In briefMesh2Shell Neural Operator for Inverse Shape Optimization
SummaryThis paper introduces the Mesh-to-Shell Neural Operator (M2S-NO), a component of a category-agnostic framework for inverse shape optimization. It separates the explicit mesh used for design from the volumetric representation used for physics prediction, mapping surface samples to a shell representation that a separate PDE neural operator uses to predict physical fields. Objective gradients update mesh vertices directly, without a latent shape decoder. Evaluated on geometric matching, jet-engine brackets, and neutron moderators, the method reduces maximum bracket displacement by 14.3% and increases neutron-sensor current by 13.8% relative to the best training designs.
In briefSemantic Interoperability in Federated Digital Twins: A Systematic Review
SummaryThis PRISMA 2020 systematic review screened 5,833 records and read 119 full texts to assess whether published systems detect and repair broken data mappings between digital twins while they operate. Across the 32 included studies, mappings are fixed before operation, and the two ingredients needed for run-time repair appear only separately. The authors derive six requirements that a system claiming run-time semantic interoperability between independently owned twins should meet.
In briefConditional calibration contracts under coupled load: a branch-aware audit of switched inference
SummaryThe paper treats conditional calibration contracts for model-serving routers under bursty load as specifications to be validated, not certified properties. It introduces CA-MHD, combining load-coupled distillation, closed-loop replay and a cost-informed Schmitt router, evaluated across 40 cells in a discrete-event simulator. The deployed wide branch has lower ordinal accuracy than the original teacher in all 40 cells, while student-only serving shows a lower historical ordinal gap in 27 of 40 cells but routing produces violation tails up to 17.5 percent.
In briefChecking concept-level audits of melanoma classifiers using generated counterfactuals and control cases
SummaryThis paper tests concept-based explanations of melanoma classifiers by building a counterfactual audit from a diffusion model conditioned on seven 7-point checklist concepts. Colour-calibrated initial noise greatly reduced colour outliers, and several concept knobs changed other concepts as well. The audit failed its pre-specified rule, but matched positive controls and permuted-label null classifiers are presented as practical checks for concept audits.
In briefCNN-Based Retinal OCT Classification and a From-Scratch Baseline (ED126-A)
SummaryThe paper reviews convolutional neural network approaches to classifying retinal OCT images and discusses methodological issues affecting fair benchmark comparison. It introduces ED126-A, a four-class CNN trained from random initialization on the Kermany labeled OCT dataset. On the held-out image-level test set, the model reached 95.03% accuracy and a macro-F1 of 0.9205, with lower precision for the Drusen class (0.7493).
In briefOrder-Only On-Policy Distillation (O³PD): training a student model to match only the teacher's token rankings
SummaryThe paper proposes O³PD, an on-policy distillation method in which the student learns to reproduce the order of the teacher's top-K tokens without fitting their probability values. The objective is invariant to teacher temperature, so no temperature tuning is needed. Across six teacher–student pairs, O³PD outperforms OPD on the average over HMMT25, AIME24, and AIME25 in every pair, by up to 2.9 points, with gains of up to 5.6 points on AIME25.
In briefConsolidating Multi-Agent Reasoning via Test-Time Training
SummaryThe paper proposes TTT-Consolidate, which treats reasoning trajectories from multiple agents as adaptation data. Through test-time training, these trajectories are consolidated into a lightweight LoRA adapter, and a new solution is generated conditioned on that information. The method meta-learns consolidation via bi-level optimization and is built on the newly constructed OpenParallelThinking corpus of 12,810 math problems. On math and STEM benchmarks it outperforms the best aggregation baselines by up to 20 percentage points.
WhyThe authors propose consolidating multi-agent reasoning trajectories into a LoRA adapter via test-time training and bi-level meta-learning. The abstract reports up to 20-point gains over aggregation b
In briefNorthCUA Technical Report: Boosting Computer-Use Agents Through Large-Scale Experience
SummaryThis paper presents NorthCUA, an integrated system that improves computer-use agents by learning from experience at scale. It combines an Experience Flywheel that generates and verifies task trajectories, staged training, and joint desktop-web reinforcement learning. The authors report that their Nano, Mini, and Pro models achieve state-of-the-art results among open-source models on GUI grounding, desktop execution, and Web interaction under the stated evaluation settings.
In briefDelay and drift combined: robot foundation model safety spans multiple system layers
SummaryThe paper argues that the physical safety of robot foundation models in closed-loop operation depends on how environmental changes, the model, the computing platform, and execution interact, not on the model alone. The authors formalize cross-layer variation propagation and run 4,000 paired closed-loop episodes of the pi0.5 vision-language-action model on LIBERO. A 400 ms inference delay raised the safety-violation rate by 3.5 points under the training viewpoint but by 37 points under a 60-degree camera shift, 33.5 points more than the two effects combined.
In briefEarthJev: Calibrated, non-generative answers to questions about Sentinel-2 land cover
SummaryThis paper presents EarthJev, a 0.8B-parameter model that answers multiple-choice questions about multispectral Sentinel-2 satellite images without generating text, reading the answer directly from the option-letter logits. Trained on 70,890 questions from BigEarthNet v2.0, it reaches 86.0% accuracy on 36,028 test questions. It answers faster than multimodal language models and provides calibrated probabilities.
Domenico Benfenati, Antonio Maria Rinaldi, Cristian Tommasino
In briefBernoulli On-Policy Distillation: Keeping Learning Signals from High-Confidence Student Tokens
SummaryOn-policy distillation trains a student language model on its own generated trajectories using token-level supervision from a stronger teacher. The paper shows that standard sampled reverse-KL distillation suppresses updates to high-confidence tokens, even though their weak gradients do not necessarily mean the teacher's supervision is weak. The authors propose Bernoulli On-Policy Distillation, which directly matches teacher and student probabilities for each token, preserving teacher-student disagreement without the confidence-dependent suppression.
In briefHQNN-CL: A Hybrid Quantum Network with Continual Learning for Market-Regime-Aware Financial Forecasting
SummaryThis paper proposes HQNN-CL, a hybrid quantum-classical neural network for financial time-series forecasting that must adapt to changing market conditions. The framework combines Gaussian HMM-based regime detection, BiLSTM encoding, adaptive-depth quantum circuits, and continual learning methods such as Elastic Weight Consolidation and experience replay, producing probabilistic forecasts. It was tested on four stock datasets (Microsoft, Systems Limited, Teradata, Tata Consultancy Services) against seven state-of-the-art models, and it outperformed the others on the Systems Limited and Teradata datasets.
In briefLocal repair method for multi-robot warehouse pickup and delivery plans
SummaryThis paper proposes Local Solution Repair (LSR), a method for improving plans used by multiple robots in automated warehouses that must pick up and deliver items without collisions. LSR detects suboptimal robot behaviors at run time and locally adjusts plans to reduce conflicts between agents. The authors apply it to repair solutions from two state-of-the-art baselines, M2M and LNS-PBS. In long-horizon simulations across three environments, it improves task throughput by up to 5.9%, reduces suboptimal instances by about 36k or 14.6%, with minor added computation time.
Ethan Schneider, Jingkai Chen, Seth Hutchinson et al.
In briefReduced Cost Influence Functions for predict-then-optimize with noisy training data
SummaryThis paper studies a predict-then-optimize setting where an upstream model predicts the cost vector of a downstream linear program, using mostly noisy training data supplemented by a small clean set. The authors introduce Reduced Cost Influence Functions (RCIFs), which combine influence functions from robust statistics with the geometry of linear optimization to estimate how each training point shifts downstream decisions. They propose a reweighting algorithm that downweights harmful training points using RCIFs to reduce downstream regret. Experiments on random linear programs and shortest-path problems show reduced regret, and the theory characterizes RCIF properties and regimes where the method is favorable over baselines.
In briefComputing black hole tidal Love numbers with physics-informed neural networks
SummaryThe paper presents a physics-informed neural network (PINN) framework that computes black hole tidal Love numbers directly from perturbation equations. It studies static perturbations of Kerr black holes, using their exact hypergeometric solutions as a benchmark for the solver. Across scalar, vector, and tensor perturbations, the method recovers the finite dissipative response with relative errors below 10^-5.
Maria J. Rodriguez, Luis Fernando Temoche, Pavlos Protopapas
In briefParetoSaddler: Pareto-guided tuning of LLM agent harnesses for quality and cost
SummaryThis paper presents ParetoSaddler, a method for optimizing the harness that wraps large language model agents, without changing model weights, to balance response quality and cost. It separates promoting intermediate candidates from final deployment selection and uses Bayesian bootstrap Pareto support to keep candidates showing credible gains in quality or cost. Evaluated on ClawMark and GAIA2, it raised quality and lowered cost per task compared with the baselines reported.
In briefAdaptive readout and decoding for diffusion driving VLA models by scene
SummaryThis paper addresses the high inference cost of diffusion driving vision-language-action models, which rerun the backbone at every denoising stage. It proposes a scene-adaptive framework with a scene-conditioned readout mixture-of-experts that combines trajectory predictions from late-layer representations after one backbone pass, plus an adaptive decoding policy that adds refinement only when a scene requires it. Experiments on the Waymo E2E and nuScenes benchmarks report improved planning quality and up to a 6.0× end-to-end speedup on Waymo E2E.
WhyThe authors propose scene-adaptive readout and decoding for diffusion driving VLAs, reporting up to 6.0x speedup and RFS gains on Waymo E2E and nuScenes.
In briefDecoupled Predictive Decision Making: a shared conditional predictor that lets AI agents learn from their own experience.
SummaryThe paper introduces Decoupled Predictive Decision Making (DPDM), which frames agent decisions as conditional prediction. Language models encode current and historical records into semantic representations, and a frozen general-purpose predictor (LimiX-2) predicts task targets from this evidence without updating model parameters. With Qwen3.5-27B as the encoder, mean accuracy across 50 benchmarks rises from 78.80% with few-shot prompting to 84.92%.
In briefSWE-Trail: Using Past Trajectories to Guide Data Collection for Coding Agent Training
SummaryThis paper presents SWE-Trail, a method that uses earlier coding agent trajectories to guide where new trajectories are collected, keeping that history out of the agent's context. Candidate actions are briefly tested from a shared checkpoint to choose which route continues. Training a 124B student model, it improves mean resolution over independent sampling by 2.33 points on SWE-bench Multilingual and 1.60 points on SWE-bench Pro.
In briefBack to the Future: Long-Horizon Prediction with Retrospective Correction for Disturbance-Aware Retrieval
SummaryThis paper introduces DARE (Disturbance-Aware REtrieval), a framework for retrieving objects from densely packed storage. It predicts how actions disturb neighboring objects through object-level message passing, and when a sparse observation arrives, a learned corrector revises an earlier predicted state and propagates the fix along the predicted contact chain. A receding-horizon planner uses these predictions to balance retrieval progress against the displacement and rotation of surrounding objects. Experiments are reported in planar and RoboCasa environments.
In briefState-aware virtual patient simulator for training psychiatrists in difficult clinical cases
SummaryThis paper presents a virtual patient simulation pipeline for training psychiatrists to handle high-risk scenarios such as suicidal ideation and hallucinations. The system keeps a patient state that evolves during the interaction, and a learned strategy selector chooses the type and level of challenge. A large language model then generates patient responses grounded in that state, the clinician's input, and clinical knowledge. After each session, the system gives the clinician feedback on their communication and management.
Harsh Mohta, C Chandrahas Reddy, Sunhith Reddy Bojja et al.
In briefLarge Geometry Models as Priors for Fast Physics Pretraining
SummaryThe paper proposes Euler-Lagrange geometry pretraining, which derives pretraining targets from the flow's Euler-Lagrange equation using shape distance and normal information rather than designing them by hand. The authors present AREALS, a masked surface autoencoder pretrained self-supervised and sign-free on 61,822 raw meshes. It is evaluated on field reconstruction and fine-tuning data efficiency across DrivAerML, DrivAerNet++, WindsorML, and DoubleDelta wing datasets.
WhyThe authors derive geometry pretraining targets from the Euler-Lagrange equation and propose AREALS, reporting reduced simulation needs across several aerodynamic datasets.
In briefFragility vs. Hardness: Conditional Causal Difficulty for LLMs
SummaryThe paper treats LLM task difficulty as the causal effect of meaning-preserving rewrites, defining a conditional causal difficulty (CCD) over tasks the model can already solve. It introduces GlassGSM and GlassMATH, benchmarks that augment GSM8K and MATH with perturbation-built twin problems that raise fragility. A predictor called Frail estimates per-perturbation break probabilities and is compared against surface-feature and supervised baselines.
WhyThe authors recast task difficulty as a conditional causal quantity and report comparisons with baselines on GlassGSM and GlassMATH.
Mohammad Reza Rezaei, Dingxian Wang, Nikhil Verma et al.
In briefDo embodied vision-language models actually use their scene-graph memory?
SummaryThis study examines whether vision-language models answering route questions for embodied agents actually rely on scene-graph memory. The authors propose a graph-swap test that gives each episode another episode's graph and counts how many answers change. On a question-answering benchmark covering routes through two to seven rooms, scene graphs delivered as text are used strongly, while graph-encoder delivery is weaker because standard pooling and late fusion discard information. Giving each node its own input token recovers part of the gain.
Shubh Garg, Divyansh Pandey, Vedant Khandelwal et al.
In briefReinforcement learning that keeps rare failure risks bounded using large deviations theory
SummaryThis paper studies finite-horizon reinforcement learning under a constraint on the probability of reaching a failure set, where failures are extremely rare but catastrophic. Instead of estimating the rare event probability, it uses the large deviations principle, bounding failure probability through the positive root of a damage variable's cumulant generating function via Cramér's theorem. The approach extends to Markov decision processes with hitting times using state-action cumulant generating functions, which are policy-independent and estimable from regular transitions. This yields a policy improvement algorithm in which every policy satisfies the failure constraint with high probability.
SummaryVision-language models sometimes describe objects that are not in an image, a problem called visual hallucination. This paper asks whether such models know which part of the image they are looking at, and formulates this as a LOOK task of identifying the visible regions of the visual input. The task is built in a self-supervised way, without manually labeling which regions the model can see. Experiments show that existing models fail to localize the visible region, and that fine-tuning on the LOOK task improves their awareness of visual input.
In briefSame Function, Different Gradients: Checking Gradient Reliability Under Reparameterization
SummaryGradient magnitudes are widely used to rank attention heads for pruning, set per-example clipping, allocate adapter rank, and debug models. This paper shows that the same input-output function can be computed in different parameter coordinates, which changes the gradients. The authors introduce Gradient Reliability under Reparameterization (GRR), which rewrites a trained model in new coordinates, confirms predictions are unchanged, and tests whether gradient-based decisions stay the same. Audits across GLUE classifiers, a ViT, GPT-2, a 1.1B Llama-style model, and LoRA adapters find that predictions are preserved while gradients and gradient-based decisions change.
In briefUniGS-SLAM: Monocular Gaussian SLAM with Better Geometry and Open-Vocabulary Semantics
SummaryUniGS-SLAM is an RGB-only monocular SLAM framework that reconstructs scene geometry and builds semantic maps. It separates open-vocabulary semantics from Gaussian reconstruction and fuses semantic observations into an independent sparse voxel map that tracks where each observation came from, so past contributions can be revised as reconstruction evolves. The fused semantics also guide Gaussian filtering and local surface refinement. Reported results on Replica, ScanNet, and KITTI-360 show strong tracking, rendering, geometry, and open-set semantic mapping, with a 2.37-fold speedup over the fastest compared semantic SLAM baseline.
In briefReinforcement learning across high-energy physics, from accelerator control to theory
SummaryThis survey reviews how reinforcement learning, a method where an agent learns decisions by interacting with an environment and receiving rewards, is applied in high-energy physics. It covers domains such as accelerator control, data quality monitoring, charged particle tracking, detector design, trigger control, analysis, and formal quantum field theory computation, comparing RL with traditional methods. The paper also discusses open challenges including reward design, sim-to-real transfer, safety, the lack of shared benchmarks, and statistical validity.
In briefBeliefs in feedback: factored representations of interacting agents in transformers
SummaryThis paper studies agents that interact through feedback loops using a computational-mechanics framework. The authors prove that any joint stochastic process can arise from such feedback-interacting agents and is generated by a hidden Markov model whose Bayesian beliefs factorize into per-agent components. Experiments show that transformers trained on these feedback sequences separate beliefs about each factor into distinct subspaces of the residual stream, with representation size scaling linearly with the number of agents. Interventions indicate these representations are causally effective.
David A. Ehrlich, Manuel Baltieri, Fernando E. Rosas et al.
In briefMETER: Matched-event method for calibrating brain-signal state transitions in short events
SummaryThis paper introduces METER, a framework for analyzing transitions between EEG microstates during short cognitive events. Each event is compared against duration-matched events from a fixed, participant-specific reference bank, with uncertainty in sparse transition tables handled explicitly. In simulations, event-level false-positive rates stayed between 0.044 and 0.056 across four null mechanisms. In 4,175 natural-reading events from 18 participants, the primary 30-ms analysis found no linguistic association that survived multiplicity correction.
In briefLoss-aware density control for 3D Gaussian Splatting
SummaryThis paper reframes adaptive density control in 3D Gaussian Splatting as local decisions that reduce the loss, replacing heuristic gradient and threshold rules. It uses Taylor-based per-Gaussian surrogates to predict loss changes from splitting, cloning, and opacity decay. The authors report improved rendering quality on standard benchmarks with comparable or smaller Gaussian budgets.
In briefECGJev: question-guided option scoring for 12-lead ECG interpretation
SummaryThis paper presents ECGJev, a framework for answering diagnostic questions about 12-lead electrocardiograms. It uses a pretrained language model as a decoder that scores all answer options in a single forward pass, adapted with diagnostic labels in two stages. ECGJev achieves the highest macro-F1 on the CPSC2018, PTB-XL and Chapman benchmarks and generalizes competitively to the small PTB cohort.
In briefSmoothness Is a Point on a Besov Spectrum, Not One Number
SummaryThis paper proposes a finite-difference roughness spectrum that measures how model outputs change under perturbations of varying size and direction, unifying Lipschitz sensitivity and directional Hessian curvature within one framework linked to classical Besov theory. In image classification, the proposed Roughness Source Control (RSC) improves noise robustness across 80 checkpoints. In post-training quantization, the mixed-precision allocation yields lower perplexity than the HAWQ-V2 baseline in all 14 settings across seven held-out large language models.
In briefReconstructing dynamic endoscopic scenes in 4D from single-camera videos using triangle splatting
SummaryThis paper presents SurgicalTS, a self-supervised triangle-splatting framework that reconstructs deformable endoscopic scenes from videos captured by a static monocular camera. The scene is represented by canonical triangles jointly optimized with a deformation network, using a quaternion-based deformation model, improved color parametrization, spatio-temporal priors, and a new initialization strategy for occlusions. Evaluations on in-vivo endoscopy videos from diverse procedures show it models soft-tissue deformation and tool-tissue interaction, outperforms state-of-the-art methods, and reduces training time.
Laura Salort-Benejam, Stamatia Giannarou, Antonio Agudo
In briefSensOFormer: A foundation model for particle accelerator sensor data
SummaryThis paper introduces SensOFormer, a denoising masked transformer with a Perceiver-style encoder-decoder for particle accelerator sensor telemetry. It handles variable sensor sets and heterogeneous sampling rates and is pretrained self-supervised on historical accelerator data. The model is transferred to missing-data imputation, anomaly detection, cavity identification, and fault identification. The authors argue a single model can learn reusable representations across these operational tasks.
In briefLARK: Using Lean to formally verify GPU kernels written by AI
SummaryLARK is an agent framework that formally verifies GPU kernels generated by LLMs against their PyTorch reference implementations in Lean 4, rather than relying only on testing. Humans only need to verify the operator-level Lean semantics in advance, while LLM agents split each kernel into segments and prove each matches its PyTorch counterpart. On KernelCraft tasks, LARK fully verified 10 of 12 Level-1 kernels and all 5 Level-2 kernels. On 10 production kernels from open-source LLM serving frameworks, it verified 9 and found a new data race bug in the tenth.
WhyThe authors formalize the GPU machine model in Lean 4 and use agents to prove kernel-to-PyTorch equivalence via segmentation; they report verifying 10 of 12 KernelCraft Level-1 kernels and finding one
In briefUncertainty-aware 3D position reconstruction for CZT drift-strip detectors
SummaryThis paper examines how to reconstruct three-dimensional interaction positions in CZT drift-strip detectors while also estimating prediction reliability. It compares deep ensembles of point-estimation networks with simulation-based inference, which estimates a posterior distribution over position. Both methods are trained on physics-based simulations and tested on synthetic and experimental data. On synthetic data, uncertainty from both methods correlates with positioning error, more strongly for SBI, while on experimental data the deep ensemble improves z-position resolution and SBI better highlights low-information regions and simulation-to-detector discrepancies.
Alejandro Valverde Mahou, Micha{\l} Kossakowski, Selina R. H. Owe et al.
In briefEvaluating vision-language models' 3D spatial understanding through scene reconstruction
SummaryThis paper introduces RECONSPATIAL, a benchmark that asks vision-language models to reconstruct 3D scenes from multi-view images by predicting oriented 3D bounding boxes in a camera-anchored metric frame. Reconstructions are scored on relative structure and metric accuracy, and 18 frontier models show a consistent gap between the two. The authors also extend evaluation to offline navigation and 3D map injection, and report that spatial fine-tuning transfers to general vision-language benchmarks.
In briefClickSight: An LLM Framework for Reading Student Click Data Through Learning Strategies
SummaryDigital learning platforms record detailed clickstream data showing how students work on tasks, but interpreting it usually takes considerable expert effort. The paper introduces ClickSight, which uses a staged large language model pipeline to find evidence tied to predefined learning strategies, judge how well each strategy is followed, and produce an interpretation grounded in that evidence. The authors report evaluations across two learning environments, 365 students, and 12 model backbones.
In briefCULTURE-CMT: A Benchmark for Context-Aware Social Media Comment Translation
SummaryThis paper introduces CULTURE-CMT, a Chinese-English benchmark for translating social media comments with their post and conversational context. It contains 572 manually verified examples grouped into Slang, Context, and Both categories. The authors also build a retrieval-augmented judge that scores cultural effectiveness of translations. Results show that interaction context helps most models but does not remove culturally ineffective translations, and that a 27B model adapted on social media data achieves the lowest invalidity rate under thinking inference with context.
In briefSwitch-SWE: Can coding agents stay focused when switching between unfinished tasks?
SummaryThis paper introduces Switch-SWE, a benchmark that tests whether coding agents can resume an earlier task after working on other unfinished tasks in the same repository. It includes 200 focus tasks from 19 Python repositories, evaluated in matched uninterrupted and interleaved sessions. Across 7 agents, interleaved work lowered focus-task resolution by an average of 11.5 percentage points. The authors also propose a context-management module that improves completion under task switching by 5.5 to 15.5 percentage points on OpenCode.
WhyThe authors report a benchmark of 200 focus tasks across 7 coding agents measuring resumption after same-repo task switching, plus a plug-in context module yielding 5.5 to 15.5 point gains.
In briefScaling Makes It Possible: How Large Models Learn Impossible Languages
SummaryThis paper examines whether language models can learn impossible languages that violate the structural, cognitive, or logical constraints of human languages. It introduces semantic coherence, a concept measuring how effectively semantic information is expressed, and tests models across multiple scales. Billion-parameter models perform similarly on impossible and possible languages, while million-parameter models degrade significantly on impossible ones, with the gap widening as semantic complexity grows.
In briefMedPsy: Small Medical Language Models Built for Efficient Edge Deployment
SummaryThis paper presents MedPsy, a family of text-only medical language models with 1.7B and 4B parameters designed to run on edge devices. The approach combines synthetic medical data, chain-of-thought targets from a 235B medical-focused teacher LLM, and a four-stage post-training curriculum of two SFT stages followed by two RL stages. On seven closed-ended medical benchmarks, MedPsy-4B surpasses MedGemma-1.5-4B by 19.34 points and matches MedGemma-27B while being 6.75x smaller. The models also produce shorter responses and show lower end-to-end latency in real-device tests on laptop and smartphone hardware.
Davide Vitabile, Alexandro Buffa, Akshay Nambiar et al.
In briefRetrieval-enhanced cultural evaluation for training social media translation models
SummaryThis paper addresses the difficulty of accurately judging cultural nuances in translations of social media user-generated content. It introduces CUE-MT, a family of translation models trained with a calibrated judge that is aided by an agent retrieving external evidence and distilling it into hints for scoring and rewards. Across three backbone sizes, CUE-MT outperforms comparably sized open-source baselines on its own benchmark and three public benchmarks. Ablations support the benefit of retrieval and hard-example mining.
In briefNAS-PO: A Policy Optimization Method for Vision-Language Models Using Native Attention
SummaryThis paper proposes NAS-PO, a reinforcement learning post-training method for vision-language models. Instead of relying on image perturbation proxies, it uses the model's native attention as a signal to encourage attention patterns that support visually grounded reasoning. Experiments on 7 out-of-domain reasoning and perception benchmarks show it substantially outperforms leading RL-tuned VLMs at both 8B and 32B scales.
In briefMIRA: A method using mutual information to make test-time model adaptation more reliable
SummaryThis paper proposes MIRA, a source-free online adaptation method that updates a deployed model using only unlabeled target data. It adds lightweight Gaussian perturbations to the last-layer classifier and uses mutual information and prediction agreement to judge which predictions are locally stable. Samples that are sensitive to perturbation are down-weighted in an entropy-minimization loss. The authors report accuracy, calibration, and negative log-likelihood improvements on distribution-shift benchmarks.
Hyeongyu Kim, Mingyeong Jang, Youngjun Song et al.
In briefSimulating Profile-Aligned Students via Hierarchical Planning in Long Learning Environments
SummaryThis paper introduces ALPS, a hierarchical simulated student that follows observable, environment-specific behavior profiles. Instead of choosing each action independently, it plans profile-aligned task segments and then generates the corresponding actions. Tested in an inquiry-based physics game and a pharmacy simulation, ALPS outperforms turn-level generation on profile consistency under scarce data, though its coverage is lower.
Bahar Radmehr, Jiaxu Zhao, Fatma Bet\"ul G\"ure\c{s} et al.
In briefTrackMem: A memory system that tracks changing beliefs in long financial conversations
SummaryThis paper presents TrackMem, a per-speaker graph memory that follows how financial figures are revised, separates views by entity and analytical lens, and keeps each speaker's actions in chronological order. The authors also introduce FinLoCoMo, a benchmark of 15 long-horizon investment conversations with 765 questions. TrackMem outperforms existing memory systems on both splits of the benchmark, reaching 73.9% macro accuracy on the failure-mode split, with the largest gains on questions about revised figures.
Austin M.Y. Cheung, Harsh Vardhan Gupta, Go Nam Lui et al.
In briefSafety4DM: A Vision-Language Framework for Risk-Aware Dynamic Maps in Vehicle-Road-Cloud Driving
SummaryThe paper proposes Safety4DM, a safety-oriented language system for dynamic maps used in vehicle-road-cloud cooperative driving. It introduces RoadRisk, a real-world road image dataset covering four road condition types and four risk levels, and fine-tunes a vision-language model to output structured risk information the map can ingest. A Safety-aware Step-wise Prompting (SafeStep) mechanism serves as the natural language query interface and generates driving advice. Experiments report 89.83% risk perception accuracy for fine-tuned Qwen3-VL-4B and over 93% query accuracy across several large LLM families.
In briefTesting Glaucoma Detection CNNs on New Datasets: Thresholds, Multi-Source Training, and Color Normalization
SummaryThis paper examines how deep learning models for glaucoma detection from eye fundus images perform on datasets they were not trained on. Five pretrained networks were fine-tuned on one of three public datasets and tested on the other two, with cross-dataset AUC falling about 0.21 below in-domain results. LAB reference normalization and strong color augmentation reduced this drop but did not remove it. The authors conclude that strong in-domain AUC does not show a model will generalize, and that studies should report external validation and operating thresholds.
In briefSynthetic Realities: Can People Tell AI Videos from Real Footage?
SummaryThis study asked 87 untrained viewers to judge 18 real and 18 AI-generated videos as either real or AI-made. The AI clips were produced using text-to-video, image-to-video, and video outpainting methods. Participants were correct 60% of the time on average, recognizing image-to-video clips most often (79%) and text-to-video clips least often (43%). Theme also mattered, as environmental clips were more often mistaken for AI-generated content.
Danyal Sarfraz, Raja Mubashar Karim, KwanMyung Kim
In briefRubricWorld: Optimizing LLM Grading Rubrics with a World Model
SummaryRubricWorld is a framework that improves the scoring instructions (rubrics) used by LLM graders while keeping the grader's parameters fixed. A learned world model predicts how each rubric edit corrects or introduces scoring errors, guiding which candidate edits to test next. Across EIR (National), ASAP2.0, and ASAP-SAS with three frozen LLM graders, it achieves the highest score in 20 of 27 metric comparisons. On ASAP2.0, gpt-5-mini accuracy rises from 18.20% to 46.70%.
In briefLarge-scale Bayesian optimisation with calibrated product-of-experts GPs and trust regions
SummaryThis paper proposes BO-pro-c-tr, a new Bayesian optimisation algorithm for large-scale and high-dimensional problems. It uses a product-of-experts Gaussian process with uncertainty calibration as its surrogate model and links this model to trust regions through a one-to-many association strategy. With the GP-LCB acquisition function, the authors report average reductions of 2.6% in simple regret and 34.9% in computational overhead compared with methods using independent local GPs and trust regions.
In briefSurgStream: A Real-Time Surgical Video Understanding Framework and Benchmark
SummaryThis paper presents SurgStream, a framework for real-time surgical video understanding aimed at intraoperative support such as workflow monitoring and procedural documentation. It introduces Workflow-Anchored Memory, which stores structured instrument-action-target observations and summarizes them by surgical phase, and Deadline-aware Transition Learning, which trains response states within clinical time windows. The authors also release SurgStream-Bench, with 11.7K streaming interactions, and report gains over prior methods, particularly on memory and forward proactive QA tasks.
In briefOpening the Black Box of Neural Auction Mechanisms: A Post-hoc White-box Auditing Framework
SummaryThis paper addresses the lack of transparency in neural auction mechanisms used on Web platforms and proposes a post-hoc white-box auditing framework. Using CITransNet as a testbed, it applies sparse autoencoders to decompose hidden representations and validates interpretable features through activation prediction and causal interventions. The authors show that these verified features can analyze misreporting incentives, alignment with Myerson's optimal auction, and the diagnosis of individual auctions.
In briefAutomated Pain Recognition from Pulse and Breathing Signals Using Hankel-FCN
SummaryThis paper proposes a deep learning framework for objective pain assessment using Blood Volume Pulse (BVP) and Respiration (RESP) signals from the AI4Pain dataset. The signals are converted into trajectory representations via Hankel matrix embeddings and processed by modality-specific Fully Convolutional Networks, then fused late to classify No Pain, Low Pain, and High Pain. The method reaches 67.09% accuracy and a 66.41% F1-score, and multimodal fusion outperforms either modality alone.
Valentina Becchetti, Muhammad Umar Khan, Stefanos Gkikas et al.
In briefMUPA²S: A Unified Multimodal Model Using Asymmetric Attention to Assess Stress
SummaryThis paper presents MUPA²S, a framework that automatically assesses stress by combining facial video and ECG heart signals in a shared token representation. It processes the data with asymmetric attention, mixing segment-local cross-attention with global self-attention. On a newly collected dataset of 58 subjects, facial video reaches 69.44% test accuracy and unified Video-ECG fusion reaches 69.58% with about half the parameters of feature-level fusion.
In briefKnowledge-guided reasoning to estimate floodwater levels in cloud-edge monitoring systems
SummaryThis paper proposes HydroReasoner, a floodwater level reasoning framework for cloud-edge collaboration that treats detection as an explicit visual reasoning problem. It is built on Qwen2.5-VL, trained with Group Relative Policy Optimization, and guided by a floodwater knowledge graph, HydroKG, through a knowledge graph path reward. The authors also apply 4-bit quantization for edge deployment, add a confidence-based cloud-edge routing mechanism, and introduce the HydroLevel dataset of 3,000 images. Reported results show accuracy gains of 4.0 and 13.6 percentage points over conventional vision baselines in in-domain and out-of-domain evaluations, respectively.
In briefBetter answers, but weaker reliance on evidence: a bypass problem in on-policy distillation
SummaryThis paper studies on-policy distillation (OPD) for scientific document question answering, where answer accuracy can rise while responsiveness to supporting evidence falls. Using matched document edits that change the supported answer, the authors identify "evidence bypass," where a model keeps its original answer after the document changes. They propose Evidence-Paired OPD (EP-OPD), tested on QASPER and PeerQA.
In briefLaunder-Bench: Can AI-driven rewrites of whole code repositories be detected?
SummaryThe paper notes that coding agents can rewrite entire code repositories while preserving tested behavior, which challenges plagiarism detection, license auditing, and provenance tracing. The authors define repository-level agentic code laundering and introduce Launder-Bench, built from 282 real repositories, with pairwise detection and source-attribution tasks. Existing methods perform far from saturating the benchmark, revealing a gap between relative code similarity and source-specific provenance evidence.
WhyThe authors introduce a repository-level provenance benchmark for agentic rewriting, built from 282 repositories, reporting a gap between similarity and source evidence.
In briefOC-OPD: Observation-Coordinated Supervision for On-Policy Distillation
SummaryThis paper studies on-policy distillation, where reverse and forward KL divergences provide complementary training signals. The authors build a four-observation matrix over student-generated prefixes and prove an identity linking the two KL gradients to a sum of contrasts, giving a shared descent direction. The proposed OC-OPD method improves mathematical reasoning across three student settings and transfers to scientific question answering and code generation.
