In briefT-JEPA: A Temporal Architecture That Learns Better Satellite Image Representations
SummaryThe paper proposes T-JEPA, which uses the sparse, uneven time gaps in Earth observation image sequences as a training signal. It predicts latent states conditioned on elapsed time and adds masked pixel reconstruction to preserve spatial detail. Under matched pre-training data and throughput, it reports leading transfer performance on static and temporal tasks, and analyses show time-gap-dependent predictability and coherent latent dynamics.
In briefFreSia: Frequency-Semantic Alignment for Multivariate Time Series Analysis
SummaryThis paper proposes FreSia, a framework that aligns large language models with the frequency structure of time series to improve multivariate time series analysis. Its FGPrompt mechanism extracts frequency-domain structure and turns it into prompts suited to LLMs, while the GCL component uses a global CLS-driven probe to bridge the time-frequency gap and fuse multimodal information. According to the paper, experiments on eight forecasting benchmarks show average MSE and MAE improvements of 13.48% and 8.06%.
In briefReconstructing hidden fields and unknown parameters of coupled physical systems from partial single-field measurements using physics-constrained neural networks
SummaryThis paper addresses how to reconstruct all coupled physical fields and identify unknown system parameters when only sparse measurements of a single field are available. The authors propose a framework in which a cross-attention encoder maps sparse sensor data onto a regular grid, and a Fourier neural operator decoder reconstructs all coupled fields. Network weights and unknown physical parameters are jointly optimized by minimizing observation loss, governing-equation residuals, and boundary/initial-condition constraints. The method is validated on 2D and 3D lid-driven cavity flows, a 2D cylinder wake, and a 2D non-ideal magnetohydrodynamics problem, showing recovery of unobserved fields and parameters.
In briefReMaD: Tuning-free Domain Adaptation for Classification and OOD Detection
SummaryThis paper introduces Reduced-rank Mahalanobis Distance (ReMaD), a prototype-based distance method for classification and out-of-distribution (OOD) detection that uses pretrained models without finetuning. It fits closed-form distribution statistics to target-dataset embeddings in the model's latent space, requiring no training or prior knowledge of OOD data. The authors modify Mahalanobis Distance by removing unused features to adapt latent distributions to new domains. Tested on four target datasets, the method shows competitive classification and OOD detection performance.
In briefRetrieval-Based In-Context Learning: A Domain Adaptation Framework
SummaryThis paper frames in-context retrieval (ICR), where demonstrations are retrieved from a database by similarity to the query, as a domain adaptation problem. The database distribution may differ from the test query-label distribution. The authors analyze ICR performance under a flexible class of distribution shifts that extends prior work, and provide theoretical guarantees quantifying its benefits and pitfalls. Experiments on synthetic and language tasks verify the theory.
In briefRotated, but by how much? Diagnosing object-rotation reasoning in VLMs
SummaryThis paper finds that vision-language models can detect that an object has rotated between views but cannot reliably estimate the rotation amount. The authors introduce OR-Bench, a fine-grained benchmark with eight tasks, and report that models' exact-angle answers concentrate on 0°, 90°, and 180°. They propose RotationCue, a lightweight decoder that recovers coarse rotation information from the model's frozen representations and feeds it back as text, which improves accuracy across all model-task combinations on three VLMs.
In briefLearning image generation and visual representations together from pixels, without labels or pretraining
SummaryThis paper introduces SCION, a single model that jointly performs image generation and self-supervised representation learning, requiring no class labels or pretrained encoders. Its core is a pixel-space encoder conditioned on the flow timestep and an embedding, trained with a contrastive loss and flow matching together in one run. On ImageNet 256x256, the authors report an FID of 8.92 without representation guidance, which they state surpasses iREPA and RCG that rely on pretrained DINOv2.
WhySingle pixel-space model jointly learns generation and self-supervised representations without labels or pretraining, reaching 8.92 FID on ImageNet and beating several baselines with solid experiments
In briefNexus: A Cloud-Edge Execution Platform for AI Agents on Devices
SummaryThis paper presents Nexus, a cloud-edge platform that treats each AI agent invocation as a persistent task. It uses an OpenWrt-based runtime for distributed serving, run-scoped delegation for authorized access to computer and mobile environments, and persistent records to track execution, failures, recovery, usage, and charging. In the evaluation, all ten Computer-Android workflows succeeded and all six revocation tests blocked later writes, while Nexus completed 24 of 24 matched tasks compared with 22 for Dify.
In briefDifference Feature Map Distillation: Compact Transformer Trackers via Inter-Sample Relations
SummaryThis paper proposes Difference Feature Map Knowledge Distillation (DFM-KD), a method for training smaller transformer-based visual object trackers. Instead of matching absolute feature values, it transfers how the teacher model relates feature differences across samples. Experiments reported in the paper show DFM-KD outperforms conventional feature-level distillation in tracking precision and success rate.
In briefSecond-Order Problem Solving for Self-Improving AI in Formal Verification
SummaryThe paper notes that standard recursive self-improvement (RSI) often patches surface-level parameters in response to failures, leading to trial-and-error without fixing root causes. The authors propose SO-RSI, which monitors execution traces for three structural anomalies and investigates why failures occur before making structural workflow edits. Under matched 24-hour search budgets, SO-RSI improves final held-out pass rates over Naive RSI by 21.8 points on Lean 4 proof generation and 25.8 points on Verus-based verifiable code generation.
In briefVoltic: Separating Volatility from Noise in Recurrent Memory Updates
SummaryThe paper proposes Voltic, a recurrent memory model that treats the memory write as a Kalman gain from Bayesian filtering, modeling how fast associations change (volatility) and how noisy observations are (stochasticity) separately. To keep parallel training feasible, it introduces diagonal and quasi-diagonal approximations of the covariance. On controlled recall tasks and a 45M-parameter language model reasoning average, the authors report that Voltic outperforms the baselines.
In briefData fusion method for errors-in-variables models that uses external repeated measurements to correct measurement error
SummaryThe paper studies errors-in-variables problems, where the target study has only one error-prone surrogate of an unobserved exposure and an external study provides repeated measurements from a different population. The authors introduce a conditional transportability assumption to identify the target measurement-error distribution and build a data-fusion estimator for a broad class of functionals. They establish consistency, convergence-rate bounds, and inference results. In simulations the Fuse-EIV estimator shows small bias for the primary exposure coefficient, and an NHANES application illustrates how population heterogeneity and error heteroscedasticity can change conclusions.
In briefSEAL: a compact speech separation model with additive reconstruction and expert routing
SummaryThis paper presents SEAL, a compact time-frequency speech separator. It uses a zero-sum additive residual bounded by local mixture amplitude for reconstruction, and routes each token to one of six residual experts using a query built from acoustic and inter-step evidence. On EchoSet, SEAL (small) beats TIGER (small) by 0.31 dB SI-SDRi with 28% fewer parameters and 2.9 times fewer MACs. SEAL (large) is within 0.07 dB SI-SDRi of TIGER (large) at 3.1 times fewer MACs.
In briefPlanetary Geospatial Foundation Models for Global Public Health
SummaryThe paper presents Google Earth AI's Population Dynamics Foundation Model (PDFM), a geospatial model meant to complement traditional epidemiological forecasting, which is limited by spatial gaps and time lags. It reports case studies across four countries (USA, Canada, Mexico, and the Democratic Republic of the Congo) covering vaccine-preventable, communicable, noncommunicable, and maternal mental health conditions. The reported results include improved US-Canada border MMR vaccination coverage predictions, cardiovascular disease nowcasting, short-term Mexican dengue forecasts, cholera hotspot forecasts, and added signal for postpartum-depression risk prediction.
Arbaaz Muslim, Aviv Slobodkin, Katherine Wheeler-Martin et al.
In briefFrom Token-Max to Outcome-Max: Rewarding Verified AI Results Over Token Use
SummaryThis paper argues that more AI usage does not automatically yield proportional productivity gains, partly because treating token consumption as productive effort encourages agents to over-compute. The authors propose outcome-max, which rewards independently verified task completion per unit cost, and test it in a three-level simulation framework covering immediate interaction, long-run adaptation, and organizational collaboration. They also introduce OutcomeShare, an incentive mechanism that, according to theory and simulation, can encourage participation and produce shared gains for employees, firms, and LLM providers.
In briefErrors in AI-assisted environmental science literature searches: abstract vs. full-text prompts
SummaryThis study compares how accurately six large language model platforms retrieve references for original articles from five leading environmental science journals published in 2024 to 2025. Fifty articles were randomly selected, and the models were prompted with either each article's abstract or its full text. Abstract-only prompts produced significantly higher retrieval accuracy than full-text prompts. Results varied significantly by platform, journal, prompt type, and output position, and overall accuracy was moderate and inconsistent.
In briefToward AI Trustworthiness: Analytically Proven Forward-Invariant Sets for AI-Controlled Systems
SummaryNeural-network controllers are widely used in nonlinear control but are hard to explain and verify. This paper proposes a framework that uses an Invertible Neural Network to map the original state space into a latent space where regular-shaped forward-invariant sets are more likely to exist. The network is trained so a hyper-rectangular candidate is invariant there and then formally verified, and the authors prove that successful verification yields analytically proven invariant sets in both spaces. Evaluated on 45 AI-controlled systems across three testbeds, the method certified invariant sets for all 45, while an adapted state-of-the-art baseline found none.
In briefAdversarial multiplayer bandit regret bound without collision feedback or shared randomness
SummaryThis paper studies adversarial multiplayer bandits with K arms and m labeled players, without collision information, shared randomness, or a communication channel. The authors design a communication and synchronization protocol built on a Monte Carlo public constructor. They prove a square-root regret bound of order K^{5/2}√T with high probability over preprocessing, holding for any oblivious reward sequence.
WhyAchieves square-root regret for adversarial multiplayer bandits without collision information or shared randomness, with rigorous probabilistic analysis of a communication protocol.
In briefGIVE-KWS: Noise-robust keyword spotting that injects lip-motion evidence into audio
SummaryThe paper studies how visual lip-motion cues can help detect keywords in noisy speech. It proposes GIVE-KWS, whose fusion module GIVE uses gated cross-attention to condition query audio on lip movements. At -10 dB, it reduces equal error rate for unseen keywords by 72.9% relative to the benchmark system, and by 62.8% on average.
Ming-Hsiang Hu, Kuan-Tang Huang, Hung-Shin Lee et al.
In briefDo Time-Series QA Systems Really Read the Data? Evidence Use and Reasoning Reliability
SummaryThis paper evaluates four time-series question answering systems (TimeOmni-1, ChatTS, TimeOmni-VL, and Time-MQA) and introduces COMMON-TSQA, a benchmark that unifies datasets and evaluation criteria. Systems are tested under the original condition and six interventions that change the input series while keeping the question fixed. The authors find that aggregate accuracy can hide how systems use numerical evidence, and some interventions trigger simple fallback behavior. They also find that rationales often include time-series claims unsupported by the input, and that agreement with the final answer does not guarantee correct numbers or valid inferences.
In briefSpending Memory on More Tokens at Lower Precision for Long Reasoning Chains
SummaryReasoning models produce a large KV cache while decoding long chains of thought, so it must be compressed online under a fixed memory budget. The paper proposes BreadthKV, which combines quantization with eviction to store more tokens at lower precision, choosing bit-width through end-to-end calibration. Across three reasoning models and four math and science benchmarks, it beats eviction alone in 17 of 18 settings and produces shorter outputs. On Qwen3-8B at the tightest budget, the share of AIME samples hitting the length cap fell from 91% with eviction to 40% with BreadthKV.
In briefTwo-Sample Testing via Path-based Inference: Using generative interpolation paths to test whether two datasets come from the same distribution
SummaryThis paper uses the generative dynamics learned by deep generative models to perform two-sample testing, deciding whether two finite datasets come from the same distribution. Using stochastic interpolants, both distributions are connected to a shared Gaussian bottleneck, and the null hypothesis is shown to hold exactly when the two halves of the path coincide, reflecting a symmetry about the bottleneck. On a synthetic benchmark and three image benchmarks, the proposed tests improve power over the strongest baseline by up to 33 percentage points at an equal total sample budget.
WhyUses stochastic interpolant paths for two-sample testing, proves null equivalence to path symmetry linked to Jeffreys divergence, and shows up to 33-point power gains on synthetic and image benchmarks
Eshant English, Wei-Cheng Lai, Yanfeng Yang et al.
In briefRESOLVE: Checking GPU Kernels with Tests and Proofs Instead of Tolerances
SummaryThis paper presents RESOLVE, a pipeline that combines testing and formal verification to validate GPU kernels. It uses binary instrumentation to expose race conditions by perturbing timing, then has an agent rewrite kernels into lower-concurrency versions that still produce bitwise-identical outputs. These simplified kernels are formally proven equivalent in F*/Pulse on real numbers, avoiding numeric tolerances. The authors apply it to KernelBench and fused GEMMs in CUTLASS, Triton, and Gluon, finding four previously unreported issues, two of them clear bugs.
WhyCombines testing, timing perturbation, concurrency reduction, and F* proofs for GPU kernel validation, finding two clear bugs with strong formal evidence.
Ashkan Vedadi Gargary, Guido Mart\'inez, Sebastian Burckhardt et al.
In briefTesting whether language models can apply traditional Chinese Bazi rules to cases
SummaryThis paper evaluates language models on 3,000 Chinese multiple-choice questions about traditional Bazi, split into 14 theory and 11 case categories. For every system tested, theory accuracy exceeds case accuracy by 16.60 to 29.56 points. Performance also varies sharply by category, with Luck Pillars averaging 84.62% while Career and Family Relations average about 37% to 38%. The authors argue for task-specific evaluation of cultural-domain applications rather than relying on aggregate scores. The benchmark measures agreement with a model-generated, model-verified answer key, not real-world predictive validity..
In briefReview of Sinhala Speech Recognition Methods, Challenges and Future Research
SummaryThis paper is the first critical review of automatic speech recognition research for Sinhala, a low-resource language. It traces the field from Hidden Markov Models through deep neural networks to self-supervised models such as wav2vec 2.0, XLS-R, Whisper and MMS. The authors compare Sinhala systems with related work on Tamil, Malayalam and Hindi and find that most reported word error rates are not directly comparable. The paper identifies six research gaps and outlines a research agenda.
In briefSharp Integrality Gaps in Calibration Distance: A Theory Paper on the Gap Between Calibration Distance and Its Relaxation
SummaryThe paper studies the offline gap between deterministic calibration distance C and its fractional relaxation L for binary unit-weight sequences under total absolute-change cost. It sharpens the known bound C ≤ L + O(√T) to a worst-case order of Θ(T^(1/3)), with explicit upper and lower bounds for T ≥ 216. It also gives a bound C ≤ L + m in terms of the number of distinct forecasts m, and a polynomial-bit-time procedure for rational inputs.
In briefEvolutionary origins of collective decision making in multi-layer voting groups, showing how individual rewards can produce accurate group solutions
SummaryThis paper develops a framework for how collective intelligence evolves in multi-layer voting populations, where individuals see limited information and groups aggregate opinions by majority rule across layers. The authors prove that single-layer voting cannot solve non-linear classification problems under any individual reward scheme. They identify a "marginal feedback" reward that pays individuals only when their opinion is pivotal in their group and every layer above, and show that peer imitation under this scheme yields accurate solutions equivalent to a multi-layer perceptron.
WhyProposes a marginal-feedback reward scheme for multi-layer voting, proving single-layer voting fails on non-linear tasks and matching a multi-layer perceptron. Novel and well-argued; impact mainly the
In briefBayesian data augmentation for retraining a vision-based UAV landing detector
SummaryThis paper addresses vision-based landing for drones in GPS-denied or cluttered urban settings. It proposes a Bayesian data augmentation framework that uses a photorealistic simulator with high-fidelity vehicle dynamics to iteratively retrain a helipad detection neural network, with landing performance as the objective. Experiments in the simulator under varied environmental conditions and vehicle states show improved landing performance and tighter confidence intervals on predicted outcomes.
In briefHow a prompt optimizer should spend a tight API budget: BudgetAPO with noise-adaptive evaluation
SummaryThis paper introduces BudgetAPO, a single-stage automatic prompt optimizer designed for settings where the number of model calls is tightly limited. It sizes each task's evaluation slice according to its measured noise from a short probe, turns accept/reject decisions into paired comparisons on a fixed slice, and rewrites reasoning strategy and output format jointly. Across seven benchmarks and five subject models, BudgetAPO ranks first on every subject, and at 250 calls it returns the seed prompt in 13% of runs versus 86% for GEPA.
In briefScalar Communication via Random Direction Refreshing for Distributed Optimization
SummaryThis paper addresses the high communication cost of distributed optimization when decision vectors are high-dimensional. It proposes a mechanism in which each neighbor message carries a single real number, the inner product of the agent's state with a shared random direction regenerated from a common seed. For an existing continuous-time algorithm with strongly convex, Lipschitz-gradient costs, the authors show that refreshing the direction at a sufficiently high rate yields exponential convergence with constant gains and no residual error.
In briefStageVLN: A Training Framework Using Spatial and Trajectory Guidance for Efficient Vision-Language Navigation
SummaryStageVLN is a training framework for Vision-and-Language Navigation (VLN) that uses spatial and trajectory guidance to shape navigation representations. A frozen geometry foundation model supplies spatial guidance, and relative-heading and expert-route progress objectives supply trajectory supervision. All auxiliary components are used only during training and removed at deployment, so inference keeps the original pathway. Reported results: 56.3% SR and 51.4% SPL on R2R-CE validation-unseen with a 4B-parameter backbone, and 54.3% SR on RxR-CE, both without an additional geometry encoder at inference.
In briefBellman-Centric Learning: Near-Optimal Regret for Linear Bandits with Memory
SummaryThis paper studies linear bandits where past actions affect future rewards through a known, bounded memory map. It proposes RSM-LinUCB, an algorithm that learns like linear bandits and plans like reinforcement learning. The authors prove a high-probability regret bound in which memory adds only an additive cost, and a matching minimax lower bound shows near-optimality. The results also extend to generalized linear rewards.
In briefGraph Data Augmentation via Contrastive Generator Inversion (DCBA)
SummaryThe paper introduces DCBA, a model-based method that infers the configuration of the ABCD synthetic graph generator from an observed network. It learns a joint representation of graphs and generator parameters using a multi-positive contrastive objective with soft negative weighting. Experiments show more accurate and robust parameter recovery than an algorithmic inverse-modelling baseline. Inferred configurations also improve community detection with PRoCD on average AMI.
Mateusz Stolarski, Micha{\l} Czuba, \L{}ukasz Krai\'{n}ski et al.
In briefTRIAGE: A direction-aware method to stabilize reinforcement learning with native NVFP4 precision
SummaryLow-precision computation can speed up reinforcement learning for large language models, but mismatches between the learner and sampler can make training unstable. The paper proposes TRIAGE, which diagnoses problems at the segment level, selectively rebalances policy-gradient updates, and applies bounded repair to remaining severe mismatches while keeping native NVFP4 W4A4 execution. Experiments on Qwen3-4B and Qwen3-30B-A3B show stable training, full-precision-level performance on five math reasoning benchmarks, and up to 2.3x higher rollout throughput than BF16.
In briefStudy of compute-optimal training and scaling for physiological waveform foundation models
SummaryThis paper trains Aether, a family of physiological waveform foundation models ranging from 20M to 2.1B parameters, on up to 36.3M hours of data. The authors build eight clinical prediction tasks from MIMIC-III and evaluate the models with linear probing. The 720M model outperforms existing baseline models on all eight tasks. The paper also proposes a scaling law linking model size, pretraining hours, and labeled patients to downstream ranking error.
In briefWhen a diffusion model commits to what it draws
SummaryDiffusion models generate images by removing noise step by step from pure noise. This paper measures, on CIFAR-10, when a trained model becomes committed to a class such as horse or truck, by freezing half-finished images and regenerating them many times to count class outcomes. Coarse decisions (vehicle or animal) settle at roughly twice the noise level of fine ones (which animal). A cheaper classifier-based proxy ranks these decisions in the right order, with rank correlation 0.73 to 0.88, but does not match their exact timing.
In briefSpacecraft rendezvous trajectory generation using composable diffusion models for flexible mission constraints
SummaryThis paper presents a diffusion-based method for generating spacecraft rendezvous and proximity operations trajectories with configurable mission constraints. Separate energy-based diffusion models are trained from optimized trajectories to satisfy individual constraints such as approach cone and sensor line-of-sight. At inference, these learned models can be combined with each other or with an analytic energy field to enforce different constraint combinations. Reported results show constraint satisfaction rates within 1 percentage point of single-constraint models or higher, without retraining for new combinations.
In briefCan CaMeLs Talk? Defending Multi-Agent Systems Against Hidden-Instruction Attacks
SummaryThis paper examines whether the CaMeL defense, which protects single AI agents from malicious instructions hidden in processed data, still holds when agents call other agents as tools. The authors show a working attack that succeeds even when every agent individually uses CaMeL. They then propose multi-CaMeL, a communication protocol that keeps trusted instructions and untrusted data separate across agents, reducing attack success to 0.0% on their MultiAgentDojo benchmark. The protocol carries a utility cost that shrinks as models become more capable.
James Peters-Gill, Avi Semler, Henning Bartsch et al.
In briefMultimodal machine learning combining genomic, epidemiological, wastewater, and clinical imaging data to detect measles and infer transmission
SummaryThis paper presents a multimodal machine learning framework that combines genomic sequences, epidemiological indicators, wastewater surveillance, and rash images to support measles detection and transmission inference. Using public data, the authors evaluated three stages: classifying transmission linkage between sequence pairs, predicting state-week outbreak risk, and classifying rash images with contrastive alignment of genomic and imaging embeddings. Genomic features mainly drove transmission-linkage prediction, while wastewater positivity was the strongest outbreak-risk predictor. The authors state that prospective validation is needed before deployment.
In briefGenAI4IDN Benchmark 3.0: A Public Tool for Evaluating Generative AI in Interactive Digital Narrative Design
SummaryThis paper presents the third version of GenAI4IDN, a framework for assessing generative AI tools that create interactive digital narratives. It adds AI-assisted evaluation through a public web application, letting the community run benchmarks, add models, and propose tasks. New fact-checking and bias-detection tasks were added, and results show the tools caught historical inaccuracies and structural biases, while limitations such as runaway thinking and overzealous safety filters were observed.
Hartmut Koenitz, Jonathan Barbara, Mirjam Palosaari Eladhari
In briefDiagnosing causal links in noisy systems via delay reconstruction of conditional moments
SummaryThe paper extends delay-coordinate reconstruction from deterministic to stochastic systems, proposing that delay vectors reconstruct conditional moments of a future distribution rather than a unique future trajectory. It supports this view using the Fokker-Planck equation and Koopman operator theory within the Mori-Zwanzig formalism. As an application, the authors revisit convergent cross mapping (CCM) and examine causal relationships in coupled logistic systems with additive and multiplicative coupling.
In briefAtomic Visual Entailment: Zero-shot image-text reasoning via fact splitting and learned selection
SummaryVisual entailment asks whether an image supports, contradicts, or leaves undecided a text hypothesis. The authors propose AVE, which splits the hypothesis into atomic facts, gets candidate predictions from frozen vision-language models for the full hypothesis and each fact, and uses a lightweight classifier to pick the final label. Without fine-tuning any vision-language model, it reaches 0.803 test accuracy on SNLI-VE.
In briefAutomated ArcGIS Pro tool for multi-source satellite land surface temperature analysis with LLM-based reporting
SummaryThis paper presents an ArcPy-based ArcGIS Pro tool that automatically extracts Land Surface Temperature from Landsat 7, 8, and 9 imagery. The tool automates radiometric calibration, NDVI-based emissivity correction, and thermal conversion, supporting both batch and single-scene processing. An optional Large Language Model helps interpret statistical results and generate reports. Tested on data from Ras Al Khaimah, UAE, the tool produced temperatures of roughly 25°C to 50°C, and the authors report it reduced processing time and human error.
In briefUndoBench: Measuring Whether Tool-Using AI Agents Can Recover From Failures
SummaryThis paper introduces UndoBench, a benchmark of 36 base workflows and 36 fault scenarios across 8 enterprise domains, which separately measures task completion and recovery from faults using paired trials. Nominal task completion reached 83.54%, while conditional recovery success was 46.72%, and naive retry caused duplicate external effects in 53.33% of trials. The authors argue that evaluating only nominal completion hides recovery weaknesses that depend on the phase of execution.
In briefFinding time-lagged cause-and-effect links in confounded time series with an adaptive, error-controlled method
SummaryThe paper introduces ORACLE-VARX, a pipeline for discovering which variables cause which others in multivariate time series and at what lags. It combines double/debiased machine learning to remove nonlinear confounder effects, adaptive causal lag estimation to choose lag orders over time, and z-tests with Benjamini–Hochberg correction to select directed edges at a target false discovery rate. On a synthetic benchmark and nine U.S. sector ETFs, the authors report better lag-order tracking and forecasting than the compared methods, with edge FDR close to PCMCI.
Hong Kiat Tan, Isaac-Neil Zanoria, James Chen et al.
In briefDREAM: Assigning Image Resolutions Dynamically in Multimodal Multi-Agent Debate
SummaryMulti-agent debate is a method that helps large language models reason better, and it is now being applied to tasks involving images. The authors note that existing approaches give every agent the same fixed image resolution and suffer from groupthink, where agents drop correct answers to follow confident but wrong peers. They propose DREAM, which runs a probe round to assign each agent its best resolution based on uncertainty, and uses uncertainty-guided rollback to restore early correct answers that were overridden. On six multimodal datasets, DREAM improves the accuracy-token trade-off by 1.5 to 3.2 percent over debate baselines without dataset-specific tuning.
Khanh-Binh Nguyen, Van Dai Do, Tien Anh Nguyen et al.
In briefSEA-LM: Spatial audio understanding for smart-glasses microphone arrays
SummaryThis paper presents SEA-LM, a model that understands spatial audio captured by wearable microphone arrays such as smart glasses. It uses FOACODER, a layout-flexible spatial audio encoder, and a multimodal large language model trained over six tasks, including sound localization and spatially selective transcription. On the authors' evaluation set, SEA-LM reports lower azimuth and elevation errors, higher temporal IoU, and lower WER on most transcription tasks than baselines, and stays robust across 1,211 array configurations with 4 to 9 microphones.
In briefKandinsky 6.0 Video: Foundation Models for Generating Synchronized Video and Audio
SummaryThis paper introduces Kandinsky 6.0 Video, a family of diffusion models that generate short video clips with synchronized audio from text or images. It offers a 3B-parameter Lite version and a 29B-parameter Pro version, producing 5-second clips with 44 kHz audio and lip-sync, upscaled to Full-HD. In side-by-side human evaluation, the Pro model clearly beats its predecessor and is competitive with leading audio-video models, especially in speech quality. Code, checkpoints, and diffusers integration are released under the MIT license.
Team Kandinsky, Julia Agafonova, Bulat Akhmatov et al.
In briefThe premise is the problem: when self-monitored adaptation's guarantees fail
SummaryThis paper examines forecasting models that are updated after deployment, and whether a monitor that uses the same prediction errors as the updates keeps its statistical guarantee. The authors study multi-step time-series forecasting and show that overlapping targets and error dependence can break a key assumption. As a result, the monitor may raise false alarms, and its protective responses can worsen prediction quality.
In briefEchoDino: A pediatric foundation model for heart ultrasound analysis across all ages
SummaryThis paper presents EchoDino, a self-supervised foundation model for echocardiography built by adapting the DINOv3 framework to 3.7 million frames from 1.7 million unlabeled pediatric heart ultrasound videos. It also introduces MEMS, a sampling method for selecting the most informative frames in videos. Across nine pediatric and adult datasets, EchoDino outperformed baseline models on view classification, structural heart disease detection, age estimation, segmentation, and ejection-fraction error, and it generalizes from pediatric to adult echocardiography.
Sheng Cheng, Donnchadh M. O'Sullivan, Daniel J. Penny et al.
In briefInference-time projection to make biomolecular diffusion model outputs physically valid
SummaryThe paper addresses physically invalid outputs from AlphaFold 3-style cofolding models, such as overlapping chains, distorted ligand bonds and angles, non-planar rings, and inverted stereocentres. The authors apply two closed-form projection operators to the diffusion model's denoised coordinate estimate at inference time: one for inter-chain van der Waals clashes and one for ligand distance geometry. The method needs no extra network evaluations or retraining. Applied to Boltz-2 and OpenFold-3 across five benchmarks, it recovers perfect physical validity while preserving structural accuracy and ligand-placement metrics, with negligible runtime and memory overhead.</summary>
Qurat-ul-ain, Yee Whye Teh, Charlotte M. Deane et al.
In briefYour Unlearning Gives You Away: Finding Which Concepts Diffusion Models Erased
SummaryExisting attacks on unlearned diffusion models assume the erased concepts are already known and focus on recovering them. This paper introduces Tracer, a framework that identifies which concepts were erased and estimates how many, without generating or classifying images. Experiments on text-to-image and text-to-video models show it identifies erased concepts in seconds with higher accuracy than prior methods.
In briefAn LLM-in-the-loop reinforcement learning method for selecting features in bioinformatics data
SummaryThe paper addresses high-dimensional bioinformatics data, where features greatly outnumber samples and cause overfitting and high computational cost. It proposes a reinforcement learning framework in which an agent selects features sequentially, while a large language model provides domain-informed exploration guidance and hybrid rewards combining data performance with knowledge-based evaluation. Experiments on diverse bioinformatics datasets reportedly show better results than baselines, stable performance across downstream models, and faster convergence than pure reinforcement learning.
In briefGaussian limits for SGD even when stationary moments fail to exist
SummaryThe paper studies Gaussian approximations for stochastic gradient descent (SGD). It builds a dependent design with standard Gaussian marginals whose stationary error has infinite positive moments, while independent observations with the same marginals give finite stationary variance. In both cases the small-step limit remains Gaussian.
WhyConstructs a design with Gaussian marginals whose SGD stationary error has infinite positive moments, yielding Gaussian and OU limits with explicit corrections.
In briefAdding New Sites to Federated Learning Without Labels or Old Data
SummaryThis paper addresses a site joining a cross-silo federated learning system after training, where its records are unlabeled and mix known and unseen conditions. It proposes an autonomous three-stage procedure run entirely at the joining site: reconstruction experts flag novelty, clustering groups flagged records into candidate conditions, and class means represent old classes. On a real industrial condition-monitoring dataset, it keeps old-class accuracy at 0.868 or above with forgetting at most 0.063, balancing old- and new-class accuracy.
In briefBeyond Words: Compositional Tokenization for More Efficient Language Models
SummaryThis paper introduces CoBPE, a compositional tokenization method that represents short phrases as a base token plus a small set of reusable modifiers, combined in embedding space at input and predicted jointly at output. Pretrained from scratch at 780M and 1.3B scales under matched training compute, CoBPE shortens sequences by 30%. It also improves average downstream performance by 1.2 points over standard BPE.
In briefMoment-accurate Gaussian mixtures for constant-step stochastic approximation
SummaryThe paper builds Gaussian mixture models for constant-step learning that match stationary energy with local Ornstein-Uhlenbeck limits, so they predict output variability and expected losses more accurately than weak convergence alone. For step size a, it proves a second-order Wasserstein error of o(√a) and derives observable covariances, expected objective gaps and first-order mean shifts. Numerical transport calculations and controlled SGD studies are used to evaluate these predictions.
In briefHow Much Is a Repeated Token Worth? Scaling Patterns in Multi-Epoch Pretraining
SummaryThis paper studies how much value repeated data adds when a language model is pretrained for multiple epochs, comparing it with one epoch on the same data and with fresh data at equal compute. It finds that repeated tokens lose half their value after a critical epoch count, which grows with training budget per parameter but barely with model size. The authors also report that, at the same epoch count, how data is replayed, how concentrated the repeats are, and the entropy of the source all affect loss.
WhyPrices repeated tokens against fresh data and single-epoch baselines, unifying critical epoch counts and size trends across several model scales with fairly broad experiments.
In briefColdDDI: Testing Whether Cold-Start Drug Interaction Models Use Mechanistic Evidence
SummaryThis paper introduces ColdDDI, a diagnostic benchmark built from DrugBank 5.1.13 with 1,900 approved small-molecule drugs and 565,731 positive drug-drug interaction pairs. It checks whether models that receive molecular, text, or knowledge-graph evidence actually use the evidence that supports an interaction, especially for drugs unseen in training. In the hardest setting, where both drugs are unseen, the main performance gap comes from whether a shared enzyme, transporter, or target is available as a mediator.
WhyProposes a diagnostic benchmark probing whether cold-start DDI models use mechanistic evidence, with broad comparisons and masking/replacement interventions across 8 methods and 13 LLMs.
In briefWildIcon: Generating wildlife videos that keep each animal's identity consistent
SummaryWildlife individual identification suffers from scarce images of the same animal in varied poses and views. The paper proposes WildIcon, an image-to-video framework that extracts high-frequency identity cues, such as fur texture and spot patterns, from a reference image and injects them into a frozen base model via lightweight adaptation. The authors also release WildlifeVid, a wildlife video dataset with individual labels. Experiments report stronger individual consistency than baselines, and filtered outputs can serve as useful training augmentations for animal re-identification.
In briefExpanding LLM Reasoning: Choosing Where to Restart Existing Reasoning Chains
SummaryThe paper studies where to begin an additional continuation inside an existing reasoning chain instead of sampling new chains. It defines expansion utility as the change in correctness from restarting at a stored step, measured across nine models and six benchmarks. Restart position matters, but a fixed always-last rule is a strong baseline that the learned router does not clearly beat.
In briefRethinking how streaming perception is evaluated on shared edge hardware
SummaryThis paper studies how to choose which accelerator (GPU or NPU) should run multi-camera streaming perception when the device is shared with other workloads. The authors test two end-to-end pipelines on one GPU-NPU platform and find that isolated single-stream tests can rank placements incorrectly, since GPU contention causes deadline misses and stale detections. They argue evaluations should report contention sweeps, deadline-miss rates on both paths, and worst-stream sAP alongside mean sAP.
In briefWhat an observability forecasting model encodes internally: a probe-based audit of Toto
SummaryThe paper audits Toto, an observability forecasting foundation model, on the BOOM benchmark using linear probes on its frozen residual stream. Toto's residuals more linearly reveal short-vs-medium cadence and metric type than the strongest raw-window baseline. However, the authors find that recoverable information is not necessarily used, since swapping residuals did not consistently change forecast burstiness.
Dhyey Dharmendrakumar Mavani, Rian Atri, Tairan Ji
In briefSteadySplats: Low-Variance Gaussian Resampling for Cleaner Stochastic Rendering of 3D Scenes
SummaryThe paper addresses visible noise in stochastic order-independent transparency rendering of 3D Gaussian Splatting models. It combines history-based spatial resampling, temporal importance resampling, and a color regularizer applied during training to reduce noise at the representation and image synthesis levels. The authors report about a 13 dB PSNR gain over previous stochastic methods at 1 sample per pixel, and convergence to sorted 3DGS with an average L1 error below 10^-4.
Felix Windisch, Thomas K\"ohler, Lukas Radl et al.
In briefCan Prosodic Style Be Inferred from Text Alone? Unsupervised Acoustic Clusters Test
SummaryThe paper tests whether written text alone carries enough information to select an appropriate prosodic style for speech synthesis. Using a 1,200-hour conversational corpus from six speakers, utterances are clustered by acoustic features and held-out clusters are predicted from text embeddings. Text beats the majority-class baseline, but most of the gain comes from word identity rather than sentence-level semantics.
WhyTests whether text predicts acoustic prosody clusters with length and majority-class controls, yielding a falsifiable, well-controlled finding relevant to expressive TTS.
In briefBetter Retrieval, Limited Clustering Gains: Multilingual Company Matching Study
SummaryThis paper studies how adapting multilingual encoders affects company name retrieval and final entity clustering differently. The authors use a controlled pipeline with multilingual E5 encoders, country-level edge budgets, unchanged pair classifiers, and guarded Union-Find clustering. On 3,633 names, direct pair recall rose from 53.74% to 76.98%, but cluster recall only moved from 32.54% to 33.33%, as most newly retrieved positives failed the existing classifier threshold.
In briefSeparating Task Difficulty from Single-Run Failure in LLM Agent Failure Prediction
SummaryThis paper examines whether predictors of LLM agent failure truly detect failing runs or simply recognize harder tasks. The authors find that most pooled AUROC comes from cross-task comparisons, and that early within-task discrimination between successful and failed runs is weak. Early stopping only becomes beneficial when within-task discrimination is much higher. The authors argue that failure prediction should be judged by whether its signal supports the intended deployment decision.
Mohsen EsfandyariDoulabi, Lawrence Arkoh, Biruk Tadesse et al.
In briefScenario-based compositional statistical model checking for safety rules
SummaryThis paper targets safety-critical fields such as autonomous driving and proposes a scenario-based compositional statistical model checking framework. It splits composite scenarios into primitive scenarios and specifications into sub-specifications, verifies each primitive independently, and combines the statistical estimates using importance sampling and kernel density estimation. The authors report that the framework accurately answers verification queries for unseen composite scenarios while reducing simulation cost through parallelization and trace reuse.
Abhinav Pomalapally, Arya Raeesi, Kevin Kai-Chun Chang et al.
In briefPoisson-GENERIC Neural Operators: Exact Metriplectic Structure via Casimir Entropies
SummaryThe paper proposes a thermodynamically consistent neural operator that uses a compatible Lie-Poisson pencil for nonlinear transport and treats the integral of latent entropy density as a Casimir, enforcing GENERIC degeneracy without projection. Across four 1D and 2D PDEs and three backbones, the model wins 61 of 72 seed-level comparisons against matched unconstrained baselines. Physical energy is bounded under coercivity assumptions.
WhyReplaces projection with a Lie-Poisson pencil and Casimir entropy for exact structure; experiments span four PDEs and three backbones, though scale and baselines are limited.
In briefFactoriax: A GPU-accelerated factory-building simulator for reinforcement learning
SummaryThis paper presents Factoriax, a factory-building simulator written in JAX that runs on GPUs. Agents must gather resources, craft machines, and place them to form automated production pipelines. The authors introduce the Easy Rocket benchmark, where agents must build a rocket within a limited number of game ticks, and report PPO training results. A 1-billion-step run takes about 8 minutes on an NVIDIA A100, but the trained agent fails to arrange machines in a functional layout and does not complete the benchmark.
Mickey Beurskens, Tristan Tomilin, Thiago D. Sim\~ao
In briefLifeLong Digital Twin: A Unified Framework for Modeling Lifelong Health Trajectories
SummaryThis paper introduces LifeLong Digital Twin, an event-driven method that organizes life events into daily health states and accumulates them into a lifelong health context. An accompanying Agent Harness incorporates multimodal evidence beyond a language model's textual context. Four language models were tested across 25 disease endpoints on three tasks: disease trajectory forecasting, disease risk ranking, and multi-horizon disease prediction. Compared with the reference condition, the approach yielded clear gains, with thyroid-disease F1 reaching 0.669.
In briefBorrowing an LLM's attention to make text embeddings follow instructions
SummaryThe paper proposes Attention Relay, which passes the attention weights of an instruction-tuned LLM to a text embedding model, letting the embedder become instruction-aware without any training. Tested across six LLMs and ten embedding models, it makes nearly every combination instruction-aware. The authors find these attention weights largely come from instruction tuning and emphasize the parts of the text the instruction asks about.
WhyProposes training-free attention relay transferring LLM instruction-following to embedders, validated across ten embedders and six LLMs with ablations.
In briefG-CARB: Graph-Localized Conformal Risk Budgets for Multi-Step AI Agent Harm
SummaryThe paper introduces CARB, a method that decides when to stop a small language model agent based on a ledger of harm incurred before stopping. G-CARB selects scoring evidence by following observable dependencies from private data sources to outgoing actions, while the ledger still covers the full execution history. On AgentDojo replay with two 14B backbones, it roughly halves the scorer-input records at intermediate risk budgets and improves autonomous task completion compared with full-prefix scoring.
In briefCutBCE: An efficient large-vocabulary loss kernel for sequential recommendation
SummarySequential recommenders with very large item catalogs struggle because standard binary cross-entropy materializes full logits tensors, causing memory overflows. This paper introduces CutBCE, an exact, hardware-accelerated BCE loss and gradient operator implemented in JAX and Pallas that keeps logits and gradients on-chip. On TPU experiments, it eliminates out-of-memory errors and, on an 876k-item SASRec training task across 8 chips, reduces peak HBM by 65.7% and increases training speed by 225.9% with comparable accuracy.
WhyFuses large-vocabulary BCE loss and gradients on-chip to avoid dense HBM logits, with 8-chip TPU experiments showing large memory and speed gains.
In briefAn equality condition for the Dobrushin bound on attention rollout, and how often it holds in trained transformers
SummaryThe paper derives an exact condition under which the layerwise Dobrushin bound for attention rollout is tight: some token pair achieving κ(A) must attend to itself at least as strongly as to each other. The authors show this condition is far from automatic for random stochastic matrices. Testing on trained vision, language and tabular transformers, they find it holds for nearly all inputs on content tokens in several models, but much less often in supervised and tabular models.
In briefMarkerless video method for assessing walking in spinal cord injury patients
SummaryThe paper presents a monocular, markerless approach that estimates lower-body kinematics and ground-reaction forces from a single sagittal-view video of people with spinal cord injury, without markers or force plates. Using the SCAI SCI Gait dataset of 239 adults, the video-based estimates agreed closely with motion-capture and force-plate measurements. The authors also report that gait waveform components were conditionally associated with functional independence and that speed-matched individuals showed distinct mechanical strategies.
WhyExtends monocular markerless gait analysis to spinal cord injury, validated on 239 patients with synchronized kinematics and force-plate data; useful for the subfield.
In briefWhen Low Prediction Error Misleads Planning: Diagnosing Latent World Model Failures
SummaryThe paper shows that the component dominating a latent world model's prediction error is not necessarily the one whose repair best improves action selection. It splits endpoint error into a candidate-pool center and action-relative responses, then compares oracle corrections across four model families and four tasks. Center error dominates MSE in most cells, yet correcting action-relative responses often gives better physical rank correlation and elite quality, motivating separate evaluation of representation, prediction, and planning.
In briefEstimating shared room decay rates and spatial amplitudes with parameterized matrix factorization
SummaryThe paper models room impulse responses as a parameterized nonnegative matrix factorization using the Itakura–Saito divergence, jointly estimating common decay rates and amplitudes. Estimation at each frequency bin yields detailed reverberation time curves directly from impulse response powers, without backward integration. The authors introduce a contribution-weighted SAGE algorithm to speed up estimation. Experiments show accurate recovery of well-separated decays on synthetic data and application to measured coupled-room impulse responses reveals frequency-dependent reverberation times and space-time contributions of shared decay components.
Jeremy B. Bai, Filip Elvander, Sebastian J. Schlecht
In briefMore Claims, Less Evidence: Bounded Verification of AI-Generated Knowledge Artifacts
SummaryThe paper proposes a Bayesian model of bounded verification centered on the Predictive Value of Pass (PVP). Under fixed verification capacity, it predicts that passing a check becomes less informative as artifacts grow, improves with larger budgets, and is most fragile when errors are sparse. Controlled experiments on FEVEROUS and FEVER support these predictions and yield the minimum verification budget needed to maintain a target PVP.
In briefSoft Strategy Selection for Batch-Mode Active Learning
SummaryThis paper addresses the difficulty of choosing an active learning acquisition strategy before labeling, since strategy performance varies across settings and existing methods spend rounds exploring. It proposes soft strategy selection, which uses multiple strategies within a single batch, and introduces FractAL. FractAL estimates per-strategy rewards via influence-function-based data attribution without retraining and allocates budget using online mirror descent. Across 7 setups spanning classification, regression, and genetic perturbation prediction, the authors report that FractAL matches or outperforms all baselines, including random sampling.
In briefSALUS: Automatically Auditing Natural-Language-to-SQL Benchmarks Using Weak Supervision From Multiple AI Agents
SummaryWidely used natural-language-to-SQL benchmarks contain annotation errors that distort evaluation. SALUS detects these errors by treating SQL generated by multiple LLM agents as noisy weak labels, combined through a label model and a learned trustworthiness estimator, without human ground truth. On 298 manually verified BIRD tasks (BIRD-Clean-xs) it achieves F1 = 0.9194. The authors estimate annotation error rates of about 37% for BIRD and 27% for Spider.
Shiyuan Zhou, Ashwin Gerard Colaco, Sainyam Galhotra et al.
In briefLiFT: Loop Flow Transformers, looped generative models that scale compute by reusing one shared core
SummaryThe paper introduces Loop Flow Transformers (LiFT), generative models that repeatedly apply a shared Diffusion Transformer core with only light architectural changes. Each loop step is trained to regress a point on a straight path from the initial estimate to the flow-matching target. A trained model can loop beyond its training depth without retraining. On ImageNet 256x256, LiFT-L/2 achieves an FID 3.34 lower than the dense DiT-XL/2 baseline with about 60% fewer parameters, 32% fewer training FLOPs, and 52% fewer inference FLOPs.
Mohammad Mahdi Derakhshani, Pedro M. P. Curvo, Gertjan J. Burghouts et al.
In briefJIVEAdapter: A multi-task adapter that separates shared and task-specific updates
SummaryThis paper introduces JIVEAdapter, a parameter-efficient fine-tuning method for multiple tasks. It splits each weight update into a Joint part shared across all tasks and an Individual part specific to each task, encouraging the two to be nearly orthogonal. The Joint part can be learned once or incrementally, then frozen and reused for new tasks. On GLUE and SuperGLUE with DeBERTaV3-base, it is reported to be competitive with strong single-task and multi-task low-rank baselines.
In briefDataset Signatures in Human-LLM Interactions and User Modeling
SummaryThis paper compares seven human-LLM conversation datasets, including in-the-wild chat logs and human preference data. Neural classifiers can identify a conversation's source dataset from user messages alone well above chance, revealing distinctive dataset signatures that persist even after matching on human-designed taxonomy dimensions. The authors also examine how these signatures affect user models trained on the data, their evaluation, and data selection for training.
In briefWhat a Harness Repairs: Visibility, Baselines, and Evaluation Defects
SummaryThis preregistered study examines where score gains from changing prompts, token budgets, or parsers around a frozen language model come from. The authors ran two studies with models from 1.7B to 8B parameters plus Qwen3-14B, reporting all 65 primary endpoints. Turning thinking off usually improved accuracy, mostly by making answers readable that the capped setting had left unparsed. For 14B, some high-scoring harnesses were not detectably above the strongest constant-label baseline.
In briefDelegationBench: Testing When AI Agents Should Ask Before Acting
SummaryThis paper introduces DelegationBench, a benchmark of 156 scenarios where AI agents choose among acting, asking permission, asking for missing information, or refusing. The authors find that agreement scores are misleading: a simple keyword rule they wrote agrees with annotators more often than eight of ten models, rephrasing the same question shifts how often models act by up to 52.5 points, and models ask users less when executing tasks with tools. They recommend reporting these properties separately rather than as one combined score.
WhyProposes a new benchmark for when agents should ask first, using matched pairs and ten models to expose evaluation distortions; evidence is solid and useful for the subfield.
In briefDeep Prior Learning for Embodied Perception: Using Camera Pose and Other Priors to Improve 3D Geometric Perception
SummaryThe paper introduces VPGGT, a VGGT-based framework that extends OmniVGGT to use camera poses, intrinsics, and depth as priors for embodied 3D perception. It simulates sensor-motivated pose noise during training and adds a parameter-free prior residual connection (PRC) to address prior dilution, where predictions are less accurate than the supplied poses. It also proposes Metric Global Attention to predict a shared metric scale. Experiments on four datasets show that PRC improves translation-direction accuracy and joint pose AUC when camera priors are given for all views, under both exact and corrupted poses.]
In briefDynaMesh: Generating Time-Varying Textures for 3D Meshes
SummaryDynaMesh is a method that generates evolving surface textures for 3D meshes from a textureless shape and a text prompt describing an effect. The geometry stays fixed while the appearance changes over time. It conditions a video model to create a reference video, then processes it with a frozen image-to-3D generator, blending conditioning over time windows and using per-shape low-rank adapters to reduce flicker and recover details. The authors report substantial improvements over recent video-to-4D and texturing methods.
In briefThis paper explores how wearable assistants can remember where objects were seen by linking what users say to 3D spatial evidence.
SummaryThe paper defines Spatially grounded Conversational Reasoning (SpaCR), covering cross-scene, recall-oriented, and counterfactual spatial queries. It proposes SpaC-MEM, an object-centric working memory that uses 3D reconstruction and segmentation to tie dialogue information to persistent physical objects. The authors also introduce Ego-SpaCR, a benchmark of 620 ScanNet video sessions with 95 conversations and 3,100 queries. Reported results show SpaC-MEM achieves the highest overall answer accuracy and needs substantially fewer input tokens than native-video baselines.
In briefWhat Does Fréchet Distance Measure? A Directional Decomposition for Interpreting Generative Model Scores
SummaryThe Fréchet distance, used as FID for images and FVD for videos, summarizes the gap between generated and reference distributions in one scalar. This paper introduces directional Fréchet distance, which projects optimal transport displacement onto a chosen direction to reveal where the discrepancy lies. Across image, video, and protein case studies, a small number of interpretable directions account for much of the distance.
In briefDistributed algorithms for computing tight α-potential approximations in general-sum games
SummaryThis paper studies how to find the tightest α-potential approximation of a general-sum game with continuous actions, within a chosen class of potential functions, when each player only knows its own utility. The authors split the problem into a global outer search over deviation tuples and distributed convex inner problems, and develop a primal-dual inner method and a projected zeroth-order outer method. Numerical experiments show the trade-off between accuracy and computation and that the optimization approach can improve on analytical α-potential constructions.
In briefRobust 2D Traversability Mapping for Construction Robots Fusing LiDAR and Camera Semantics
SummaryThe paper addresses navigation difficulties for autonomous mobile robots on construction sites, where LiDAR alone can miss hazards like wet mud and can misread speed-breakers and inclines as obstacles. It proposes a real-time pipeline on an NVIDIA Jetson AGX Orin that uses LiDAR as the main geometric estimate and monocular semantics as a selective corrective signal. The authors also release a multimodal dataset from two active sites with 506 annotated frames across 28 classes.
In briefFunctional structure of recovery after compressing LLMs with low-rank methods
SummaryDifferent low-rank compression methods yield compressed LLMs that respond differently to the same recovery procedure, so their advantages at the compression endpoint may shrink, grow, or reverse. The paper formalizes this heterogeneity as a reproducible compression-induced functional structure called recovery pressure, with a standardized characterization within each backbone. It finds that endpoint recovery pressure relates to later recovery response, its module-wise structure reorganizes non-uniformly during recovery, and tracking it offers a complementary view of progress beyond scalar loss.
Zishan Shao, Liang Tian, Georgiy Zemlevskiy et al.
In briefA lightweight EEG model that transfers across neurological disorders with frozen weights
SummaryThis paper introduces USE-FM, an EEG foundation model with only 1.46 million parameters, pretrained by self-supervised signal reconstruction on TUEG. After freezing the encoder, it is evaluated on abnormal EEG detection (TUAB) and epileptic seizure recognition (TUEP) against models such as LUNA-Base and CBraMod. The authors report strong sensitivity and F1-score on TUEP, competitive AUC on TUAB, and organized latent representations in clustering and visualization analyses.
In briefLogic-Logit: A rule-based way to model and explain people's choices
SummaryThis paper introduces Logic-Logit, an interpretable choice model that explains human decisions using logic rules. It combines column generation and the Frank-Wolfe algorithm to extract these rules efficiently. Experiments on synthetic data and real commercial and healthcare datasets show it outperforms baseline models in interpretability and accuracy.
In briefLEON: Learning Location Embeddings from OpenStreetMap Neighborhoods with Hexagonal Graph Masked Autoencoders
SummaryThis paper presents LEON, a self-supervised framework that adapts Graph Masked Autoencoders for learning geographic region representations. It builds hexagonal grid graphs using H3 indexing and learns spatial embeddings from OpenStreetMap amenity distributions. LEON was evaluated on EuroSAT satellite image classification and geographic prediction tasks, with accuracy on EuroSAT improving by up to 1.87%.
Szymon Soltysiak, Radoslaw Malek, Jedrzej Kusnierz et al.
In briefWhen Does Retrieval Help? In-Context Adaptation in Vision-Language-Action Models
SummaryThis paper studies how retrieval methods affect in-context adaptation in vision-language-action (VLA) robot models, which retrieve expert demonstrations at test time within the RICL framework. It compares four methods: image-based retrieval, retrieval with VLA state, retrieval using VLA backbone features, and random retrieval. No method consistently wins on task success, random retrieval achieves a non-trivial success rate, and standard retrieval-quality diagnostics do not reliably predict downstream performance.
In briefRobust Surgical Robot Instrument Tracking Using Multi-Cue Fusion and Self-Training
SummaryThis paper addresses the difficulty of tracking surgical robotic instruments during minimally invasive surgery. It proposes a tracker-guided self-training framework that adapts a model pretrained on synthetic images to unlabeled real-world videos, using an uncertainty-aware EKF and an RTS smoother to refine instrument poses and generate pseudo-labels. Experiments on real videos show consistent gains across all keypoint metrics, with the model outperforming prior approaches in accuracy and runtime.
In briefSkillGATE: Gate-Aware Tree Search for Finding Relevant Skills
SummaryThis paper addresses Skill Retrieval, the task of finding the most relevant skills in large external skill libraries. The authors propose SkillGATE, a graph-guided hierarchical framework that uses Gate-Aware Monte Carlo Tree Search to adaptively navigate regions and select skills. Experiments on six benchmarks show consistent gains across various retrieval and reranking backbones, with a reported 16.3% improvement in overall R@1 over the strongest retriever-based baseline.
In briefWhy Softpick Fails at Initialization: A Normalization Analysis
SummaryThis paper examines why Softpick, an attention operator that rectifies scores instead of using softmax, loses its advantage at larger scale. The authors attribute the failure to its normalization denominator, which they split into positive and negative sums and treat as independent choices. They report that the failure begins at initialization, where some rows have a zero positive denominator and gradient norms exceed 10^12. Only Softpick and a stop-gradient variant train from scratch, and at 230M parameters the stop-gradient variant performs better on quantization and passkey retrieval.
