In briefLearning visual speech features by cross-modal distillation and joint face-lip modeling
SummaryThis paper proposes FLARE, a framework for learning visual speech representations for lipreading. It uses Whisper representations as a teacher signal, converted into discrete targets by a random-projection quantizer, and pretrains with masked prediction. A dual-stream encoder processes both lip regions and full faces. Word error rates are 11.7% on LRS2, 13.2% on LRS3, and 34.3% on WildVSR.
WhyCombines Whisper distillation with dual face-lip streams, a substantive extension of prior work, validated across LRS2, LRS3, and WildVSR with ablations.
In briefTEX-MEX: A Contrastive Learning Method Linking Shape Similarity and Time for Time Series
SummaryThis paper introduces TEX-MEX, a self-supervised learning method that combines structural similarity and temporal closeness into one relationship for time series contrastive learning. It is evaluated on PPG, EEG, ECG, accelerometry, and continuous glucose monitoring data, where the authors report broadly state-of-the-art results against several competing approaches. The authors also train a 560K-parameter PPG foundation model and state they will release it open-source and open-weight.
WhyUnifies structural and temporal similarity into one contrastive relation, with broad multi-modality benchmarks showing strong evidence and meaningful impact on time-series self-supervised learning.
Aditya Radhakrishnan, Maxwell A. Xu, Wanting Mao et al.
In briefPredicting Suspected Overdose Deaths from Investigation Reports, with Evidence
SummaryFinal confirmation of US overdose deaths is often delayed by the complexity of certifying drug-involved deaths. The paper proposes an automated framework that classifies suspected overdoses from death investigation reports and produces explanations guided by expert knowledge. It compares pipeline models that rationalize before predicting with self-rationalizing models, finding that a pipeline with explanations grounded in the reports performs best. The best model outperforms a rule-based system in deployment by 15% and allows expert intervention on final decisions.
Raavi Gupta, Ziqing Yang, Karli R. Hochstatter et al.
In briefTest-Time Forecasting: Recurrent Depth for In-Context Dynamics Prediction
SummaryThis paper asks whether foundation models for dynamical systems can truly learn generalizable in-context forecasting strategies for unseen systems. It shows that looped transformers can emulate dynamical system propagators in context, linking linear attention to (extended) dynamic mode decomposition. The authors introduce a Test-Time Forecasting framework where a one-step forecaster uses gradient descent iterations as recurrent depth. Lightweight models trained on ODEs and PDEs are reported to generalize to out-of-distribution dynamics during autoregressive rollout after one-step training.
In briefTextVerifier: Checking whether generated Chinese characters are drawn correctly
SummaryText-to-image models can produce readable Chinese text that still has small stroke or component errors, which OCR and general vision-language models often miss. This paper presents TextVerifier, which uses a character's Ideographic Description Sequence to judge whether an image crop is a correct rendering of the target character. The authors also introduce TextRenderBench-ZH, a per-character benchmark of 7,358 characters, and report an F1 of 0.844, above the baselines. The score can also serve as a reward to reduce rendering errors during Flow-GRPO post-training.
WhyFormulates rendering verification as target-conditioned, using IDS and hard same-character pairs, usable as RL reward; F1 0.844 beats OCR, VLM, and TextPecker on its benchmark.
In briefLabCompiler: A framework that turns flawed lab protocols into verifiable human-machine workflows
SummaryThe paper introduces LabCompiler, which converts natural-language lab protocols that may contain defects into verifiable workflows executed jointly by humans and machines. A protocol firewall diagnoses and repairs defects before compilation, and a device-agnostic two-level graph separates experimental semantics from hardware implementation. The authors also present a protocol recovery benchmark, and report that LabCompiler outperforms frontier LLMs at protocol error correction.
In briefGraphRAG-LTM: A Layered Threat Model for Attacks on Graph-Based Retrieval-Augmented Generation
SummaryThis paper introduces GraphRAG-LTM, a framework for testing attacks on graph retrieval-augmented generation, which builds answers from knowledge graphs derived from documents. It models three levels of attacker access, from query edits guided by ranking feedback to gradient-based query attacks and deletion of bridge triplets from the evidence graph. The results show that each level degrades answer quality, and bridge-triplet deletion further impairs retrieval of supporting facts.
In briefCalibrating Predictors Using Multiple Noisy Labels
SummaryThe paper studies how to evaluate and recalibrate a fixed binary predictor when each example has several annotator labels that may be incorrect, as in medical annotation and crowdsourcing. It proposes Noise Corrected ECE (NC-ECE), a binned estimator of expected calibration error under class-conditional noise, with non-asymptotic bias bounds, along with algorithms for estimating unknown noise probabilities. Experiments on semi-synthetic benchmarks show that the estimator closely tracks its clean-label counterpart and that the recalibration methods improve calibration against latent labels, while uncorrected recalibration on noisy annotations can make it worse.
In briefDefect detection for circuit boards that works across different lighting and production conditions
SummaryThe paper treats industrial board defect detection under changing lighting, surface contamination, and production line conditions as a domain generalization problem. For the IDA 2026 Track 1 challenge, the authors propose four complementary modules covering robust training, small-defect-aware learning, score calibration with threshold selection, and metric-aligned checkpoint selection. On a single RTX 4090 GPU, the model reaches a Track 1 score of 0.564 and a screening score of 0.696 on the test dataset.
In briefTwo-Stage Residual Learning for Strength-Informed Thermal System Optimisation
SummaryThis paper proposes a two-stage residual learning (TSRL) framework built on artificial neural networks for optimising thermal systems. A base network predicts the process output, and a residual network learns the signed errors left by the base network. The framework was tested on a crude distillation unit and a 395 MW industrial gas turbine, with lower test RMSE reported for both. Each computed solution is labelled as a good or weak modelled region to inform safety-critical decisions.
Talha Ansar, Waqar Muhammad Ashraf, Abdulelah S. Alshehri et al.
In briefRDCompGS: Rate-distortion optimization for compressing 3D Gaussian Splatting with image codecs
Summary3D Gaussian Splatting enables high-quality real-time novel view synthesis but needs substantial storage and transmission. This paper proposes RDCompGS, which performs overall rate-distortion optimization of Gaussian representations for subsequent coding by combining a differentiable coding-rate estimate with rendering distortion. A rate constraint adjustment model further regulates the trade-off under different compression conditions. On three benchmark datasets, it reports average BD-Rate reductions of 37.84% with H.265/HEVC and 49.31% with JPEG XL.
In briefTyped decision models for tabular data, compared with LLMs on robustness
SummaryTyped decision models answer questions from a fixed set of options by returning a probability for each option rather than generating text. The paper tests four such models against three large language models on row classification, fact verification and anomaly detection, using 21 content-preserving rewrites of each request. The decision models match the LLMs at lower cost, and three of them are much more robust to question formulation. However, no tested model is robust to how the table is organized, and at least a third of row decisions change under some rewrite for every model.
In briefA unified framework modeling student answers to questions and tutor-posed subtasks in dialogue on one IRT scale
SummaryThis paper proposes a unified framework that models student responses to questions and tutor-posed math subtasks during tutoring dialogues on a shared item response theory (IRT) scale. The method identifies subtasks in dialogues, rewrites them as self-contained questions, estimates difficulty from text, and regularizes difficulty using tutor moves between consecutive subtasks. Evaluated on two real-world math education datasets with human- and AI-tutored dialogues, the framework outperforms strong baselines on difficulty prediction and response correctness prediction.
Zhangqi Duan, Simon Woodhead, Maya Miller-Vedam et al.
In briefReducing the gap between spatial and space-filling curve neighbors in 3D point cloud segmentation
SummarySpace-filling curves offer a lightweight, accurate way to search for nearest neighbors in 3D point clouds, but they create a mismatch between neighbors along the curve and neighbors in the original Euclidean space. This paper proposes a loss function that reduces the effect of this mismatch by training the model to align features of point pairs that are far apart in 3D space but adjacent along the curve. The method is added to a strong 3D point cloud segmentation model and tested on three datasets. The authors report that it outperforms conventional methods on both indoor and outdoor point cloud data.
In briefRetrieve only when useful: agentic reinforcement learning for selective retrieval-augmented translation
SummaryThis paper introduces SRAT, a framework in which a translation model learns when to translate directly or retrieve external knowledge, what to retrieve, and how to use it, optimized for final translation quality. The authors also propose retriever training aligned with translation quality rather than simple similarity. Experiments on idiom, knowledge-intensive, and document-level translation show SRAT consistently outperforms always-retrieve and never-retrieve baselines.
In briefHyrCap: Ranking-Calibrated Action Proposals for Understanding Events in Video
SummaryThis paper addresses poor localization and low-quality proposals in temporal action detection on long untrimmed videos. It proposes HyrCap, a query-based framework that combines complementary video features and calibrates proposal ranking confidence using cross-representation consensus, localization quality estimation, and redundancy suppression. The calibrated proposals are also used as inputs for LLM-based semantic action understanding, with a staged evaluation protocol. The authors report state-of-the-art results on public benchmarks.
In briefDiagnosing and Correcting Reward Hacking in Verifiable Instruction Following
SummaryThis paper shows that verifiable instruction-following benchmarks can miss whether a response actually answers the request, so some frontier models receive full credit with little substantive content. Using the same verifier as a training reward causes reward hacking: after reinforcement learning, constraint satisfaction rises while response quality falls. The authors propose CALIF, which adds a calibrated likelihood-based bonus to verifier rewards using only the existing frozen reference policy, with no extra judge or learned reward model. On IFBench, CALIF improves constraint satisfaction while keeping response quality near or above pre-RL levels.
In briefHankelKANet: a lightweight network that estimates radio path loss using physics-inspired convolutions
SummaryThis paper proposes HankelKANet, a lightweight neural network for estimating path loss in wireless network planning, with electromagnetic propagation characteristics built into its architecture. It combines a residual UNet backbone with a convolutional Kolmogorov-Arnold network, using kernels parameterized by Hankel function amplitudes. On the simulated RadioMapSeer dataset, it reduces parameters by 78% and RMSE by 16.96% compared with the baseline PEFNet, and it also transfers across datasets.
In briefSpectral World Renderer: Unified 3D Hyperspectral Unmixing
SummaryThis paper presents the Spectral World Renderer (SWR), which combines 3D world generation with hyperspectral unmixing to identify what materials surfaces are made of, even when they share the same colour. A coding agent builds executable 3D worlds rendered with measured spectra, producing hyperspectral images with exact labels and sub-pixel proportions, released as SWR-Bench with eight worlds. SWR decodes each Gaussian of a compact 3D representation into material abundances during rendering, yielding spectra and material maps in one pass.
Chia-Ming Lee, Yu-Jou Hsiao, Ming-Ching Chang et al.
In briefRevisiting Triplet Contrastive Learning for 3D Shape Understanding: Learning Transferable 3D Representations from Language Only
SummaryThis paper argues that supervision from rendered 2D images is unreliable for multimodal 3D pre-training, because synthetic views suffer from domain gap, self-occlusion, and rendering artifacts. It proposes learning 3D representations from language alone, embedding them in hyperbolic space with an entailment loss to preserve the hierarchy between general language concepts and specific 3D shapes. The method, Hyperbolic Contrastive Learning (HyperCL), is reported to improve 3D shape classification by up to 4.6% and cross-modal text-to-3D retrieval by up to 5.5% on two widely used benchmarks and a newly introduced one.
In briefVLRLA: Adapting Vision-Language-Action Models at Deployment via In-Context Reinforcement Learning
SummaryVision-language-action (VLA) models link visual perception and language understanding to robot control, but pretrained VLA policies cannot gradually adapt from experience gathered during deployment. This paper proposes VLRLA, a framework that adds in-context reinforcement learning to existing VLA models, learning from histories of observations, actions, rewards, and termination signals. Model parameters stay fixed while the policy progressively refines its actions and improves task success. The authors study its learning properties in simulation and show it can adapt to unseen scenarios in real-world deployment.
In briefHigh-probability regret bounds for Feel-Good Thompson sampling
SummaryThis paper studies Feel-Good Thompson sampling (FGTS), which adds to standard Thompson sampling an exploration term favoring reward functions with high predicted optimal values. Earlier analyses of its two standard variants only gave regret guarantees in expectation. The authors derive the first high-probability regret bounds for both cumulative and current-step FGTS in contextual bandits with general realizable function approximation. In the d-dimensional linear setting, both variants satisfy a regret bound of O(d√T log(2T) + √T log(1/δ)) with probability at least 1−δ.
WhyGives first high-probability regret bounds for both FGTS variants, matching UCB-style dimension and horizon dependence in linear bandits; theoretical with no experiments.
In briefOne Block Ahead: Decoding the Next Text Block Early in Diffusion Language Models
SummaryThis paper proposes One Block Ahead (OBA), a training-free decoding method that denoises the next block of tokens concurrently with the current one in block-diffusion language models. It promotes the next block's sampler state, including the self-conditioning signal and denoising time. On DiffusionGemma, averaged over ten benchmarks, OBA raises tokens per forward pass by 42.3% and single-request throughput by 16.3%, with a small accuracy drop.
In briefCktHGVAE: A graph-based AI model that learns and designs transistor-level analog circuits
SummaryThe paper proposes CktHGVAE, a heterogeneous graph variational autoencoder for transistor-level analog and mixed-signal CMOS circuits that handles both design and state variables. Without topology-specific tuning, its generated operating states reach median R2 above 0.9 for several amplifier and inverter circuits when compared with SPICE simulations. The model supports latent-space search for designs meeting specifications and can transfer zero-shot to unseen topologies, though transfer quality depends on the pretraining topologies. The authors suggest it could underpin automated design of AMS CMOS circuits.
In briefStructure Before Intelligence: Controlling Which Results Agentic Science Workflows Release
SummaryThe paper addresses agentic scientific workflows that can run to completion yet release wrong results due to silent errors, stale inputs, or faulty shared steps. The authors place a harness around the workflow's DAG that enforces mandatory checks, contains failures, recomputes from tracked versions, adds validators, and releases results only when a risk model estimates low invalidity probability. In 60 held-out paired fault-injection instances, validation provided safety and risk estimation improved efficiency, while the language model showed no measurable benefit.
In briefSKG: A Law-Driven Engine Linking Inference and Induction
SummaryThis paper proposes SKG (Symplectic-Kernel Gradient), a reasoning engine in which laws serve as priors, data supply instances, and inference is geometric evolution under law constraints. It unifies symplectic geometry, GENERIC, kernel methods, RKHS, group-equivariant networks, and constrained mechanics, and defines an octuple structure matching conservation laws, constitutive relations, symmetry groups, and axiom systems. The paper reports theoretical results on metric reconstruction, kernel bandwidth, error decomposition, and extrapolation bounds. It validates these on a 1D damped oscillator and a 2D rotation-invariant system.
In briefHoli-Spatial 360: Turning Sparse Home Panoramas into Annotated 3D Indoor Scenes
SummaryThe paper presents Holi-Spatial 360, a framework that converts sparse panoramas from real-estate listings into annotated multi-room 3D environments. It uses correspondence-guided geometric optimization to build a 3D Gaussian Splatting representation, then restores appearance and completes trajectories with adapted image and video priors. The authors release Holi360, with over 1K scenes, more than 100K instance annotations, 2M spatial QA pairs, and a benchmark for evaluating VLM spatial understanding.
In briefDiscourse-parsing-based, structure-preserving document-level data augmentation
SummaryThis paper proposes a data augmentation framework for documents that preserves discourse structure, combining large language models with document-level discourse parsing. It identifies rhetorical relations between sentence pairs and replaces key phrases with coherent but topically shifted content. Experiments on text summarization and question answering show the augmented data gives complementary supervision and improves combined-data training.
In briefRobota-Embed: A Unified Robot Embedding Model for In-Context Learning and Evolution
SummaryThe paper introduces Robota-Embed, a vision-language-action embedding model that jointly encodes a robot's task, execution stage, and future continuation for retrieval. It outperforms the action-free Gemini Embedding 2 on the authors' benchmark, and removing action input degrades retrieval more than removing vision. The authors apply retrieved experience to policy conditioning (Embed2ICL) and feedback-driven task memory evolution (Embed2Evolve), reporting higher success rates on simulated tasks.
In briefReal-time camera pose estimation made sturdier using insole sensors
SummaryThis paper studies how single-camera human pose estimation fails when the lower body is occluded or briefly leaves the frame. The authors collected a synchronized multimodal dataset and proposed a part-decoupled temporal fusion model that adds plantar pressure and acceleration signals from smart insoles to the lower-body branch. Tests on three real-time visual backbones show improved pose accuracy, smoother motion, and better gait measurements, with the largest gains under occlusion and short visual dropouts.
In briefEfficient Streaming Inverse Text Normalization Using a Pretrained Text-to-Text Model
SummaryThis paper studies streaming inverse text normalization, which converts spoken-form speech recognition output into formatted written text. The authors propose an efficient streaming end-to-end system that adapts a pretrained text-to-text model, using architectural changes, a specialized training strategy, a Read-Tag-Write decoding policy, and inference optimizations. On Vietnamese datasets, it reaches accuracy comparable to non-streaming baselines, outperforms hybrid methods, and meets real-time latency requirements.
In briefTransformer neural operators for digital self-interference cancellation in full-duplex radios
SummaryThis paper treats digital self-interference cancellation in in-band full-duplex radios as an operator-learning problem. A Transformer neural operator learns how the residual self-interference channel evolves from past observations and, with the known local transmit signal, reconstructs the current residual interference. The method is trained without a desired signal and then applied when one is present. Experiments in static and time-varying channels report stable cancellation, with 5 to 15 dB improvement over the considered conventional methods at 30 dB SIR.
In briefState-Matched Sequence Conformal Prediction: choosing past errors from similar states to calibrate prediction intervals
SummaryThis paper proposes SMSCP, a sequential conformal prediction method that builds a state from pre-target information, selects matched past calibration scores by state similarity, and applies the ordinary unweighted conformal order statistic. The authors analyze control of the resulting conformal cutoff under temporal dependence and characterize predictive information the chosen state does not capture. Simulations and experiments on solar irradiance and bike-demand benchmarks report better interval scores than several comparators and coverage closer to the nominal 0.90 level on five of six benchmarks compared with ResCP.
In briefIMU-Bench: A benchmark for wrist motion sensor representations across kinematic, semantic, and clinical tasks
SummaryIMU-Bench is a benchmark of ten tasks drawn from eight datasets, designed to test how wrist-worn inertial sensor data supports multiple tasks and transfers between them. The authors compare six self-supervised methods pretrained on unlabeled real and simulated wrist motion. Their masked autoencoder, InrtlWrist, achieves the best or joint-best frozen-probe results on nine of ten tasks, and fine-tuning beats training from scratch on nine of ten tasks. The study also finds that supervision benefiting one task can reduce performance on others.
In briefAgriJev: Efficient Agricultural Option Scoring with Multi-Scale Adapted Models
SummaryThis paper introduces AgriJev, a family of Qwen3.5-based models with 0.8B, 2B, 4B, and 9B parameters for agricultural decision tasks. Given agricultural text, a judgment question, and candidate definitions, the models return scores and relative probabilities in a single forward call instead of generating an answer. Across 17 external test views, AgriJev-9B reaches 75.44% mean accuracy and outperforms the original generator on 13 views, with median scoring latency about half that of same-scale answer generation.
In briefWhat LALM post-training actually learns about audio understanding
SummaryThis paper examines why post-training improves large audio-language models on audio understanding and reasoning benchmarks. Through controlled experiments, the authors separate genuine audio understanding gains from adaptation to specific question formats. They find that much of the reported improvement persists without informative audio but largely disappears when the question format changes. The authors conclude that post-training often teaches models to exploit task-specific textual cues and answer patterns rather than generalizable audio reasoning.
In briefActICL: Post-training an agent to retrieve its own in-context examples for multimodal tasks
SummaryThis paper introduces ActICL, which trains a multimodal large language model agent to decide what demonstrations it needs and retrieve them progressively using tool-augmented reasoning. The agent identifies information gaps, filters and refines candidate demonstrations, and then passes the final context to an answering model. A two-stage post-training procedure jointly optimizes acquisition and answering. The authors report an average gain of 11.6 percentage points over standard similarity-based retrieval across seven benchmarks and three MLLMs.
In briefEnabling AI Agents to Aggregate Past Events via Executable Memory
SummaryThis paper addresses agent questions that require aggregating events from history, which retrieval alone cannot solve. It proposes TallyMem, which extends Trajectory Memory with a reusable event ledger, model-generated event specifications, and deterministic execution. On 776 questions from 99 held-out MINTEval repositories, it raises Qwen3-32B exact match from 35.2% to 49.0%.
In briefMemCode: Discrete Semantic Representations for Long-Term Agent Memory
SummaryThe paper proposes MemCode, a long-term memory framework for LLM agents that rewrites dialogue into self-contained atomic memory units and quantizes them into a learnable codebook. It adds codebook co-occurrence topology as a retrieval path alongside vector search, and separates online writes from asynchronous offline consolidation. The authors report the best overall accuracy on LongMemEval-S and LoCoMo, with per-dialogue memory construction taking roughly 40 to 130 seconds.
In briefDeBoneDiT: A depth-guided diffusion Transformer that removes bones from chest X-rays
SummaryBones in chest X-rays can obscure lung findings and complicate diagnosis. The paper proposes DeBoneDiT, which uses a depth-aware autoencoder for latent compression and a source-conditioned Brownian bridge diffusion process to suppress bones while keeping anatomy plausible. The authors also release SZCH-X-Rays, a dataset of 741 paired chest X-rays and dual-energy subtraction soft-tissue images. They report superior bone suppression with lower computational overhead.
In briefPFedMA: Choosing the Next Agent in Evolving Federated Multi-Agent Collaboration
SummaryThis paper proposes PFedMA, a federated framework that lets LLM-based agents collaborate over multiple rounds while keeping client data decentralized. It tracks each agent's historical capability and reliability to guide selection, and refines queries based on accumulated feedback. The framework also lets privacy rules evolve in both directions between local and global levels while preserving personalized client constraints. The authors report improved task performance, lower measured privacy leakage, and better robustness in reasoning, privacy-sensitive, and adversarial experiments.
In briefMIPS: Reducing Mode Collapse in RL over Large Outcome Spaces
SummaryThis paper addresses mode collapse, where reinforcement learning with expected-reward maximization concentrates probability on a few outcomes even when many good solutions are needed. The authors propose Marginal IPS (MIPS), which estimates each outcome's marginal probability by combining sampled trajectories with a learned model of trajectories leading to that outcome. On Hypergrids, MIPS stays close to the reward-distribution target as the outcome space grows, while frequency-based IPS degrades. In molecular synthesis and phylogenetic inference, MIPS maintains diverse, high-quality sampling where GRPO and frequency-based IPS collapse.
In briefAdCoEvo: A Co-Evolving Multi-Agent System for Ad Delivery
SummaryThis paper presents AdCoEvo, a multi-agent advertising delivery system in which agents with specialized roles coordinate decisions across different timescales. The system updates its model parameters based on actual delivery outcomes, with planning, adjustment, and diagnosis agents providing mutual feedback to improve strategies over time. Experiments report that, compared with a rule-based baseline, AdCoEvo raises advertiser value by 14.02% and lowers cost per acquisition by 7.07%. Against a fixed-parameter multi-agent system, the gains are 3.96% and 1.24% respectively.
In briefAre We Reading the Right Internal Signals to Detect Prompt Injection?
SummaryThis paper compares five methods that detect prompt injections from a language model's internal state, such as residual-stream probes and attention scores. The methods are tested on Qwen3.6-35B-A3B and Gemma-4-26B-A4B-it under a common indirect injection setup. Methods detecting injection presence separate unsuccessful attacks better, while attention, exposure, and role-confusion methods separate successful attacks better. Attention Tracker performs most consistently on successful injections.
In briefxLightGCN: an explainable graph neural network recommender that identifies the most influential peers
SummaryThis paper presents an exact, graph-based method for generating explanations of recommendation system decisions. The authors argue that explainability is essential for establishing trustworthiness and validating effectiveness in risk-sensitive, user-facing AI such as recommenders. The technique is built on LightGCN-style architectures and is said to provide interpretable explanations along with a theoretically guaranteed measure of explanation fidelity. The paper concludes by discussing limitations and directions for future research.
In briefConvergence of the final iterate in random-reshuffling gradient descent with large steps
SummaryThe paper studies the final iterate of random-reshuffling stochastic gradient descent (RR-SGD), a widely used method for finite-sum optimization. For unconstrained convex problems, it proves an expected objective-gap bound for the last iterate with a constant stepsize that no longer needs to shrink inversely with the number of components. The rate matches existing averaged-iterate results up to logarithmic factors.
In briefMore Than Seeing: Vision as Scalable Anchors in Multimodal Pre-Training
SummaryThis paper studies how visual content affects native multimodal pre-training, which trains one model from scratch on text and images. Controlled ablations replacing image embeddings with content-free padding tokens show that retaining visual content improves text prediction on image-text examples without materially changing text-only loss or aggregate language scores. The authors also evaluate spatial reasoning from text descriptions alone, finding that both conditions improve with scale, with visual gains most sustained in localization, navigation, and symbolic planning.
In briefCombining simulation and step-by-step reasoning so AI models can answer physics questions better and cheaper
SummaryMultimodal language models often must reason about how the physical world behaves to answer questions. This work systematically compares video-based physics simulation with language-based chain-of-thought reasoning on tasks such as fluid dynamics and elasticity. It proposes a cognitively inspired hybrid model that chooses the appropriate method at test time to balance accuracy and computational cost. The model can be trained with a 4B-parameter vision-language model and used to improve much larger models exceeding 3T parameters, such as Gemini 3 Pro.
In briefProgress-Aware Distillation to Curb Stagnation in Small Language Model Agents
SummaryLLM-based agents are powerful but costly in tokens and latency. The paper finds that when agent abilities are distilled into small language models, student agents often stall in unproductive loops of failed strategies during long trajectories, a pattern seen across models and tasks. The authors propose Progress-Aware Distillation (ProD), which detects and penalizes stagnation and encodes the student-teacher stagnation gap as an adaptive margin. Experiments on 8 SLMs from 0.6B to 8B parameters across in-domain and out-of-domain benchmarks show reduced stagnation and higher success rates and efficiency.
In briefLearning Feature Selection via Manifold Deep Unfolding PCA
SummaryThis paper proposes MDUFS-Net, an unsupervised feature selection network that adds an ℓ2,0-norm to principal component analysis. It unfolds an ADMM-based iterative algorithm into an end-to-end trainable network so parameters are learned automatically, and uses manifold optimization for numerical robustness. An enhanced variant, MDUFS-Net+, adds an ℓ0-norm for more discriminative features. Experiments reportedly show both outperform state-of-the-art methods on unsupervised feature selection and perform well in clustering.
In briefEvoProbe: An Evaluation-Driven Framework for Evolving Benchmarks for LLM Agents
SummaryThe paper presents EvoProbe, a closed-loop framework that generates tasks, runs diagnostic evaluation, and regenerates tasks based on feedback to keep benchmarks for LLM agents current. It combines verifiable subtask outcome scoring with multi-judge trajectory evaluation to identify phase-level bottlenecks and success and failure patterns. In two rounds of cybersecurity experiments on eight models and 66 tasks, the updated tasks yielded lower scores, with mean Outcome decreasing by 11.8% and mean Pass^3 by 40.0%.
In briefRecognition Is Not Grounding: A Benchmark for Multilingual Financial Evidence Extraction
SummaryThis paper introduces MULTIFINGROUND, a benchmark testing whether language models can both recognize shared accounting concepts across languages and exhaustively recover the matching numerical facts in multilingual financial reports. It is built from FY2025 disclosures of 90 issuers in English, French, and Greek, aligned through a shared accounting taxonomy. The authors find that concept recognition is highly accurate, yet many correctly recognized groups still miss at least one target fact.
In briefA comparison of four reinforcement learning methods for self-driving car decisions
SummaryThis paper compares four reinforcement learning methods for autonomous vehicle decision-making at an unsignalized intersection. The methods are FVI, RCPI, DQN, and a discrete-action version of D-DDPG, evaluated on optimality, safety, training stability, and cost. RCPI is closest to a reference solution and most stable but least data-efficient, while D-DDPG is most data-efficient but least stable. The authors also offer guidance on choosing a method based on model availability, training budget, and stability needs.
In briefHow LLMs Understand Legal Philosophy: Testing Their Legal-Philosophical Reasoning
SummaryThis paper examines whether large language models can understand legal philosophy and apply it in legal reasoning, a capability the authors call legal-philosophical reasoning. The authors introduce LePhilBench, a benchmark with two evaluation tasks and a manually constructed dataset. Its test set contains 220 U.S. Supreme Court opinions, annotated by trained annotators on fourteen questions designed by a law professor, yielding 9,240 labels.
In briefTesting AI agent skills in base models before post-training
SummaryThe paper introduces ASAP, a lightweight framework that elicits agentic behavior from pretrained base models on end-to-end tasks without converting them into chat-style agents. It adapts existing agent scaffolds using trajectory continuation, short reference prefixes, and minimal recovery mechanisms. On software-engineering and deep-research benchmarks, ASAP distinguishes performance across base models and pretraining checkpoints and exposes execution trajectories for analysis.
In briefVideoHarness-SRSI: Structured Recursive Self-Improvement for Video Understanding Harnesses
SummaryThe paper studies how recursive self-improvement can automatically evolve the executable harnesses that organize LLM information gathering, tool use, reasoning, and memory. The authors first present VHarness-BRSI, finding it competitive but prone to validation overfitting, search-budget dilution, and the early-experience trap. They then propose VHarness-SRSI, which grounds every search decision in transferable, sufficient evidence gathered before committing. On four video understanding benchmarks it improves the equally weighted mean held-out score by 6.2% over VHarness-BRSI and reduces the validation-test gap by 37.9%.
In briefBeyond Trial and Error: Predicting and Auto-Tuning Distributed LLM Training Configurations
SummaryThe paper presents Primus Projection, which predicts distributed LLM training throughput by combining small-scale GPU microbenchmarks with analytical simulation, and can also run entirely on CPU. A Primus Tuning Agent uses this projection as a fast scoring function to search configurations and filter out memory-infeasible ones before using GPU time. Projections stay within about 10% of measured multi-node throughput on Llama 3.1 and Mixtral 8x22B. In a Mixtral 8x22B case study, the agent found a configuration with 27% higher throughput than the published 4-node BF16 reference.
In briefAgentSafeEvo: Adversarial Co-Evolution for Safer LLM Agents
SummaryThis paper introduces AgentSafeEvo, a co-evolutionary reinforcement learning framework that improves the safety of LLM agents through an attacker-defender game. The attacker learns to build adversarial tasks that expose weaknesses in the defender's tool use, while the defender learns to complete tasks safely under such attacks. The authors also propose learnability-guided task weighting, which adjusts each task's influence on updates using defender-reward variability within the task. Experiments reportedly show improved adversarial robustness while general agent capabilities are maintained.
In briefFrom Failures to Learnable Worlds: Policy-Conditioned Navigation Environment Generation
SummaryThe paper introduces WEDGE, a framework that uses a pretrained navigation agent's failure modes to guide generation of training environments. An LLM proposes weakness features from successful and failed rollouts, which are validated and mapped to controllable environmental factors. Across three agents, weakness-conditioned adaptation outperforms matched random and hard-example baselines on ProcTHOR, with gains also extending to HM3D for some agents.
In briefLatent annealed sampling method for inferring sparsely observed random fields
SummaryThis paper proposes latent annealed posterior sampling (LAPS) to infer posterior distributions of random fields from sparse observations. The field is mapped to a lower-dimensional latent space, where a flow matching prior is learned and observations are incorporated progressively via annealing, with a variance-preserving bridge controlling covariance inflation. Experiments report accurate posterior approximation, recovery of multiple modes, and conditional sampling time that stays nearly constant as the grid is refined.</summary>
In briefWinARDet: Compressed Window Auto-Regression for Long-Horizon Vision-Language Grounding
SummaryThis paper addresses a failure in multimodal LLM-based detectors, where in dense scenes they repeatedly output identical bounding boxes. The authors propose Compressed Window Auto-Regression (WinAR), which splits long answer sequences into windows, applies full autoregression only to the current window, and compresses earlier windows into a fixed set of latent tokens. On RSDense, with 200 to 2,000 objects per image, the WinARDet model raises RexOmni F1 (IoU=0.5) from 32.16 to 39.36 and lowers the redundancy rate from 35.48% to 13.00%.
In briefRecursive Formation and First-Class Reuse of Compositional Capabilities
SummaryThe paper presents FormCraft, a framework that treats acquired operations as first-class capabilities that can be recursively composed, directly invoked, and selectively stored in a persistent library after domain-specific verification. It separates capability well-formedness and public callability from later admission to storage, while preserving formation provenance and operational identity. Experiments across continuous, discrete, propositional, and sequence domains show instance-independent reuse and reduced search work when reusing capabilities.
In briefW-JEPA: A self-supervised wavelet-based model for wearable activity recognition
SummaryThis paper presents W-JEPA, a self-supervised framework for recognizing human activities from wearable inertial sensors. It splits sensor windows into wavelet tokens indexed by scale, axis, and time, and trains an encoder to predict masked tokens from a target encoder, learning cross-scale and cross-axis motion structure. Across six benchmarks, it outperforms prior methods under both random-window splits and held-out-subject evaluation, improving average macro-F1 by 8.7 percentage points over the strongest published baseline.
In briefForkGRPO: Efficient Exploration and Fine-Grained Credit Assignment for Diffusion Language Models
SummaryThis paper addresses improving reasoning in diffusion large language models using reinforcement learning. It proposes ForkGRPO, which samples multiple continuations from high-entropy intermediate states along a reference trajectory and assigns credit at the segment level. The method reports the best performance on three of four benchmarks, improving over JustGRPO by up to 1.2 points while using up to 27.8% fewer generated tokens.
In briefVision-language models combine conflicting evidence like biased Bayesian observers
SummaryThis paper studies how vision-language models decide between conflicting text and images. The authors propose a Bayesian framework in which models combine conditionally independent evidence from each modality with prior beliefs, plus modality-specific biases. This two-parameter cue combination model predicts model preferences under unseen input degradations and transfers to new forms of unreliability without refitting.
In briefScaling LLM Inference via Parallelism: Throughput-Latency Trade-off
SummaryThis paper presents a systematic framework for analyzing how tensor, pipeline, and hybrid parallelism trade off throughput against latency in LLM inference. By constructing TTFT-ITL feasible sets and Pareto frontiers, it shows that optimal configurations shift from tensor-parallel-dominant to pipeline-parallel-dominant as workloads become more prefill-heavy. The authors also use goodput to select parallelism under SLO constraints and report that pipeline-parallel configurations scale and meet SLOs more favorably in long-context settings.
In briefActMOF: Using active learning to map how metal-organic frameworks form
SummaryDesigning metal-organic frameworks (MOFs) is limited by sparse experimental data and noisy crystallization outcomes. This paper presents ActMOF, an experiment-based benchmark for comparing search strategies in aluminum-carboxylate MOF synthesis, built from measured microwave syntheses with crystallinity scores from diffraction data. Gaussian process-guided policies ranked among the leading methods compared with tree ensembles and large language models. The approach then guided microwave syntheses of six related aluminum MOFs, and all six yielded crystalline products within 12 experiments each.
Hongrui Sheng, Shutian Liang, Reid A. Coyle et al.
In briefTrain to Stitch: A Topology-Preserving Method for Making Independently Trained Models Work Together
SummaryThis paper studies how two neural networks, such as autoencoders, can be trained independently and still work together without coordination during training. The authors propose topological regularization to align the latent spaces of the two models. They give a conditional theoretical analysis and report reconstruction experiments on multiple datasets, showing a substantially narrower interoperability gap.
Akshay Malhotra, Nicolas Zilberstein, Shahab Hamidi-Rad
In briefKoAGE: Can Large Language Models Know, Learn, and Create Korean Neologisms?
SummaryThe study introduces NeoTrace, an automated framework that collects and dates 4,271 Korean neologisms across genres, and uses it to build KoAGE, a benchmark with three acquisition tasks and two generation tasks. Tested on 21 models, the models defined only 13-30% of cutoff-year words in a five-model temporal analysis, yet could infer meaning from a single usage. Even correct definitions gave little improvement in analyzing how pre-cutoff neologisms were formed. Single models often followed human word-formation patterns, while multi-agent coinage favored abbreviation and depended on how interaction was framed.
WhyIntroduces a rerunnable neologism pipeline and the first benchmark comparing model coinages with attested words, with 21-model experiments separating recall, contextual learning, and coinage.
In briefSERO: Learning from Executable Expert Continuations in Long-Horizon Agent RL
SummaryThe paper proposes SERO, a method for long-horizon agentic reinforcement learning that addresses the limitation of policies only learning from futures they themselves visit. It restores a state reached by the current policy and lets a training-only oracle continue interacting with the real environment, using those downstream observations and oracle actions for learning. Each oracle step is weighted by how compatible the frozen behavior policy is with the preceding oracle actions. Evaluated on ALFWorld and WebShop with Qwen2.5-1.5B and 7B models, the authors report that learning the full executable continuation outperforms supervising only the root oracle action, and that the weighting further improves out-of-distribution performance.
In briefEpiGuard: An Autonomous Agent Platform for Early Warning of Infectious Diseases
SummaryThis paper presents EpiGuard, an autonomous agent platform that monitors official surveillance records and keeps improving itself to give early warnings for over 50 infectious diseases. It changes its analysis strategy depending on how much evidence is available and saves successful data-collection and analysis procedures as reusable standard operating procedures. Tested on public data from the China CDC and U.S. CDC, it improves forecasting performance over state-of-the-art systems by 9.68% and 14.04%, respectively.
In briefSelective next-location prediction using a lightweight confidence metamodel
SummaryThe paper applies MetaModelSelect, a lightweight post-hoc confidence metamodel, to decide when cellular next-location predictions are reliable enough to act on. It learns whether a frozen recurrent network's top-1 prediction is correct using only the model's outputs, without retraining the backbone. On Changchun cellular mobility trajectories, it reports lower risk-coverage metrics than maximum softmax probability and other baselines.
In briefSequential Gaussian process regression for learning from pairwise preference comparisons
SummaryThis paper studies continuous global optimization where only noisy pairwise comparisons, indicating which of two design points performs better, are observed instead of direct objective values. Such comparisons produce non-Gaussian posteriors, so the authors use a moment-matching approximation that gives fast closed-form updates of the model's mean and covariance. They propose DOVE, a fully sequential sampling policy combining an information gain criterion with randomized sampling. Under standard regularity conditions, they prove convergence rates for statistical error and expected simple regret that depend on kernel smoothness.
WhyProposes moment-matching sequential GP updates and DOVE policy, with first convergence rates for this setting; solid theory, narrower practical scope.
In briefThe Sum of All Summaries: PROPALIGN, a Graded Method for Rating Proposition Salience
SummaryThis paper introduces PROPALIGN, a method that estimates how salient a proposition is using multiple summaries, and releases PROPALIGNBANK, a dataset of 138K alignments over 35K propositions. Across diverse spoken and written English genres, the authors find that asking models for direct judgments is inconsistent, while graded, summary-based metrics are more stable. Single models underperform humans, but an ensemble with linguistic features approaches manual annotation quality ceilings.
In briefSTRATA: A Self-Maintaining Experience Graph That Helps Language Agents Improve Over Time
SummaryThis paper proposes STRATA, a memory system for large language model agents that stores past experience in a three-layer graph of skills, workflows, and meta-frameworks. The system keeps itself up to date through several mechanisms, including garbage collection and contrast edges that flag misleading, opposite-direction experiences. Tested on four interactive benchmarks with two frozen backbones, it achieves the best average score while using the lowest latency and fewest environment steps among memory-augmented methods.
WhyThree-layer experience graph with contrast edges for negative transfer is novel; extensive comparisons across four benchmarks, two backbones and seven baselines show strong evidence, with a large Scie
In briefRA²: A Method for LLM Unlearning That Preserves Benign Ability Near the Forget Boundary
SummaryThis paper examines a failure mode that standard large language model unlearning evaluations can miss: models may perform well on ordinary retain prompts but degrade on benign reasoning near the forgotten domain, which the authors call forget-adjacent utility loss. The authors propose RA² (Retain-Anchored Attraction), a latent-space method that redirects forget-token states toward nearby retain-supported representations while preserving higher-layer behavior on retain data. On WMDP and MUSE, RA² achieves higher forget-adjacent utility than strong baselines at similar broad-utility levels, indicating that where forget representations are redirected affects benign behavior near the forget boundary.
In briefixi-sLM: An On-Device Small Language Model for AI Call Agents
SummaryThis paper presents ixi-sLM, a 1.2B-parameter Korean small language model designed to run on mobile NPUs for AI call summary and call answering services. The authors trim the vocabulary by call frequency, retrain the tokenizer on call speech, and apply task-aware mixed-precision quantization calibrated on real calls, yielding an 803MiB model. They report 94% of a state-of-the-art cloud LLM's task-weighted quality and 23% of CPU per-call energy, with deployment across Qualcomm and Apple NPUs serving several million subscribers.
In briefEvoWeave: Reusing Agent Experience by Composing Operations in Graphs
SummaryThis paper presents EvoWeave, an experience memory framework for interactive agents that represents abstract states as nodes and reusable operations as directed edges with input and output contracts. A planner splits a task into a dependency graph, and retrieval with bounded directed search assembles relevant experience fragments. Tested on four benchmarks, five settings, and three language model backbones, it improves AppWorld Normal and Challenge goal completion by 6.99 and 4.99 points over a no-memory baseline. The authors also note tradeoffs in retrieval overhead and conditions for valid reuse.
In briefLearning What to Keep: Transfer-aware personal skill learning from user feedback
SummaryThis paper presents SkillGrain, a framework that learns reusable personalized skills from user feedback. It splits feedback into independent rules, predicts which rules will transfer to future requests, and keeps only those. The method jointly optimizes skill induction and skill-conditioned generation with reinforcement learning, and is tested on a benchmark of five personalized writing tasks. Experiments with three base models in in-domain and out-of-domain settings report the best overall average performance among compared methods.
In briefRobust Explanatory Analysis for Structured Optimization Networks (REASON) in Learning
SummaryThis paper proposes a hierarchical learning framework that resolves residual error in two loops, inspired by System 1 and System 2 thinking. A fast loop uses available evidence to resolve what it can, while a medium-speed loop acquires additional evidence for remaining residual error and measures its value by residual reduction. The authors state that, over nested solution spaces, the medium-loop optimum is no worse than the fast-loop optimum. Experiments reportedly show consistent gains over the fast loop, including under unseen corruptions, at about half the computation of the medium loop.
In briefDeLoop: Using an Agent's Own Repeated Actions as Self-Supervision to Cut Loops in Long Tasks
SummaryThis paper addresses language-model agents that repeat a tool call and receive the same observation during long tasks, a failure it calls a loop. It proposes DeLoop, a self-supervised method that treats a looping action as a negative example and the first later non-looping action in the same rollout as a positive one, trained with direct preference optimization without task rewards, gold answers, or LLM judges. On CodeGym under compression, DeLoop reduces the share of Qwen3-8B tasks with loops from 57.3% to 6.1% and uses about a third as many steps, with similar success. On Qwen3-32B, it lowers loops and steps while raising task success from 37.7% to 44.8%.
In briefWhy one-step prediction accuracy may not reflect how well JEPA world models plan
SummaryThis paper questions the common practice of selecting latent world models by one-step prediction loss. The authors propose a frozen-decoder protocol that reports one-step error in physical units and measure goal-conditioned planning success on the same checkpoints across three continuous-control environments. They find almost no correlation between one-step error and planning success (r = -0.010), and that the relationship between the two varies unstably across environments.
In briefTesting whether compressed memory really helps language models
SummaryThis paper asks whether a language model's compressed memory actually holds useful document information when it answers questions. The authors zero out memory vectors and compare against separately trained models that cannot access stored text or receive a constant input for every document. Across nine systems, some models depend on memory yet still underperform a trained alternative. On TriviaQA, memory taken from another question's passage raised F1 over no document, but the gain does not require the question's own passage.
In briefPolicy-Verified Self-Training: Teaching Agents to Follow Rules Using Verifier Judgments
SummaryRuntime policy verifiers usually only block disallowed agent actions, and their decisions are then discarded. This paper proposes Policy-Verified Self-Training (PVST), where an agent generates its own trajectories, a deterministic gate compiled from a written policy judges each action, compliant trajectories are reinforced, and violations serve as negatives for preference optimization. On held-out AgentDojo tasks under prompt injection, violating trajectories and successful attacks drop substantially, though utility falls on travel tasks.
WhyTurns runtime policy verdicts into training rewards without stronger teachers, reducing violations and attack success on held-out AgentDojo tasks.
In briefHow detector design, model size, and crowd density affect tracking multiple pedestrians in video
SummaryThis paper examines how detector architecture, model size, scene crowding, and inference mode affect tracking of multiple pedestrians in video. The authors compare YOLO11, RT-DETR, and YOLO26 detectors with ByteTrack and BoT-SORT trackers on MOT17, MOT20, and SOMPT22 benchmarks. RT-DETR achieved the highest mean HOTA scores on all three benchmarks, with a larger advantage on MOT20 but at higher computational cost. On MOT17 and SOMPT22, YOLO26 variants, especially YOLO26-L with NMS, offered a better balance of quality and speed, while end-to-end inference did not consistently improve tracking over NMS.
Pedro Manoel Herminio Alves, Dalton C\'ezane Gomes Valadares, Cl\'audio de Souza Baptista
In briefBeyond Corrected Actions: Training LLM Agents with Textual Critiques
SummaryThe paper studies whether LLM agents learn better when expert models provide textual critiques explaining their mistakes, not just corrected actions. It proposes Critique-Augmented Agentic Fine-Tuning (CAFT), which uses expert corrections plus explanations of student errors as fine-tuning targets. Across four student models on τ²-Bench Airline+Retail and WebShop, adding critiques improved performance by 8% on average over action-only distillation, and CAFT also gave a better warm start for reinforcement learning.
In briefSeeing, Maintaining, and Learning: How Memory in Multimodal AI Agents Evolves
SummaryThis paper examines how multimodal AI agents, which process information from several types of data such as images and text, can keep useful memories over long tasks and lifelong use. The authors propose a two-axis framework covering how memory is formed and represented, and how memory management changes with experience, divided into static, dynamic, and self-evolving tiers. Based on this framework, they summarize recurring design patterns and principles, and discuss open challenges and future directions.
In briefDANGERLENS: Probing Hazardous Concepts Still Inside Safety-Aligned Models
SummaryThis paper asks whether language models that have been safety-aligned to refuse harmful requests still internally retain hazardous concepts from earlier training. The authors introduce DANGERLENS, a probe that maps hidden states into vocabulary space and uses blinded semantic judgments to separate harmful goals and means from hazard recognition and safety control. In experiments on three Qwen1.5 scales, checkpoints with prior harmful exposure generally kept greater hazardous concept accessibility after the same alignment, most clearly in intermediate and later layers.
In briefDynamic Graph World Modelling: predicting how graphs evolve from past context
SummaryThe paper introduces DyG-WM, a framework that predicts future node and relation representations in dynamic graphs from temporal context. It combines a historical node-relation context encoder, a path-signature encoder for ordered structural changes, and a dual-level joint-embedding objective that avoids reconstructing the raw graph. Experiments on 12 link-prediction and two node-classification datasets show the method is effective on both tasks.
In briefTail Competition Explains Scaling in Best-of-N Sampling for Verifiable Problems
SummaryThis paper treats Best-of-N sampling as a competition between the extreme reward tails of correct and incorrect responses. Using extreme value theory, it identifies three asymptotic regimes and derives finite-sample scaling laws for representative reward distributions. It also proposes a lightweight tail-tension criterion to estimate when additional sampling may become harmful.
In briefVLX-Go: Robust Robot Tracking of a Target Person Amid Distractors
SummaryThis paper presents VLX-Go, a vision-language waypoint policy that helps a robot follow a designated person through environments containing distractors. The model is trained in stages: vision-language alignment, multimodal reasoning mid-training, and tracking specialization on 19,532 oracle episodes. The full recipe reaches 55.80% distractor-tracking success, compared with 41.54% for the earlier OmTrackVLA model. A preliminary on-policy refinement step raises this to 58.22% and reduces collision rate.
In briefSteering vectors for rival ethical frameworks are interchangeable: a competing-target test for activation steering
SummaryThis paper examines whether activation steering, which adds a vector to a model's hidden states to control behaviour, controls its target specifically. The authors introduce a competing-target test that compares a vector with one built the same way for a competing target, using an equivalence test against a margin. On held-out situations in open-weight models, vectors for Kantian deontology and act utilitarianism both raised their own framework's score and lowered the other's, and neither met the margin. The vectors share most of their direction, so steering may deliver a mixture unless it is tested against the competing target.
Weilun Xu (EPFL), Zipeng Sun (McGill University, Mila)
In briefLong-Term Convergence Behavior of Transformers When Generating Text
SummaryThis paper studies how transformer networks behave when generating text over long periods. The authors show that even a simple three-token, single-layer attention-only transformer can fail to converge. For a broad class of transformers with arbitrary finite depth, layer normalization, MLPs, RoPE encodings, and mixed local and global layers, they prove that each model has a softmax temperature threshold above which its empirical cache distribution converges to a unique limit independent of initial conditions. However, the supremum of these thresholds is infinite, so no single global minimum temperature guarantees convergence for all transformers.
In briefBFHD: A bidirectional method to harmonize heterogeneous clinical assessments
SummaryClinical records often mix measurements from different assessment systems used across centers and time. The paper treats harmonization at the measurement level as a bidirectional recoverability problem and proposes BFHD, a framework that identifies measurements translatable in both directions and separates non-translatable components. Experiments on synthetic data and real clinical assessment pairs report broader feasible harmonization coverage and improved subset stability than baselines.
In briefWhere to Start and How to Move: Curriculum Analysis for GRPO-Trained Long-Horizon LLM Agents
SummaryThis paper studies how starting positions, handoff units, and movement rules affect unassisted task completion when training long-horizon LLM agents with GRPO. It proposes Guide Assisted GRPO, which replays a successful reference prefix before the learner continues. Experiments on BFCL across three model sizes find that starting near completion works better under a fixed turn-level threshold rule. Guided starts also improve AppWorld success on both evaluation splits.
In briefIntegrated Systems for Detecting Novelty in Sensor Networks and Computer Vision
SummaryThis paper studies novelty, or anomaly, detection in two settings: identifying blocked rain gauges in weather sensor networks and handling multiple types of novelty in computer vision. The authors evaluate four existing quality control methods and show their limitations, then propose three new techniques combining probabilistic fault detection with consistency-based diagnosis. They also introduce the NOVEL-SS dataset, derived from Snapshot Serengeti camera trap data, and the DCA baseline system, which detects, characterizes, and accommodates multiple novelties.
In briefIs Smaller Greener? Choosing Right-Sized Language Models for AI Agents
SummaryThis paper examines which language model an AI agent should use to complete tasks reliably while minimising energy, cost and time. It evaluates models per successfully completed task rather than per call, and presents a decision model that identifies crossover task lengths where smaller models stop being the better choice. A pilot study of 15 open-weight models on two multi-step agent tasks with measured GPU energy broadly matched the predicted results.
In briefTarget-token mix may falsely look like layer specialization
SummaryLayer-ablation studies often read a larger early-layer damage effect on one corpus as evidence of task specialization. This paper shows that differences in the target tokens each corpus asks the model to predict, such as whitespace and numbers in math solutions, can create this appearance. Splitting the interaction by token type and recomputing it under a shared target mixture reverses the positive effect in most models, mainly when whitespace is removed. A tokenizer-only screen also anticipates these risks.
In briefSeparating Selection from Redistribution in Training-Data Gate Effects
SummaryThe paper proposes Counterfactual Gate-Support Decomposition (CGSD), a budget-matched four-arm design that splits a training-data quality gate's effect into within-support selection, between-support redistribution, and their interaction. On Korean-English and Korean-Japanese retrieval, only the within-support selection component exceeds its calibrated null band. The authors also characterize conditions under which the design is constructible, including cases where it is undefined.
In briefWhy Agent Teams Fail Across Boundaries: A Failure Taxonomy for Private-State Agent-to-Agent Systems
SummaryThis paper studies how agent-to-agent systems fail when agents act for different principals and hold private information and authority. The authors analyze over 2,500 recorded runs across 488 scenarios and derive A²-Failures, a taxonomy of 20 failure modes in four categories. They find that task phrasing drives discovery and interaction, that agents do not verify peers' factual claims on their own, and that aggregate outcome scores hide coordination and conduct failures, so each agent should be evaluated against its own principal. Run logs are released publicly.
WhyProposes a failure taxonomy for agent-to-agent systems with private states, derived from over 2,500 runs across 488 scenarios; moderately novel, well-evidenced, and useful for the subfield.
Sarath Shekkizhar, Prathyusha Jwalapuram, Zixuan Ke et al.
