In briefRouteREPA: Routed Representation Alignment for Pixel-Space Image Transformers
SummaryThis paper proposes RouteREPA, a method for improving the training of Just Image Transformers (JiT), which generate images directly in pixel space. The authors find that applying standard representation alignment (REPA) to JiT conflicts with pixel reconstruction and lowers the reconstructability of intermediate representations. RouteREPA uses lightweight, representation-dependent routers to adaptively direct task-oriented representations for alignment and generation, applied across multiple network depths. Experiments on ImageNet-1K at 256x256 resolution show it consistently outperforms standard REPA across model scales and training budgets.
In briefPlagiaCorpus: A bilingual dataset for testing plagiarism detection, including AI-generated and cross-language text reuse.
SummaryPlagiaCorpus is a bilingual Arabic-English dataset for benchmarking plagiarism and text-reuse detection, including reuse by large language models and across languages. It contains 198,643 records, with 34,885 source passages and 163,758 queries spanning five reuse classes and a Safe class of non-reused text. The train and test splits are grouped by source passage so no passage appears in both. The dataset is released in JSON and Parquet formats under a CC BY-SA 4.0 license.
In briefWhy table foundation models work well on unseen relational databases
SummaryThe paper studies how to predict missing or future values in a column of a relational database without training a new model for each task. The authors first confirm with more diverse experiments that single-table foundation models paired with parameter-free relational encoders remain competitive, even without any relational pre-training. They then offer several complementary explanations, arguing that these encoders map relational effects into fixed-length embeddings that resemble the distributions single-table models were designed for, so their properties carry over to relational settings.
In briefILR-SMO: A Robust Method for Integrating Spatial Multi-Omics Data Through Iterative Refinement
SummaryThe paper proposes ILR-SMO, a self-supervised framework that integrates spatial multi-omics data by progressively refining a shared latent representation rather than fusing modalities in a single step. It adds reliability-aware modality gating and a joint objective to handle modality imbalance and noise while preventing representation collapse. According to the authors, ILR-SMO outperforms seven state-of-the-art methods on five spatial multi-omics benchmarks and shows strong robustness across diverse settings.
In briefTraining-time domain transfer for diagnosing new imaging types without generation at inference
SummaryThe paper proposes training-time domain transfer, which generates synthetic target-domain training data from archived routine samples so models can diagnose real novel-modality inputs directly, without image generation at inference. Unstained prostate histopathology serves as a proof of concept, compared against inference-time virtual staining and conventional H&E-based cancer detection. Training-time transfer showed higher concordance with the conventional benchmark than virtual staining for both raw and deparaffinized tissue.
Julia A. Mielcarz, Andrea Camilloni, Nita Mulliqi et al.
In briefEvaluating 3D Human Pose Estimation Models for Analyzing Fitness Exercises
SummaryThis paper tests recent 3D human pose estimation methods on Fit3D, a dataset of synchronized multi-view recordings of many fitness exercises with accurate annotations. It compares two monocular lifting methods, TCPFormer and PoseMamba-L, with the multi-view model MVGFormer, and examines the effect of 2D detectors and three temporal filters. The authors conclude that 2D detection quality, temporal modeling, and post-processing should be considered together in exercise analysis pipelines.
N\'estor Ortega, Agust\'in Salgado, Javier S\'anchez
In briefMaking Contrastive Regularization Work for Diffusion Transformer Training
SummaryThis paper asks whether standard InfoNCE contrastive learning can regularize diffusion transformer training. The authors find that naive InfoNCE gives only marginal gains because near-orthogonal noise components create a shortcut, especially at high noise levels. They propose CRAN, which anchors contrastive dispersion to distances between clean-sample representations to suppress this noise shortcut. Experiments on ImageNet show CRAN consistently improves training across model scales and training budgets.
In briefHow Unimodal Augmentation-Based Self-Supervised Objectives Lead Separate Models to Align Across Modalities
SummaryThis paper examines why independently trained single-modality models come to show implicit cross-modal alignment. The authors show that self-supervised objectives based on unimodal data augmentation induce this alignment: low unimodal losses produce strong cross-modal alignment, while high unimodal losses prevent it. They theoretically link the widely used InfoNCE loss to local CKA, a measure of cross-modal alignment, and support the claim with controlled experiments and tests on 53 frozen text, vision, and audio encoders.
In briefDefining Autonomous Weapons: A flexible framework for purpose-based definitions
SummaryThe paper reviews and critiques the many definitions of autonomous weapon systems (AWS) proposed by scholars, governments and international regulators. The authors argue that a single definition is insufficient and that a definitional framework is needed to classify different AWS within one conceptual family. Building on a broadly accepted definition, they propose a schema whose variations can be easily modified to classify various concepts within the AWS category.
In briefDependency-aware state revision for long-term agent memory
SummaryLong-lived agents store memories of users, tasks, and decisions, but when an upstream fact changes, existing systems revise only the contradicted memory and keep derived beliefs, plans, and actions. The paper formulates dependency-aware state revision and proposes StateGraph, which maintains a state dependency graph recovered from interaction histories. Counterfactual verification separates support dependence from semantic relatedness, invalidating dependent states only when a strict dependency fails and no independent support path survives. The authors also introduce StateChangeBench, a 50-case benchmark, where StateGraph shows gains over a CUPMem reimplementation.
In briefCloudra: A multi-agent framework for checking reaction mechanisms with geometry and validation
SummaryThis paper presents Cloudra, a multi-agent framework that organizes chemical objects in a Delivery DAG, linking species nodes with elementary-step transition states. Cheminformatics methods fix molecular identities and atom mappings, while frequency and IRC calculations confirm which species a saddle point connects. On CloudraBench's 50 tasks, the Guided workflow reaches a mean composite score of 0.805, versus 0.584 for the same-model Standard baseline.
In briefHow user attributes and political leaning affect LLM receptiveness in online political debates
SummaryThis paper studies how large language models respond when they take part in online political debates. The authors reconstructed 5,376 Debate.org debates as LLM conversations and ran nearly 150,000 controlled multi-turn conversations that varied user political orientation, expertise, rhetorical strategy, and defended position. They find that LLMs are systematically more receptive to left-leaning than right-leaning users, an asymmetry that persists even when opposing users share a conversation. Repeated interaction raises receptiveness without removing this ordering, and fine-tuning on different political corpora can largely erase or reverse the pattern.
Akram Elbouanani, Yousra Taleb, Gianluca Demartini et al.
In briefWhat makes a mathematical theorem interesting, and how to measure it for open-ended exploration
SummaryThis paper studies what makes mathematical theorems interesting, by observing which results mathematicians single out in formal libraries and papers. Each selection pattern is modeled as an evaluator that weights features of a result's statement, proof, and dependencies, split into shared and subjective parts. In formalized mathematics, ex-ante features predict these selections about as well as a large language-model judge, and the two complement each other. Much of interestingness is shared, carried mainly by structural features such as statement and proof length.
In briefMaswao: A South African Sign Language dataset built automatically and benchmarked for translation.
SummaryThis paper proposes a framework that automatically collects and prepares sign language datasets with minimal human labour, using South African Sign Language (SASL) as a testbed. It gathers SASL videos from public sources such as news broadcasts, transcribes their audio with speech recognition, and uses motion-based detection to find sentence boundaries, producing the weakly sentence-aligned Maswao corpus. Benchmarks on sign-to-text translation show that Transformer models outperform recurrent baselines, with the best achieving BLEU-1 to BLEU-4 scores of 13.78, 5.21, 2.52 and 1.54.
Thapelo Sindane, Unarine Netshifhefhe, Vukosi Marivate et al.
In briefTailorRig: Controllable Auxiliary Skeleton Generation for Garment-Aware Character Rigging
SummaryTailorRig is a framework that treats rigging humanoid characters with complex garments as a controllable, structured generation task rather than automatic rig recovery. Artists can specify a body skeleton template and semantic labels for garment bone chains, and a hierarchical method first assigns vertices to kinematic chains before refining bone weights to reduce weight bleeding. The authors introduce Human-Mix, a dataset of over 38,000 humanoid models, and report improvements in skeleton accuracy, deformation stability, controllability and usability over existing methods.
In briefControl foundation model for dynamic block diagrams
SummaryThe paper proposes a control foundation model that represents connections, signal values, and missing values as dynamic graphs, and frames response prediction and control-input generation as one masked prediction problem. The Block Diagram Transformer implements this formulation by combining time-axis attention with a graph neural network that follows block-diagram connections, and it can be initialized from pretrained time-series foundation models. Trained on control-experiment data mixing multiple settings, the model stabilized an unstable system through feedback control, which the authors report is hard to achieve by fine-tuning existing time-series foundation models.
In briefPhotoGuider: Composition-aware cropping and 360° framing with explainable photo scoring
SummaryPhotoGuider is a method that first finds the best viewing direction in a 360° panorama and then frames the shot, and it also handles standard image cropping. It scores images using shared composition-aware features grouped into four photographic tokens (Subject, Composition, Aesthetics, Integrity) and uses a fine-tuned multimodal model to explain composition. It achieves the highest average mIoU on two standard 2D cropping benchmarks and ranked first in 53.2% of cases in a blind study of 120 participants over 100 panoramic scenes.
Yuan-Kang Lee, Kuan-Lin Chen, Chia-Che Chang et al.
In briefLANCE: Linking Low-Level Image Fusion to High-Level Semantic Understanding
SummaryThe paper proposes LANCE, a framework for remote sensing that uses structured language as an indexable semantic state to connect image fusion with semantic interpretation. A frozen vision-language model produces a closed-vocabulary text record, a language-conditioned generator fuses the imagery, and a verifier re-describes the output to close the loop. Across seven test sets on GF2, QuickBird, WorldView-III and WorldView-II, LANCE improves reduced-scale reference metrics by up to 11.7% over the strongest learning baseline, and independent evaluation shows gains of +6.4 mAP and +0.072 open-vocabulary F1.
Guangyuan Dong, Lingdong Shen, Xudong Zhang et al.
In briefVoice-native LLMs detect frustrated voices in audio but fail to act on them
SummaryThis paper examines whether voice-native language models, which process speech directly rather than through transcripts, notice frustration in a speaker's voice. Using a 2x2 design that pairs frustrated or neutral wording with frustrated or neutral voice over 353 tasks, the authors release VoxAffect, a dataset of 1,412 clips. Most models rarely label the frustrated voice as frustrated, yet linear probes can decode it from their audio encoders, suggesting the models hear the voice without using it.
WhyThe authors apply Brunswik's lens model to separate perceiving from using vocal frustration, testing eight models on 1,412 clips.
In briefSequential quality control for human-AI data labeling when shared bias is present
SummaryThe paper proposes JABER, a bias-aware sequential quality control framework for annotation that jointly models label correctness and shared biases to decide whether to gather more evidence or stop. It develops a finite-horizon planning algorithm to minimize expected quality control loss under resource limits. Experiments report reductions in expected loss compared with methods that model label and bias separately and with one-step lookahead, and zero labeling errors in real-API experiments on FEVER.
In briefGibbs variational last layers for learning long-horizon system dynamics
SummaryThis paper improves long-horizon probabilistic prediction of dynamical systems using variational Bayesian last layers (VBLLs). The authors reinterpret VBLLs through Gibbs variational inference and propose multi-step training losses that use continuous ranked probability score (CRPS) and full autoregressive rollouts. Experiments on synthetic, chaotic, and real-world data show that these multi-step CRPS objectives outperform single-step likelihood training in long-horizon accuracy and distributional fit.
In briefTug of Watts: Energy-Accuracy Trade-offs of LLMs in Medical QA
SummaryThis paper studies the trade-off between energy consumption and accuracy for large language models on medical question answering. The authors estimate the energy use and carbon emissions of 13 LLMs from 3 model families, ranging from 270M to 14B parameters, on tasks from the OpenMedical-LLM leaderboard. They argue that reporting energy and accuracy separately is insufficient, and that joint trade-off metrics need to be systematically constructed and interpreted.
In briefA universal policy wrapper that guarantees goal reaching in control tasks
SummaryCALF-Wrapper combines a fallback policy that provably reaches and maintains a goal set with a pretrained reinforcement-learning base policy, using the base policy's critic estimate to choose between them during deployment. Under the stated assumptions, the paper proves that the wrapper inherits the fallback's goal-reaching property and gives probability bounds on settling time and maximum distance from the goal set. Experiments in four simulated control environments show effective task performance and the highest observed goal-reaching rate, without extra dynamics models, online learning, or constrained optimization at execution time.
In briefCombining Domain Adaptation and Retrieval for Saline-Alkali Land Reclamation Planning
SummaryThis paper presents Wenyao, an agricultural decision-support system that combines domain adaptation with retrieval-augmented generation for planning saline-alkali land reclamation. The authors built SA-Bench from 8,664 bilingual research papers, comprising 4,860 paper-level planning instances. In a blind assessment by 31 experts, Wenyao with retrieval received the highest mean scores for objectives and projected outcomes, while GPT-5 scored highest for cost estimates.
In briefExpanding Aircraft Coverage of the Poll Schumann Model with Public Data
SummaryThe Poll Schumann model needs airframe and engine parameters, yet published tables cover only part of the aircraft fleet. This work reconstructs missing parameters from certification documents, manufacturer manuals, and ICAO engine records, using published equations for physical parameters and regression for the propulsion efficiency scale. The inventory covers 90 type designators, and the reconstructed values are compared against pycontrails and recorded A320 flight data.
In briefProtoFusion: Protocol-guided prompt refinement for generative infrared and visible image fusion
SummaryThis paper presents ProtoFusion, a framework that treats infrared and visible image fusion as generation constrained by a structured protocol covering environment state, grounding principles, and style. The protocol is optimized through a closed-loop process of generation, evaluation, and revision, then fixed at inference. The authors report improved quantitative results on standard benchmarks, favorable aesthetic evaluation, and visually coherent fused outputs.
In briefSigned causal learning with latent SDEs to infer dynamic brain effective connectivity
SummaryThis paper proposes ISCL, a latent stochastic differential equation framework that derives time-varying effective connectivity between brain regions and separates excitatory from inhibitory influences. The method is evaluated on the synthetic Lorenz96 and NetSim benchmarks. It is then applied to HCP Social Cognition task-fMRI data to analyze the causal roles of networks during social cognition.
In briefAGiR: Reducing Over-Reliance on Gifts in Mixed-Motive Games
SummaryThis paper studies mixed-motive games (MMGs), where self-interested agents must both cooperate and compete. The authors find that existing gifting methods can make agents overly dependent on received gifts, so they fail to learn reward-earning behaviors. They propose a refusal-augmented gifting framework with an adaptive mechanism that lets agents selectively reject some gifts based on policy-update signals. The mechanism works with various gifting schemes as a plug-in and supports decentralized learning.
In briefTesting super-resolution methods on multispectral crop images
SummaryThis study examines whether single-image super-resolution methods built for RGB images can sharpen red, green, and near-infrared (RGN) crop imagery. Eight pretrained methods are applied directly to RGN images without fine-tuning, at 4x and 16x magnification, on whole images and segmented leaf regions. Transformer-based methods significantly outperform GAN- and diffusion-based approaches, with HAT, HiT-SR, and SRFormer performing best.
Abhiram Karukayil, Jo\~ao F. C. Mota, Mateo V\'asquez-Guevara et al.
In briefSolv3R: a 3D looped transformer that refines multi-view geometry iteratively
SummarySolv3R is a looped transformer that solves multi-view geometry by continuing from its current estimate, refining camera poses and scene structure. Its update operator does not depend on the number of iterations, so already-solved and newly added views can be updated together, and representations of solved views can be cached for reuse. The authors report that one learned solver supports offline reconstruction, incremental reconstruction, relocalization, and streaming without task-specific training.
Jeong-gi Kwak, Andrea Tagliasacchi, Sho Kagami et al.
In briefKolmogorov-Arnold Networks vs. MLPs: Statistical Benchmark on Clinical Tabular Data
SummaryThis paper benchmarks nine Kolmogorov-Arnold Network (KAN) variants against a tuned standard MLP and a Random Forest across seven medical tabular datasets. It uses repeated stratified cross-validation, nested Bayesian hyperparameter tuning, and AUPRC, Brier score, and efficiency metrics. On severely imbalanced datasets, some KANs such as TabKAN and ReLUKAN achieve higher AUPRC than MLPs, but overall discrimination is comparable while KANs incur longer inference latency and more parameters.
In briefAdapting frozen feature upsampling by guiding spatial attention
SummaryThis paper examines whether gains from adapting features before a frozen JAFAR upsampler come from changes in feature values or spatial aggregation. The authors introduce a readout-null intervention that keeps coarse classifier logits fixed while letting spatial attention weights adapt. In three-seed experiments on PASCAL VOC and Cityscapes, adaptation before upsampling improves mIoU by 0.807 and 1.413 points over adaptation after upsampling, and spatial guidance is found to be the useful adaptation target.
In briefScaling action-free video prediction for transferable driving visual representations
SummaryThis paper studies scaling action-free video latent prediction as a pretraining objective for end-to-end autonomous driving. The proposed ScaleDrive framework is pretrained only on large-scale internet open-world driving videos, with the video encoder kept frozen during downstream planning training. Ego-motion and depth probes show that larger-scale pretraining progressively strengthens the temporal and geometric information in the frozen representation. ScaleDrive consistently improves downstream performance across different end-to-end driving policies and closed-loop benchmarks.
In briefMethod That Decides When to Request Extra Reasoning Proposals
SummaryThe paper introduces ValueFormation, a test-time policy that, after each sample from a primary model, chooses to stop, keep sampling, or acquire one proposal from a heterogeneous reasoning specialist. All three actions are compared on a shared cost-regularized value scale learned from vote-state features, which is frozen across tasks and models. On held-out MATH-500, a Llama-fitted calibration beats the strongest of eight baselines by more than fifteen points for three 8B primaries and transfers unchanged to Qwen3-14B.
In briefTLAPS-Bench: Machine-checked correctness proofs for real system specifications in TLA+
SummaryTLAPS-Bench is a benchmark and harness for using AI agents to build machine-checked proofs of real-world system specifications written in TLA+. It supports iterative proof development on the TLA+ Proof System (TLAPS) and includes safeguards against reward hacking. The authors used AI agents to prove new properties for cache coherence, Byzantine consensus, and write-ahead log specifications. On 70 of its 265 proof targets, the best-scoring model proved 31 (44.3%).
Qian Cheng, Ruize Tang, Saad Mohammad Rafid Pial et al.
In briefSpecOPD: A Parallel Training Method That Speeds Up Speculative Decoding for Language Models
SummaryThis paper proposes SpecOPD, a branch-parallel on-policy distillation framework for training the draft model used in speculative decoding. It runs draft sampling and target supervision in parallel across branches anchored at reference points, sharing target-prefix computation through proposal trees. The authors report up to 41.45% higher mean acceptance length, up to 16.09x supervised-token throughput over sequential-rollout on-policy distillation, and up to 3.56x generation speedup over autoregressive decoding.
In briefAdaptive Classifier for Urban Point Cloud Segmentation Using Aerial Context
SummaryUrban photogrammetric point clouds are usually processed as small local 3D patches, which limits the context available for telling similar categories apart. The paper presents ACASAC, a classifier that adds decision-level adaptation on top of feature fusion of aerial orthophoto and 3D data. It uses a point-aligned aerial region to link the full aerial view with local patches and models the relations between aligned 2D and 3D representations to build a classifier for each sample. It improves a fixed dual-representation head by 2.1 and 2.7 mIoU points on SensatUrban and UrbanBIS.
In briefA general AI agent for formalizing mathematical foundations with proof assistants
SummaryThis paper presents MFF-Agent, a general-purpose agent that combines large language models with the Rocq proof assistant to formalize mathematics. It includes task classification, tool-assisted execution, verification feedback, and state-based task management to generate, refine, and verify formal proofs. The agent is tested on case studies covering Morse–Kelley set theory, real numbers, propositional logic, and equivalences of the axiom of choice.
In briefNIR: Approximate Neural Instance Retrieval for ALCIQ
SummaryThis paper presents NIR, a family of neural instance retrievers for the ALCIQ description logic that output graded membership scores from knowledge graph embeddings. The authors compare token-based encoders (GRU, LSTM, Transformer) with a compositional encoder that mirrors description logic operators through primitive sub-networks. On eight benchmarks, all variants closely approximate the Pellet reasoner on complete knowledge bases, and the Transformer is the most robust under incompleteness. The source code is publicly available.
N'Dah Jean Kouagou, Pritilata Saha, Caglar Demir et al.
In briefIterated Width-4 ReLU Networks: Turing Completeness and Verification Complexity
SummaryThe paper studies iterates of a fixed width-4 ReLU network acting as a discrete-time system on the cube [0,1]^4. It shows that every deterministic Turing machine can be encoded by such a network, constructible in polynomial time, whose iterations exactly simulate the machine with iteration count proportional to running time. Reachability for these iterated networks therefore spans NP-complete, NEXP-complete, and undecidable (RE-complete) cases depending on the time bound.
WhyThe authors report a width-4 iterated ReLU construction simulating Turing machines on convex cubes, yielding a full complexity hierarchy and extensions to smooth activations and Neural ODEs.
In briefCARE: A framework for an in-house customer-service AI agent that improves itself from its own failures
SummaryThe paper presents CARE, a framework that lets production customer-service agents recursively improve from their own accumulated experience. After a one-time cold start on governed production traces, an in-house mid-sized model serves traffic alongside a proprietary incumbent and is updated using its own diagnosed failures via on-policy self-distillation. Candidates are promoted only after simulation and inspection scores, an online comparison with the incumbent, and human approval.
In briefOmni-Life: A Benchmark for Real-Time AI Assistants in Daily Life
SummaryThis paper introduces Omni-Life, a benchmark that tests whether real-time omnimodal AI assistants can give timely and appropriate help in everyday situations. It contains 400 videos, 290 requiring assistance and 110 requiring silence, combining synthesized and real-world footage. Evaluating twelve model configurations, the authors find that at most 4.48% of videos needing help are handled fully correctly. Task reminders improve recall but add unwanted responses, and increasing latency tolerance yields only modest recall gains.
In briefTiming Health Messages Within WhatsApp's 24-Hour Window: Evidence from a Nigerian Maternal Health Chatbot
SummaryThe paper analyses 98,607 WhatsApp messages sent to 3,696 mothers by a maternal health chatbot in Nigeria. Free-form messages never failed while the 24-hour customer service window was open, but failures jumped by 93.0 percentage points exactly when it closed. The authors propose window-aware scheduling and compare sending approved templates, waiting, and sending reply prompts through offline evaluation and simulation.
Jeremiah Moyondafoluwa Adepoju, Samuel Omolaiye, Thanni Adewuyi et al.
In briefThink with Skills: pseudocode planning to improve how reasoning models use skills
SummaryThis paper introduces Think with Skills, a method that trains reasoning models to convert relevant skill descriptions into task-specific pseudocode plans before execution. The pseudocode makes operations, conditions, intermediate results, and dependencies explicit. The approach uses supervised fine-tuning for initialization and Group Relative Policy Optimization for end-to-end optimization. On SkillsBench, task success rises from 35.0% to 62.0%.
In briefGraTED: A graph-tensor model decoding passenger EEG to detect hazards in self-driving cars
SummaryThis paper proposes GraTED, a graph-tensor EEG decoding model that predicts passenger risk and identifies danger in autonomous-vehicle scenarios. It represents electrode scalp locations as an undirected graph, using graph convolution for spatial dependencies and low-rank tensor classification to preserve multidimensional time-feature structure. On the PassengerEEG dataset of 11 participants and 45 hours of recordings, GraTED achieves the highest balanced accuracy across all three evaluation settings while using the fewest learnable parameters among seven baselines.
In briefRxnMiner: A multimodal agent that extracts structured reaction data from chemistry figures
SummaryThis paper presents RxnMiner, a multimodal agent system that combines an expert-informed extraction procedure with a chemistry toolbox to read molecular structures, link shared and entry-specific evidence, and build source-linked reaction records from literature figures. The authors also introduce ChemRxn-Bench, a benchmark of 603 reaction figures with 8,547 manually annotated and reviewed records. On this benchmark, with Qwen3.8-Flash-Next as the backbone, RxnMiner raises stereochemistry-aware reaction macro-F1 from 21.0% under direct prediction to 71.0%, versus 25.0% for the strongest existing system.
In briefA reinforcement learning method that schedules multiproduct pipelines in detail while keeping hydraulic accuracy
SummaryMultiproduct pipelines ship refined products to delivery stations in sequential batches, and their detailed scheduling must meet demand under nonlinear hydraulic constraints. The paper proposes an action-decoupled proximal policy optimization framework, casting scheduling as a semi-Markov decision process with flexible action durations and using a cascaded dual-actor design to separate discrete operating durations from continuous flow decisions. Validated on a real multiproduct pipeline in China, it achieves delivery accuracy close to the exact optimum, improves flow rate stability, and substantially reduces solution time compared with a linearized MILP solved by Gurobi.
In briefDisen-Forcing: Separating Semantic Anchoring from Motion in Autoregressive Video Diffusion
SummaryAutoregressive video diffusion models can generate long videos in real time but degrade quickly when inference runs longer than training. The paper finds that initial frames keep long-range consistency mainly through a sparse subset of attention heads, while most heads capture local motion, and that existing frame-sink methods over-condition all heads, causing motion to collapse during distillation. The authors propose Disen-Forcing, a distillation-time plug-in that uses disentangled context routing and motion-disentangling regularization to reinforce this specialization. On long-horizon video benchmarks, it improves both long-range consistency and motion dynamics of base models.
In briefTransformers learning least squares regression in context via masked linear attention
SummaryThis paper constructs a transformer with masked linear self-attention that learns the least squares estimate in a simple regression task within its context. It obtains a closed-form (analytical) solution using layer normalization rather than an approximate gradient descent solution. An experiment accounts for the non-uniqueness of transformer representations and uses this closed-form implementation in a transformer trained with L1 regularization to output the least squares estimate.
In briefUsing LLM agents to find layered governing equations for community-structured network dynamics
SummaryThis paper introduces SAGE, a large language model-guided self-reflective agent framework that discovers governing equations for complex networks. It iterates through a Library-Discovery-Review process, first identifying community-level shared dynamics, then modeling node-specific residual dynamics, and adding missing candidate terms when needed. Tested on two synthetic and four real-world networks, including mobile communication, brain electrophysiology, power grids, and urban rail transit, the authors report compact interpretable equations that outperform representative baselines.
In briefPredictive safety guardrails that foresee an AI agent's actions before they happen do not always make it safer
SummaryThis paper examines world-model-driven predictive safety guardrails, which forecast the consequences of an LLM agent's actions before they are executed. Across seven policy models and two safety environments, the authors analyze how predictions are integrated into decisions, how reliable they are, and whether their content is faithfully used. Stronger reliance on predictions did not yield consistent safety gains, and it could lower task completion while substantially raising inference cost. Safety predictions were also notably less reliable over longer horizons than near-term forecasts.
In briefWhen Tracking Is Affordable: A Movement-Budget Threshold for Dynamic Regret
SummaryThis paper studies dynamic regret for online learners whose total movement is capped by a hard budget. It proves that when the budget is at most a constant multiple of the comparator's path length, every budget-feasible algorithm suffers linear regret, via a deterministic adversarial construction. Conversely, projected online gradient descent with a budget-aware step size achieves sublinear regret once the budget is large relative to the diameter plus path length. The tight minimax rate above the threshold is left as an open problem.
In briefTRIO: A method that refines retrieval-augmented generation tasks with separately optimized agents.
SummaryThis paper proposes TRIO, a retrieval-augmented generation method that helps language models answer questions using external knowledge. Three agents are each optimized separately: one expands ambiguous queries, one judges whether the expanded query is sufficient for retrieval, and one produces the final answer from the evidence. The authors report that experiments on multiple benchmark datasets show TRIO outperforms state-of-the-art approaches.
In briefOffline approximate value iteration: a reinforcement learning method for planning satellite constellation deployment
SummaryThis paper proposes a reinforcement learning approach based on approximate dynamic programming to optimize deploying Walker-type satellite constellations under uncertainties such as launch failures and satellite malfunctions. The problem is modeled as a Markov decision process covering low-injection orbit launches, continuous low-thrust orbit raising, and variable satellite loads. Linear function approximation with basis functions and action pruning are used to reduce the computation and memory needed to find optimal policies. Preliminary results show convergence to policies with lower expected costs and deployment times than a rolling-horizon baseline.
Tristan Leuridan (LGI), Yiping Fang (LGI), Adam Abdin (LGI)
In briefBuilding hybrid interactive story systems with LLMs, from formal description to evaluation
SummaryThis paper presents an LLM-based narrative platform in which interchangeable story managers operate over a shared symbolic world model of facts, operations, and goals. Because each operation has explicit preconditions and effects, sessions can be analysed as plans and assessed with shared indicators of player agency, controllability, author intent, and robustness. The authors implement three contrasting managers and compare them on an interactive Little Red Riding Hood scenario. They find that no single manager performs best on every property.
In briefCARE: Contrastive learning that resists environment changes to predict behavioural intentions
SummaryThis paper introduces CARE, a self-supervised framework that learns representations from trajectories to predict behavioural intentions. It combines temporally asymmetric contrastive views with an IRM-style gradient penalty across environments, plus masked reconstruction and evidence consistency objectives. On an IMPTC-derived four-class task, CARE reports 0.840 accuracy under IID fine-tuning and 0.806 accuracy on a held-out environment.
In briefOmniOPD: Stable On-Policy Distillation for Joint Audio-Video Generation
SummaryThis paper proposes OmniOPD, a modality-wise on-policy distillation framework for joint audio-video generation. It transfers the abilities of specialized audio and video reinforcement learning teachers into a single student model. The method uses asymmetric noise scheduling, grouped multi-noise sampling, and timestep-wise importance weighting to address unstable audio learning caused by sparse audio supervision. Evaluations on JavisBench and VBench show that it preserves audio quality while significantly improving video generation.
In briefSynForecast: A Controllable Library of Synthetic Time Series for Forecasting Research
SummaryThis paper presents SynForecast, an open-source Python library with 32 synthetic time series generators, 19 of which implement mechanisms with published sources in fields such as econometrics and finance. Users can vary perturbations and mechanism families while holding everything else fixed, and the output is reproducible. Evaluated on M4 at five frequencies, a synthetic prior helped a neural network with few observations but hurt it with full history. At the hourly frequency, adding daily cycles reduced zero-shot error by 43 to 63%.
Olivier Sprangers, Cristian Challu, Jan Rathfelder et al.
In briefFormal Analysis of the Narrative Paradox in Interactive Narratives
SummaryThe paper addresses the narrative paradox, the tension between player autonomy and narrative control in interactive stories. It proposes a formal framework using an extended state-transition system, defining properties such as player agency, EM controllability, author intent, and robustness with quantitative indicators. A case study compares different Experience Manager policies, revealing correlations, conflicts, and compromise zones among these properties.
In briefSelfImprove-Bench: Testing Whether Self-Improving AI Agents Generalize
SummaryThis paper introduces SelfImprove-Bench, an evaluation framework for recursively self-improving (RSI) AI systems. It organizes tasks into three capabilities (exploring, planning, and reasoning) and implements them as games for controllability and reproducibility, with scalable generation driven by templates and difficulty coefficients. The framework grades generalization from level G0 to G4 to measure how far improvements transfer from training conditions. Experiments show that the evolutionary gains of existing RSI methods are inconsistent across these generalization levels.
In briefSomaAlign: Learning Reliable Multisensor Representations for Soft Robot Grasping
SummaryThe paper introduces Soma, a dataset of 100 objects, 1,000 grasping trajectories, and 74,914 multimodal samples, plus Soma-Success, a dataset of 998 retained grasp trials. It proposes SomaAlign, a three-stage method that estimates modality reliability, aligns modalities with adaptive temperatures, and enforces cross-modal consistency. SomaAlign reaches 90.66% average Recall@10 across 30 cross-modal retrieval tasks, and frozen encoders with light probes achieve 94.59% post-hoc and 84.07% early grasp outcome accuracy on Soma-Success.
In briefDrawOffer: Evidence-Based Tracking of User Ability for Career Support
SummaryThis paper presents DrawOffer, a system for career-support assistants that separates language-model observation extraction from deterministic inference over source-linked records. It maintains uncertain ability estimates, experience assessments, and revisable issues, and withholds judgments when evidence is insufficient. Evaluated on 102 fictional personas with 580 scored checkpoints, DrawOffer passes 169 of 210 checkpoints on 37 test personas.
In briefPGuard: A built-in, streaming plug-in safety guardrail for multimodal language models
SummaryThis paper introduces PGuard, a safety guardrail built into a language model that checks each generated token as it streams, using the model's own internal representations without changing its outputs. It adds only 0.82-2.26% end-to-end overhead when serving the Qwen3.5-122B-A10B model with SGLang. The authors report state-of-the-art results among native guardrails across more than 30 benchmarks, with an average F1 of 0.836 on text queries using Qwen3.5-9B.
In briefMultiMental-IAS: A Multimodal Driver Mental-State Dataset for Context-Aware Non-IID Evaluation
SummaryThis paper presents MultiMental-IAS, a multimodal dataset of driver mental states collected in a CARLA-based cockpit-in-the-loop simulation. It combines vehicle telemetry with scenario context, EEG and PPG signals from a Muse 2 headband, and cabin video from 23 participants across 148 trials under six induced states. The authors provide a benchmark with a unified evaluation protocol, including participant-disjoint cross-subject folds. For unseen drivers, fusing driving dynamics with normalized EEG/PPG raises macro-F1 from 0.408 to 0.485.
Baran Can G\"ul, Nagendra Pal Singh, Anjana Arvind Naik
In briefPOD-CNF: A generative reduced-order model for parametrized stochastic PDEs
SummaryThis paper introduces POD-CNF, a data-driven generative reduced-order model for uncertainty quantification in parametrized high-dimensional systems. It uses Proper Orthogonal Decomposition to obtain a low-dimensional representation of the solution manifold, then trains a Conditional Normalizing Flow to approximate the density of the reduced coefficients given the input parameters. Numerical experiments show it reproduces output distributions accurately at substantially lower computational cost, and a proof-of-concept application illustrates its use in Bayesian inverse problems.
Andrea Violante, Nicola Rares Franco, Andrea Manzoni
In briefPoint cloud super-resolution using multi-scale spatial priors and local-global modeling
SummaryThis paper proposes a point cloud super-resolution method that reconstructs sparse outdoor LiDAR scans into denser point clouds using range images. It combines multi-scale spatial priors with joint local-global modeling, including a spatially gated fusion module, a multi-granularity local-global perception module, and 2D median-residual denoising. Experiments on KITTI-360 and nuScenes report IoU scores of 0.4703 and 0.3593, respectively.
In briefHLSClaw: Bringing vibe coding to high-level synthesis via a skill-based agent harness
SummaryThe paper presents HLSClaw, the first skill-based agentic harness designed specifically for holistic optimization in high-level synthesis (HLS). An LLM agent uses modular, executable optimization skills and a feedback-driven orchestrator to jointly navigate code restructuring and directive search. On real-world HLS benchmarks, it reports a 39.12x geometric-mean latency speedup, beating the ScaleHLS compiler baseline by 1.98x and Claude Code and Codex by 1.86x and 1.70x.
In briefHistory Forcing: Distilling History-Dependent Video Continuations for Long-Horizon World Models
SummaryThis paper proposes History Forcing, a unified on-policy distillation framework that strengthens historical supervision for few-step autoregressive video generators, helping long videos stay consistent. It decouples supervision using the residual between teacher predictions with and without historical context, then adds the weighted residual to the distillation target. The authors report state-of-the-art results among 1.3B models, with a VBench score of 85.07 for 15s videos and 79.82 for 60s videos, and better scene consistency for out-of-sight content.
In briefLookahead-guided visual refocusing for interleaved multimodal reasoning
SummaryThis paper addresses a limitation of interleaved multimodal chain-of-thought reasoning, where visual evidence chosen from history may lag behind what the next reasoning step needs. The authors propose LPAVR, a training-free test-time framework that uses Multi-Token Prediction's draft tokens, verified in parallel by the target model, as a lookahead signal. The verified prefix guides new visual evidence, which replaces history-based evidence when the latter does not cover the indicated regions. The paper reports improved reasoning accuracy and higher acceptance rates for speculative tokens on multiple multimodal reasoning benchmarks.
In briefStop Handcrafting Engines: AI-Generated, Formally Verified Database Systems
SummaryThis paper presents GenDB-V, a system in which AI agents synthesize database engines tailored to specific hardware and workloads, with correctness formally verified. It formalizes SQL semantics via SQLSpec and jointly synthesizes fast code and formal proofs. On TPC-H and SEC-EDGAR, verified template implementations achieve 22.0 to 61.3 times geometric-mean speedups over DuckDB and Umbra, and a verified general-purpose engine reaches 1.27 to 3.4 times speedups on TPC-H.
In briefSpectral collapse of self-attention under isotropic score geometry
SummaryThis paper studies the spectral structure of self-attention in Transformers using an idealized model of the pre-softmax score matrix. Diagnostics on pretrained vision and language transformers motivate the model. The authors prove that row-wise softmax preserves dominant modes while concentrating the bulk spectrum at zero. Stable rank stays bounded for bidirectional attention but grows logarithmically for causal attention.
SummaryThis paper presents LipLingo, a lip-reading system built on the LipNet architecture with standardized preprocessing, a rotational validation strategy for unseen speakers, and a flexible sentence-level similarity evaluation. The model combines 3D convolutional layers with bidirectional recurrent layers, trained using the Adam optimizer and CTC loss. On the GRID dataset it reports 95.0% character-level accuracy, 87.2% sentence-level accuracy, and a 13.1% word error rate, outperforming the baseline LipNet.
Mahmoud S. Mayaleh, Liban Ali Achkir, Nijat Mehdiyev et al.
In briefRisk-constrained evaluation framework for financial document recognition systems
SummaryThis paper treats financial document recognition as a risk-constrained evidence-recovery problem, arguing that standard character and word error rates do not capture the consequences of errors in financial workflows. It defines typed, evidence-linked financial facts and a multi-layer quality model, and introduces acceptance gates based on upper confidence bounds for critical-field errors. The work is theoretical and presents no empirical benchmark results.
In briefDIKT: A diagnostic, interpretable knowledge tracing framework for prediction and learning analysis
SummaryThis paper proposes DIKT, a knowledge tracing framework that combines response prediction, skill-level diagnosis, and model-based learning analysis. It generates bounded cognitive-state trajectories and decomposes the final response probability into skill-wise probabilities to reveal students' strengths and weaknesses. On three benchmark datasets, DIKT achieves an average absolute AUC gain of 2.04 percentage points over all evaluated baselines.
In briefStage-aware multimodal brain graph contrastive learning for diagnosing neurodegenerative disorders
SummaryThis paper proposes smBGC, a stage-aware multimodal brain graph contrastive learning method for diagnosing neurodegenerative disorders. It uses a stage-aware feature enhancement strategy to capture features specific to each progression stage, and a stage-adaptive brain graph augmentation within a graph contrastive learning framework to fuse multimodal data. The authors report that smBGC outperforms state-of-the-art methods on various diagnosis tasks across two real-world datasets.
In briefUsing Mixture of Task Adapters to Unlock Vision Foundation Models for Multi-Task Dense Prediction
SummaryThe paper addresses conflicts between semantic recognition and geometric inference in multi-task dense prediction with vision foundation models. It proposes D2MoE, which adds task-generic low-rank adaptation and a sparse task-adaptive mixture of experts to a frozen DINOv3 backbone, plus a Semantic Prior Injector that couples task features with frozen foundation priors. Experiments across diverse benchmarks report robust multi-task improvements with a small trainable footprint.
In briefGaussFlaw: Fine-grained diagnosis of rendering flaws in 3D Gaussian Splatting
SummaryThis paper addresses localized rendering artifacts in 3D Gaussian Splatting (3DGS) by introducing GaussFlaw-80K, a large-scale dataset of over 80,000 artifact regions with pixel-level masks and eight flaw labels. It also proposes GeoCritic, a multi-task model that combines RGB appearance with depth-derived cues to localize artifacts and classify their types. The authors further apply targeted refinements guided by the diagnosis, including Gaussian densification, attribute correction, and outlier pruning.
In briefEffComp: A Hybrid Reinforcement and Supervised Learning Method for Efficient Prompt Compression
SummaryThis paper proposes EffComp, a framework that compresses prompts for retrieval-augmented generation (RAG) question answering. It uses BERT-style reranker and sentence selector models to perform fast, sentence-level extractive compression, which avoids hallucinations in the compressed prompts. Training combines reinforcement and supervised learning to optimize the compression ratio while preserving LLM accuracy. Experiments on several QA datasets show better compression-accuracy trade-offs than prior methods with minimal latency.
Romtham Tungsununtham, Piyawat Lertvittayakumjorn, Atiwong Suchato et al.
In briefTrace4D: A method that teaches vision-language models to track 3D object changes over time for grounded reasoning.
SummaryThis paper introduces Trace4D, which trains vision-language models to predict sequences of 4D states describing how objects and their spatial relations change in 3D over time. The model reasons over these state changes together with temporal visual observations, using training on about 600K 4D perception examples and about 70K grounded QA examples. At inference, it predicts 4D states and reasoning directly from video without reference annotations or an external geometric model. Experiments on Trace4D-Bench and VLM4D show improved question-answering accuracy over base vision-language models.
In briefIs your LLM an aggressive narcissist with depression? A study of psychological profiles in large language models.
SummaryThis paper reports a large-scale experiment in which 22 large language models completed 11 psychological inventories. The authors also examine correlations between human and model responses. The results show that models display coherent, relatively stable psychological patterns, with personality traits largely consistent while emotional and symptomatic dimensions vary more, and anxiety and neuroticism emerging as dominant tendencies.
Polina Fedorinina, Anna Fattakhova, Nadezhda Romanova et al.
In briefDINOFAS: Face anti-spoofing via bonafide-centered representation learning
SummaryThis paper presents DINOFAS, a framework for detecting face spoofing that treats the problem as learning representations centered on genuine (bonafide) faces. It combines low-rank adaptation of DINOv3, an attention-based style-content disentanglement module, and a dual-branch design separating prototype-based liveness learning from attack-semantic supervision. Under the standard OCIM leave-one-domain-out protocol it reports an average HTER of 9.09%. When trained only on SynthASpoof and tested on four unseen real-world datasets, it achieves the lowest average HTER (21.41%) and highest average AUC (86.14%) among compared methods.
Shakeel Muhammad Ibrahim, Muhammad Sohail Ibrahim, Muhammad Shahid Jabbar et al.
In briefRotation-robust keypoint matching via pair-relative partial equivariance
SummaryThis paper addresses keypoint descriptors trained on upright images, which often fail under large in-plane rotation. It introduces pair-relative partial equivariance, measured over the empirical distribution of relative rotations in image pairs, and PaSteer, a relaxed descriptor-steering framework with identity-preserving residual and free-form Fourier variants. The authors report that PaSteer consistently improves rotation-robust matching over an exact steerer on frozen DeDoDe descriptors across Roto-360, MegaDepth-1500, AIMS, and IMC-PT, especially under large relative rotations, with minimal upright degradation.
In briefEffects of patch size on time-series foundation models
SummaryThis paper studies how patch size affects patch-based time-series foundation models by comparing encoder-only and decoder-only forecasting models trained on the same data and budget. Across 97 GIFT-Eval tasks, encoder-only models benefit from smaller patches, while decoder-only models perform best at an intermediate patch size. The preferred decoder patch size grows with the forecasting horizon, and Pareto analyses of training compute and CPU latency show decoder-only models as low-cost options while encoder-only models reach the highest accuracy.
Jad Yehya, Andrei-Tiberiu Pantea, Aymeric Dieuleveut et al.
In briefWhat Drives the Sudden Solve? Diagnosing and Steering Hard-Prompt Learning in RLVR
SummaryThis paper studies the sudden "zero-to-one" jump in which hard math prompts, unsolved for many training steps, abruptly become solved under reinforcement learning with verifiable rewards. The authors measure time-to-first-positive and find that failed attempts before success show higher early generation entropy. Based on this, they propose NAPO, which integrates with GRPO by ranking negative rollouts in unsolved groups and assigning graded advantages. Under fixed rollout budgets, NAPO improves Pass@256 by 10.0 points on AIME24 and 6.6 points on AIME2025.
In briefReplay Method for LLM Continual Learning via Multi-Teacher On-Policy Distillation
SummaryThis paper studies how a student language model can keep earlier domain abilities when domains arrive one at a time and earlier expert models are no longer available. It uses replayed prompts from earlier domains, with the previous-stage student as teacher, and compares this against sequential distillation, recovery from stored snapshots, and joint multi-teacher distillation. Experiments use five domains, a 1.5B student, and five public experts. Replay without stored snapshots closes part of the gap to joint training, and a fixed replay share performs about as well as recovery from all snapshots.
In briefSubspace Methods for Transferring 3-D Inspection Models from Synthetic to Real Data
SummaryThis paper studies how subspace representations help models trained on synthetic defects transfer to real 3-D surface inspection, where only scans of good parts are available before deployment. The authors test nearest-subspace classifiers, one-class template subspaces and the generalized difference subspace on MVTec 3D-AD and Real3D-AD scans. They find that a subspace helps under shift in proportion to the shift it spans, and that in reference-based inspection the reference and its registration matter more than the subspace.
Abdulla Nasir Chowdhury, Naish Ibrahim, Gulpi Qorik Oktagalu Pratamasunu et al.
In briefBeyond Monotonicity: Learning Stochastic Online Optimization From Partial Feedback
SummaryThis paper studies stochastic online optimization where, each round, a policy adaptively observes n random variables with unknown distributions and learns only from the observed values. The proposed algorithm runs dynamic programming using the most optimistic expected values over a confidence set, with regret governed by level-set complexity. The authors also prove a matching lower bound and apply the results to several nonmonotone problems.
WhyThe authors drop FSD monotonicity, give regret bounds depending on level-set complexity, and report a matching lower bound across several nonmonotone problems.
In briefLLM-based multi-agent system that coordinates bridge maintenance tasks
SummaryThis paper proposes a knowledge-enhanced method for formalizing autonomy and uses it to build BOMMA, a multi-modal multi-agent system for bridge operation and maintenance. The system combines heterogeneous diagnostic actions, physics-mechanics-informed representations, long-term memory, and structured experiential knowledge in a coordinated workflow. On 3,492 multi-source instructions, it reports scheduling recalls of 85.37% for simple tasks and 99.40% for composite tasks, and a 95.35% BERT F1-score for maintenance action reasoning.
In briefMoviEdit: A framework for editing long multi-shot videos
SummaryThis paper presents MoviEdit, an end-to-end framework for editing long multi-shot videos while keeping edits coherent across scene and viewpoint changes. The authors build about 87K paired editing videos through automatic data-generation pipelines, without manual annotation. The model uses a cascaded bidirectional-autoregressive design that first creates a globally coherent edit draft and then propagates and refines it over the full video. They also introduce MoviEdit-Bench, a benchmark for this task, and report that MoviEdit clearly outperforms open-source models and performs on par with industry frontier models.
In briefChoosing signal features for LSTM-based finger angle estimation from muscle signals
SummaryThis paper studies how to extract features from surface electromyography (sEMG) signals so that an LSTM network can continuously estimate finger-joint angles, for controlling dexterous myoelectric prostheses. It systematically compares window sizes, step sizes, and six time- and frequency-domain features on the Ninapro DB2 dataset. A 300 ms window with a 100 ms step, using either Mean Absolute Value or Root Mean Squared features, gave the best performance, with a test PCC of about 0.779.
In briefAttention residual U-Net for locating polyps in endoscopy images across datasets
SummaryThis paper proposes a polyp detection framework built on an attention residual U-Net to locate lesions in endoscopy images. Rather than regressing bounding boxes directly from global features, it learns lesion-focused spatial representations and suppresses irrelevant background responses before reconstruction. The method was evaluated on multiple public endoscopic datasets captured with different devices, and it reportedly outperforms relevant state-of-the-art methods, with an AUC improvement of up to 22.0%.
In briefAffToM: A Benchmark for Recursive Emotion Reasoning in Multi-Party Interactions
SummaryAffToM is a benchmark with 1,922 test pairs and 19,759 training pairs for testing how well models reason about emotions in social situations. It covers individual emotion inference and multi-party reasoning that tracks participants' nested beliefs and emotions as events unfold. The authors also propose AffToM-R1, a two-stage training method combining supervised fine-tuning with perspective-aware reinforcement learning. Experiments show multi-party reasoning remains hard, with the best evaluated model under 60% accuracy, while AffToM-R1 improves that accuracy by 13.38 percentage points over its base model.
In briefStaged Pandora's Box: Reservation Indices for Multi-Fidelity Bayesian Optimization
SummaryThis paper frames staged search, deciding which candidate to advance and when to stop, as a staged Pandora's box problem with a designated set of collectable stages. For independent finite-stage candidates, it derives an exact covered-call recursion yielding reservation indices, optimal values and losses, and shows that advancing the index-maximal candidate above the incumbent is optimal. The paper also examines information sharing and changes to collectable stages, and checks its representation and bounds with exact finite-law experiments.
Long Vu, Jongseong Brad Choi, Binoy Ravindran et al.
In briefIssueLighter: An LLM Agent That Judges Whether Software Issue Reports Are Valid, Using Issue-Type Awareness
SummaryInvalid issue reports burden maintainers of open source software. The paper manually annotated 520 reports from three projects and identified 13 causes of invalidity tied to issue types. It proposes IssueLighter, an agent-based framework with a type-aware validity assessment strategy. Experiments report an average F1 gain of over 20 percentage points over baselines, along with evidence-backed rationales.
In briefMURFormer: A Lightweight Attention-Based Framework for Denoising PPG Signals
SummaryThis paper proposes MURFormer, a lightweight framework that removes motion-induced noise from photoplethysmography (PPG) signals by mapping noisy recordings to clean ones. It uses a Multi-Scale Attention-Conv Fusion block and a two-pass refinement guided by uncertainty that prioritizes unreliable regions. Evaluated on the CIME dataset and a Daily Activities dataset of 24 subjects, it improves signal-to-noise ratio and reduces heart-rate mean absolute error relative to the raw noisy input.
In briefHuman-supervised AI research in power systems: knowledge, physics, simulation and governance for LLMs
SummaryThis review examines applications of large language models in power-system research alongside physics-grounded power AI studies, and discusses conditions for using such models in science. It proposes a knowledge-physics-simulation-governance framework in which LLMs translate information and coordinate tasks, simulators and validation tools test models, and experts retain authority over key assumptions and final conclusions. The paper also proposes a staged development roadmap and requirements for research reporting.
In briefMicroMouse: Masking Harmful Long-CoT Supervision via Relative Potential
SummaryThis paper proposes MicroMouse, a framework that automatically identifies and masks harmful supervision signals in long chain-of-thought (CoT) data during supervised fine-tuning (SFT). It treats a long CoT as a search trajectory toward the target answer and introduces Relative Potential (REP), which measures how well each reasoning state supports the gold answer by contrasting it with a matched foil. Experiments across multiple student model scales and reasoning benchmarks report gains over Full-CoT SFT and existing selective-SFT baselines, along with shorter output lengths.
In briefCredit assignment over trees for training long-task AI agents
SummaryThis paper addresses reinforcement learning for long-horizon LLM agents, where successful experiences are rare and sparse outcome rewards provide little learning feedback. The authors propose TriGRPO, a tree-based framework that uses hindsight-guided branching and credit assignment within and across rollout trees to turn scarce successes into exploration and learning signals. Experiments report consistent gains in task performance and rollout efficiency across model scales.
In briefTesting Whether AI Reasoning Adapts, Beyond Answers and Efficiency
SummaryThe paper argues that current AI evaluations, based on final-answer accuracy, reasoning traces, process supervision, and efficiency, do not show whether reasoning adapts when tasks or evidence change. It separates across-task adaptivity, where abstractions transfer to new problems, from within-task adaptivity, where hypotheses and actions are revised after new observations or counterexamples. Using ARC-AGI as a diagnostic case study, it proposes intervention-oriented evaluation that tests abstraction transfer and hypothesis revision under controlled changes, with explicit reporting of test-time procedures and budgets.
In briefCorrect ranking, wrong scale: reproducing a robot world model and testing its uncertainty
SummaryThe paper reimplements the dynamics model of the Robotic World Model (RWM) and tests the uncertainty estimate of its follow-up, RWM-U, which penalizes reward in offline policy optimization. On held-out data, training on the model's own rollouts outperforms one-step training at the preregistered long horizon but not on short one-step windows. RWM-U's ensemble disagreement ranks rollout errors well, yet its spread substantially understates error, with one-sigma coverage of 4.6% at 100 steps. The study covers one robot, one gait and one dataset, and trains no policy.
