIG) module that emphasizes edges and contours from the frequency domain, which are less sensitive to appearance changes. The authors report competitive results on standard benchmarks and strong robustness with unknown orientations and limited fields of view.
In briefPersonalization without stereotyping: when speech AI models should and should not use a speaker's demographic cues
SummaryThis paper introduces BASIS-Bench, a benchmark that separates beneficial personalization from stereotypical behavior in speech LLMs across three axes: demographic robustness, stereotypical bias, and personalization. Evaluating eight recent speech LLMs, the authors find that most models recognize demographic cues but fail to use them appropriately for personalization. Explicit demographic labels improve personalization but also amplify stereotypical associations already present in the text backbones.
Kshitij Ambilduke, Caroline Brun, Thibaut Thonet et al.
In briefELMA: A Benchmark for Long-Term Text-to-Motion Generation with Anaphoric Instructions
SummaryThis paper introduces a task of generating long-term human motion from multi-segment text instructions that contain anaphora, where constraints from earlier segments must persist. The authors present ELMA, a dataset of continuous 30-second clips extracted from in-the-wild videos, with dense segment-level annotations preserving cross-segment dependencies. Experiments show that current motion generation models struggle with context-dependent instructions, and fine-tuning on ELMA improves global physical realism and semantic alignment.
In briefAuditing vision-language model benchmarks using saved evaluation outputs
SummaryThe paper adds an audit layer to VLMEvalKit that reuses per-question outputs saved from existing evaluation runs, applying six detectors plus one text-only rerun with images removed. Auditing MMStar, a benchmark built so images are essential, with three models, the authors find zero aggregate accuracy gain from the image on 20.2% of questions. After a permutation-null correction, the excess is concentrated in mathematics and science categories.
WhyReuses existing evaluation artifacts to audit benchmarks, quantifying image-gain excess with item-matched permutation nulls; novel framing with fairly solid supporting analysis.
Denis Degtiarev, Anna Gromova, Sergey Katkov et al.
In briefMOSAIC: Object-Centric Slot Adaptation for Remote Sensing Image Super-Resolution
SummaryThis paper proposes MOSAIC, a framework that adapts a pretrained diffusion Transformer to remote sensing image super-resolution using compact object-centric structural conditions. It organizes degraded images into structural slots, integrates them across resolutions, and retrieves recurring priors from a shared prototype memory. These representations are injected into a frozen diffusion Transformer via lightweight residual modulation. Experiments on UCMerced, AID, and RSSCN7 report consistent PSNR gains at ×2 and ×4 and favorable LPIPS at ×4.
In briefConfidence-aware densification for 3DGS under a fixed budget
SummaryThe paper addresses densification in 3D Gaussian Splatting under a fixed primitive budget, arguing that placing new primitives by image error rather than opacity, as 3DGS-MCMC does, leaves background regions under-reconstructed. The method uses a learned confidence to direct new primitives and relocates low-contribution primitives to poorly explained regions. Averaged over 24 scenes, it outperforms 3DGS-MCMC in PSNR, SSIM and LPIPS at every budget from 100k to 2M primitives.
Manuel Hofer, Thomas K\"ohler, Felix Windisch et al.
In briefUnsupervised detection and suppression of traits transferred via subliminal learning in model distillation
SummaryWhen a student language model is distilled from a teacher sharing its base parameters, it can inherit the teacher's behavioural traits, including misalignment, even from semantically unrelated data. This paper proposes a method that requires no trait labels, recovering such subliminally transferred traits from parameter updates alone by exploiting their consistency across diverse teacher-generated datasets. The recovered weight-space direction can be amplified to interpret the trait or projected out to suppress it. On existing subliminal-learning benchmarks, the method recovers animal-preference traits and suppresses them with almost no impact on task performance.
Joseph Egan, Ciar\'{a}n M. Gilligan-Lee, Michael O'Riordan et al.
In briefKnowing Whether, Not Just Where: Anchor-Probing Distillation for Open-Set Remote Sensing Grounding
SummaryThis paper introduces AnchorProbe-Distill (APD), a teacher-student framework that treats open-set remote sensing visual grounding as learning visual support. The teacher uses the original query as a reference anchor to interpret query perturbations, and this reasoning is distilled into a student model. The student performs localization and rejection from a single image-query pair. Experiments show the model can reject absent targets while also improving localization of present ones.
In briefMAGMA-Bench: A benchmark testing long-running robot agents on changing tasks and execution failures
SummaryThis paper introduces MAGMA-Bench, a benchmark for testing long-horizon embodied agents in dynamic multi-robot environments, independent of any specific architecture. It evaluates agents on longer tasks and online disruptions such as execution failures, interruptions, and mission updates, including scenarios requiring coordination among robots with different capabilities. Results show that even the strongest models fail on many clean episodes and perform worse under disruptions, while supervised adaptation helps a small model but performance stays lower on longer and multi-robot missions. The authors release the benchmark, 261,646 task-level training examples, and an expert-trace generation pipeline.
Loan Bernat (LAAS-GEPETTO), Matthieu Grard, Ariane Herbulot (LAAS-RAP) et al.
In briefS-Space: A Spatial Workspace Inside Multimodal Language Models
SummaryThis paper identifies S-Space, a linear subspace in multimodal language models where object representations encode continuous horizontal, vertical, and distance coordinates. The authors argue it functions as an internal spatial workspace used in perception, viewpoint-changing reasoning, and verbal responses, supported by causal interventions. S-Space is also used as a diagnostic tool, revealing that models can form spatial representations but struggle with transformations needed for reasoning across viewpoint changes..
In briefSimulating loss spikes in deep network training with layer-wise random feature models
SummaryThis paper studies the intermittent loss spikes that appear during pre-training of large language models. The authors show that simple quadratic models with heavy-tailed, power-law distributed inputs can reproduce these spikes. They introduce a theoretical framework and simulation algorithm that model per-layer training dynamics of deep neural networks using such models. In a case study with an Adam setting known to cause spikes in Transformer language models, the simulation predicted spiky behavior from a given checkpoint, and a β2 schedule derived from observed metrics empirically improved performance.
In briefBeyond Successful Demos: Learning Robot Interactions and Failure Dynamics from Human Videos
SummaryThis paper introduces HF-WM, an action-conditioned world model that predicts the consequences of robot actions. It bridges the gap between human hands and robot manipulators with a dual-branch framework using stabilized pointflow as a shared motion representation. It also applies GRPO post-training on perturbed actions from successful demonstrations to learn failure dynamics. The authors report state-of-the-art action-conditioned video generation and useful policy evaluation in downstream robotic tasks.
In briefCross-Teacher Guarding for Multi-Teacher On-Policy Distillation (Road MOPD)
SummaryThis paper addresses capability degradation in multi-teacher on-policy distillation, where gradient conflicts arise between domain-specific teachers sharing student parameters. The proposed Road MOPD method uses non-routed teachers as auxiliary cross-domain critics on the same student trajectories, coordinating supervision via teacher-agreement-based token credit assignment and multi-teacher gradient projection. The authors report mitigated degradation in the disadvantaged domain, over 5 percentage point gains under high conflict, and overall improvements in low-conflict and three-domain settings.
In briefWhether causal effects can be identified across causal abstractions, with a strict hierarchy of criteria
SummaryThe paper studies whether causal effects can be identified from observational data at the level of a collection of compatible causal graphs, produced by causal abstractions, rather than a single graph. It formalizes several identifiability notions and organizes them into a hierarchy that clarifies how they relate and differ. The authors show that an effect can be identified over a graph collection without a single derivation valid across all compatible graphs.
Cl\'ement Yvernes, Marianne Clausel, Emilie Devijver et al.
In briefA stochastic-analysis view of diffusion models: score functions, time reversal, and sampling
SummaryThis paper reviews score-based generative models from a stochastic-analysis perspective, linking classical sampling theory, score estimation, and diffusion time reversal. It explains how score functions connect Langevin sampling with score matching, and how Gaussian smoothing leads to denoising score matching and multiscale sampling methods. It distinguishes dynamics that preserve a fixed distribution from dynamics that transport a prescribed family of marginals, and treats diffusion time reversal in path space. The paper ends by connecting ideal reverse diffusion to practical generative algorithms.
In briefAsymptotic analysis of discrete-time self-attention dynamics in transformers
SummaryThis paper studies the self-attention mechanism of transformers by treating the layer index as discrete time. It classifies all equilibria of this discrete-time dynamics and analyzes their stability for both single-head and multihead cases. The main attractor is a consensus equilibrium aligned with the principal eigenvector of the value matrix, and this behavior is confirmed using weight matrices from a real transformer.
In briefIs This Alarm Worth Acting On? Counterfactual Evaluation of VLA Runtime Monitors
SummaryRuntime failure monitors for robot policies are usually judged by how well they separate failed from successful runs, but a late or unnecessary alarm can be unhelpful or disruptive. The paper introduces RESCUE, a counterfactual framework that restores simulator and policy states at checkpoints, compares alternative actions, and tests whether acting on alarms reduces expected task-completion cost. Experiments with OpenVLA-OFT and π0.5 on LIBERO-Spatial and LIBERO-Plus use 128,340 counterfactual branches.
David Ahmedt-Aristizabal, Abdelwahed Khamis, Mohammad Ali Armin et al.
In briefPRISM: A Retrieval-Augmented Multimodal Model for Clinical Prognosis
SummaryThis paper introduces PRISM, a multimodal language model for clinical prognosis that combines per-feature numerical tokenization, recency-weighted aggregation of longitudinal ECGs, and a population memory bank. Retrieval uses mean-pooled embeddings of similar training-cohort encounters rather than individual records. On the 1,443-task MDS-ED benchmark, the authors report AUROCs for diagnosis, deterioration, ICU admission, and mortality, along with ablations of each component. Supervised fine-tuning on the external MC-MED cohort reaches 0.8611 AUROC.
In briefFedPCD: Priority-Constrained Federated Aggregation
SummaryThis paper proposes FedPCD, a federated learning method that separates the global optimization objective from protection of individual clients. It keeps conventional federated aggregation as the main goal and applies only the minimum correction needed to meet client performance constraints. According to the authors, it outperforms state-of-the-art baselines on heterogeneous benchmarks, achieving the highest global accuracy and converging in fewer communication rounds.
In briefBeyond Temporal Proximity: Explanatory World Models for Credit Assignment in Reinforcement Learning
SummaryThis paper introduces Explanatory World Models (EWM), which uses explanations from a language model to identify earlier events that plausibly led to a delayed outcome and redistributes reward accordingly. It targets the difficulty standard reinforcement learning has with rewards that arrive long after the actions that caused them. EWM outperforms baselines such as PPO and LaRe on Crafter, WordCraft, and KuaiRand, and improves Recall in an offline evaluation of an industry generative recommender.
WhyUses LLM explanations for causal credit assignment with conservative reward redistribution, validated against baselines on several long-horizon benchmarks and an industrial recommender.
Oriol Corcoll, Qiyao Wei, Ciar\'{a}n Mark Gilligan-Lee
In briefSplatQ: Interaction-Aware Pruning of 3D Gaussian Splatting via Quadratic Selection
SummaryThis paper presents SplatQ, a method for pruning 3D Gaussian Splatting (3DGS) scenes under a fixed budget. It groups Gaussians into size-limited spatial and appearance blocks, smooths per-block retention quotas, and models individual importance, pairwise deletion interactions, and quota penalties in a QUBO objective solved with simulated annealing or a coherent Ising machine. The authors report that on 5 benchmarks, SplatQ outperforms state-of-the-art pruning methods in reconstruction quality, spatial quality balance, and rendering speed.
In briefStride: Scale-Aware Multi-Agent Evolutionary Refinement for Generating Infographics
SummaryThis paper presents Stride, a multi-agent framework that iteratively refines infographics produced by a text-to-image generator. It separates checking whether constraints are met from locating where they fail, then sends global fixes to a composer agent and local fixes to a retoucher agent. A candidate is kept only if it raises the verifier score, and the authors report results that beat several systems on BizGenEval, IGenBench, and WISE.
In briefHumanize MLE Agents: Two AI Agents Taking Turns on Machine Learning Tasks
SummaryThis paper proposes Humanize MLE Agents (HMA), in which two LLM agents alternate sessions on a shared workspace, with each session limited to at most k experiments and context reset between sessions. Across 75 Kaggle tasks in MLE-bench, HMA reaches a state-of-the-art any-medal rate of 78.2%, exceeding the strongest open-source baseline. The authors also report applicability to a kernel optimization case.
WhyProposes session alternation with context resets, backed by theoretical analysis and strong results across 75 Kaggle tasks, with solid evidence and moderate field impact.
In briefNeural block-wise auto-exposure for capturing high dynamic range scenes in one shot
SummaryThis paper proposes a neural algorithm that sets exposure times for each non-overlapping 16×16 block, using a temporary image and its exposure time, so a high dynamic range scene can be captured in a single shot. The temporary image can be sRGB or raw. The authors explore raw input because it measures radiance in highlight and shadow regions more accurately, helping capture more detail. Experiments validate the proposed algorithm.
In briefProactive Agentic Orchestration: Sub-Agents Request Mid-Task Adjustments to Their Execution Setup
SummaryThe paper introduces POrchestra, a framework that lets sub-agents proactively communicate with the orchestrator at any logical point to request harness adjustments while keeping local execution continuous. It also proposes On-Policy Orchestration Distillation (OPOD), which attributes these revision requests to earlier orchestration decisions and uses them as process supervision alongside the final reward. The method is evaluated on three long-horizon agent benchmarks covering deep research, dynamic digital assistance, and software engineering.
In briefBranching On-Policy Distillation: branching at uncertain decisions to widen teacher supervision coverage
SummaryThis paper proposes BranchOPD, an on-policy distillation method. It selects decision points using student uncertainty and teacher-student disagreement, resamples student continuations from shared prefixes, and applies token-level teacher supervision to both main trajectories and branches. The authors report that BranchOPD outperforms Vanilla OPD and Multi-Rollout OPD (n=7) across evaluated benchmarks in mathematics and agent tasks, and that branching at uncertain decisions expands subsequent behavior coverage.
In briefA scalable learned solver that speeds up real-time nonlinear model predictive control
SummaryNonlinear model predictive control is costly to solve online, which limits its use for fast or large systems. This paper proposes a learned iterative solver that exploits the temporal structure of the problem so its network size does not grow with the prediction horizon, and uses a closed-loop resampling strategy for training. On an economic NMPC benchmark of a CSTR cascade, it scales to larger problems than the monolithic LISCO and reaches median speedups of up to four times over IPOPT. In closed loop, its control performance is comparable to IPOPT with no observed constraint violations.
In briefA local agentic AI framework that generates edge-case driving scenarios as traceable safety evidence
SummaryThe paper presents a laptop-based agentic AI framework that uses local language models to adaptively generate edge-case scenarios for training, testing and safety assurance of autonomous systems. A second independent model critiques and refines each scenario against domain criteria, and the refined scenarios are run in simulation. Clustering of metric scores and thematic analysis of scenario descriptions then identify functional gaps and their triggering conditions, producing traceable evidence for a safety assurance case.
In briefWhether reinforcement-learning-driven reasoning can be physically implemented
SummaryThe paper examines whether a reinforcement-learning-driven reasoning process can be implemented in hardware using a physical neural network model with a reward external potential energy. The authors report a negative result. They argue that, in this hardware implementation, a chain of thought in the system's dynamics conflicts with Lyapunov stability.
In briefA Better Metric Will Not Save You: Comparison Choices Matter More Than the Measure
SummaryThis paper examines the limits of representational similarity measures such as RSA, CKA and Procrustes. The authors argue that these methods all take a given set of conditions for granted, and that this set is determined by which inputs are admitted and how they are grouped, choices no invariance property controls. Using the known acoustic organisation of vowels across fifteen languages and twenty-seven representations, they find that changing which categories are compared shifts the gap between two representations by more than twice the measured effect, while changing the number of categories has no effect.
In briefGraftOOD: Grafting out-of-distribution examples into advantage-weighted correction for video reward optimization
SummaryThe paper observes that optimizing video generation with reward signals can raise proxy scores while reducing motion, a problem it calls motion collapse. It proposes GraftOOD, which replaces part of the online rollouts in selected prompt groups with out-of-distribution video examples, and builds GraftSet-Video with 9,588 exemplars for 2,397 prompts. On Wan2.1-1.3B, Dynamic Degree rises from 4.86 to 44.41 compared with the reward-only FlowAWR, while HPSv3 falls from 8.10 to 7.32.
In briefA real-world CAD segmentation benchmark comparing rule-based and learning-based hole detection methods.
SummaryExisting CAD segmentation datasets are mostly synthetic or automatically labelled, so their relevance to real engineering tasks is unclear. This paper focuses on fastener-related hole detection, introducing two rule-based methods and a human-annotated benchmark built from engineering CAD data, which also trains the learning-based UV-Net and AAGNet. Results show rule-based methods generalise well without training data, while learning-based methods perform strongly with sufficient annotations but are less robust under domain shift.
In briefWhy multimodal LLMs miss text humans can read by squinting, and a method to fix it
SummaryThis paper studies a failure in which multimodal LLMs miss text that humans can recover by squinting, sometimes reporting that no text exists. The authors introduce SquintBench, with 1076 moderation-derived images, and SquintCtrl, a controlled synthetic suite for diagnosis. They propose Squint OPSD, where a frozen teacher with a downscaled view and ground-truth transcription supervises a student that reads only the original image. Squint OPSD reaches 88.61 ANLS on SquintBench, versus 26.50 for the unadapted backbone.
In briefA closed-form solution for computing Gaussian BROJA information decomposition
SummaryThe BROJA partial information decomposition splits the information two sources carry about a target into unique, redundant, and synergistic parts, but it is hard to compute for high-dimensional continuous variables. This paper derives a closed-form solution to the optimization problem in the general multivariate Gaussian case, so the decomposition can be calculated directly. The authors report speedups of over two orders of magnitude for high-dimensional Gaussian BROJA PID.
In briefSqueezeBench: A Math Benchmark That Resists Saturation
SummarySqueezeBench is a benchmark for measuring AI mathematical ability that does not saturate, because each success raises the bar rather than closing the problem. Each problem concerns an integer function known only for small n, and submissions must provide machine-checked Lean proofs that a new bound is valid and strictly better than the current record somewhere. The authors release ten extremal combinatorics problems with twenty initial records. In baseline runs, agents based on a frontier model had 82 submissions accepted and improved every record, and two newer models later improved 19 of the 20 records.
WhyProposes a saturation-resilient benchmark with machine-checked Lean record updates, a novel framework; baseline experiments are modest in scale.
In briefEnriching design-time models using evidence from simulation and operation
SummaryThis paper addresses how knowledge gained from simulation and operation of cyber-physical systems can be systematically fed back into design-time models. It proposes an evidence-driven workflow and demonstrates it on an Influence Model, using a sign-following robot use case. The authors show that simulation evidence can support participant pruning, interaction characterisation, influence-function abstraction, and validity-scope information. The results indicate the workflow is feasible for simulation-based evidence and may improve design-time reasoning in later iterations.
Barbara da Silva Oliveira (UniCA, Laboratoire I3S - COMRED, KAIROS) et al.
In briefAgentic Large Language Models: How LLMs Can Act in the World and Interact with Each Other
SummaryLarge language models can communicate in human language and reason about the world, and are widely used for writing and answering questions. This book focuses on agentic LLMs, which combine these language abilities with the ability to act directly in the real world. It discusses how agentic LLMs work, how they act, how they interact with each other, and what this means for users.
Max van Duijn, Michiel van der Meer, Aske Plaat et al.
In briefO'ido: Open-vocabulary speech recognition on a cheap microcontroller without a neural chip
SummaryThis paper presents O'ido, an inference engine that runs open-vocabulary English speech recognition entirely on an ESP32-S3 microcontroller without a neural accelerator. It executes a 13M-parameter Conformer-CTC model using int8 quantization and a schedule designed around flash and PSRAM bandwidth. On LibriSpeech, it makes fewer errors than the vendor's recognizer on the same chip, but is not yet real time.
WhyOpen-vocabulary speech recognition runs on a $5 MCU; LibriSpeech WER measured on one chip, though not yet real time.
Daniel Varela S\'anchez, Guillermo Mart\'in-Coello Ju\'arez
In briefLearning by Compression: Risk Bounds Reach the Law-of-Large-Numbers Rate
SummaryThis paper studies data-dependent risk bounds for learning algorithms derived from compression schemes. Existing bounds carry an extra O(√(ln N/N)) term, and the authors show the logarithmic factor is not intrinsic. They prove that with probability at least 1−δ, risk is bounded by k/N + O(1/√N), uniformly over all values of k. The leading term k/N is shown to be necessary, and the paper examines when the empirical risk keeps the O(1/√N) rate and when it converges more slowly or becomes inconsistent.
WhyRemoves the logarithmic factor from compression-based risk bounds, matching the law-of-large-numbers rate with full proofs and characterizing regimes where empirical risk fails.
In briefRe-examining DistilHuBERT Accent Classification with Speaker-Disjoint and Cross-Corpus Tests
SummaryThis paper re-evaluates DistilHuBERT-based accent classification using L2-ARCTIC and several external English accent datasets. It compares file-level splits with speaker-disjoint splits and tests balanced external-data training and teacher-student pseudo-labeling. The authors find that file-level evaluation substantially overestimates performance, while speaker-disjoint evaluation gives a more realistic picture of generalization to unseen speakers.
