In briefHow influence from one LLM agent shifts through a chain of agents
SummaryThis paper studies how evaluative judgments from LLM agents travel through multi-agent workflows before reaching a person. On 32 Taiwan 2025 recall cases, authentic comments shown to one agent moved its stance by +0.171 toward the injected position, and more than half of that shift was removed by the third agent in a chain. The authors find that information architecture, including source access, handoff format, and which model reads untrusted content, governs how far manipulation spreads, and that transcripts do not reveal the attack.
In briefKeep Your Receipts: Reliable Cleanup of Agent-Built Knowledge Graphs
SummaryThe paper addresses how to clean up agent-built knowledge graphs when a source is retracted or overruled, noting that deployed methods delete only directly extracted facts and that rebuilding produces many inconsistent artifacts. The authors propose PolyGraph, which has the constructor emit alternative bundles of source evidence for each artifact, so cleanup becomes set arithmetic without further model calls. A re-derivation audit checks these self-reported receipts, finding that they are more trustworthy when shaped like the narrow claims they support.
WhyNovel receipt-based cleanup of agent knowledge graphs, verified by re-derivation audits across models and domains; impact is mainly confined to knowledge graph maintenance.
In briefECG-RSI: Self-improving ECG models via targeted waveform synthesis
SummaryThis paper proposes ECG-RSI, a recursive self-improvement framework that combines verified electrocardiogram waveform synthesis with feedback-guided intervention search. Across seven benchmarks, three rounds raised Qwen3.5-35B-A3B's mean question-answering accuracy from 34.57% to 56.04% and mean classification F1 from 24.27 to 34.03. Continued feedback correction produced diminishing external gains.
In briefMissPath-FM: Flow Matching with Structured Priors for Incomplete Time Series
SummaryThe paper argues that standard flow matching starts from uninformed Gaussian noise, wasting capacity transporting entries that are already observed. It proposes MissPath-FM, which uses interpolation and Gaussian process posterior priors as structured source distributions, plus a missingness-aware path that allocates transport differently to observed and missing regions. Across 48 evaluation settings, it reports the best result in 45 of them.
In briefKeeping Fine-Tuned Models from Forgetting: A Concept-Guided Regularizer for Efficient Adaptation
SummaryParameter-efficient fine-tuning (PEFT) adapts models cheaply but often causes them to forget pre-trained knowledge. This paper proposes Concept-Guided Retention (CGR), a plug-and-play regularizer for any PEFT method that uses a frozen sparse autoencoder to limit weight updates to task concepts and anchor general concepts in activations. On a CLIP vision benchmark, CGR improves retention at comparable accuracy for every base method.
In briefAdaptive-Group Process Rewards for Training GUI Agents
SummaryThis paper addresses slow learning in reinforcement learning for GUI agents, which receive only sparse outcome rewards. The authors propose a framework that splits long GUI trajectories into semantically coherent subgoal groups and estimates process rewards at both the action and group levels. These process rewards are combined with GRPO's outcome advantage to give finer-grained training guidance. Experiments on process reward benchmarks and downstream RL training are reported to show the framework's effectiveness.
In briefMeasuring how much each demonstration state matters for imitation learning
SummaryImitation learning usually treats all demonstration states equally, even though small action errors matter far more in some states than others. The paper defines state criticality as the worst-case drop in downstream value caused by a bounded perturbation of the expert action in a given state. It proposes SCRIM, which estimates this with an adversarial disturbance policy and a value critic, and uses the scores to reweight behavior-cloning losses and select compact, high-value demonstration subsets. Experiments on simulated and real-world robotic manipulation tasks are reported.
Yusuf Umut Ciftci, Joonwon Kang, Ajay Suresha Sathya et al.
In briefGlobal-to-Local Attention for Detecting Sarcasm in Image-Text Posts
SummaryThis paper proposes GLDA, a framework for detecting sarcasm in social media posts that combine images and text. It first uses a global fusion module to capture high-level semantic interactions between the two modalities, then applies a local gated cross-modal attention mechanism to filter out redundant visual and textual features. Experiments show that GLDA improves sarcasm detection and achieves competitive results compared with existing approaches.
In briefFrom Token Correction to Claim Admission: Training-Free Method to Reduce Hallucination in Vision-Language Models
SummaryLarge vision-language models often generate objects not supported by the input image, a problem called object hallucination. The authors propose EGVCA, a training-free framework that accepts a semantic claim only when it has enough image-dependent support, measured by comparing its likelihood under the actual image and a visual-null input. Experiments on CHAIR, POPE, and MME across three model backbones report substantial reductions in hallucination while preserving caption quality.
In briefA differentiable graph-level metric for 3D scene graph supervision
SummaryStandard 3D scene graph prediction uses local object and relation classification losses, which do not measure the overall semantic and structural distance between a predicted graph and its ground truth. This paper proposes Contextual Hellinger Triplet Geometry, which treats classifier outputs as an aligned probabilistic scene graph field and trains with a differentiable graph-level metric, without changing existing network designs. Experiments on 3DSSG show improved graph-conditioned 3D scene generation, and ablations and user studies report better descriptiveness, relation plausibility, and structural preservation.
In briefPushback in a Haystack: How language models respond to factual errors with ambiguity and little effect
SummaryThis arXiv paper examines whether language models push back when users state false facts, and how they do so. The authors introduce a phrase-level, human-validated labeling framework and apply it to model responses in English, German, Chinese, Hindi, and Yoruba. They find that pushback often uses indirect phrasing mixed with affirmation, more so in Hindi and Yoruba. A human study with 300 participants suggests this ambiguous pushback may fail to correct users' factual misconceptions.
In briefA node embedding framework for model-centric conformance checking in process mining
SummaryThis paper proposes a model-centric framework for evaluating process models discovered from event logs. It compares node embeddings of discovered and ground-truth Petri net models, aligned with Procrustes alignment and measured with a parameterized cosine distance to give a bounded, unified metric. Experiments on nine public event logs and on synthetic logs with controlled noise indicate the method is more robust to log deviations than traditional metrics while remaining sensitive to model degradation.
In briefGene Expression-to-Image Learning: How Spatial Layout Affects Information Fidelity
SummaryThis paper studies Gene Expression-to-Image (GE2I), a method that arranges single-cell RNA sequencing gene values on pixel grids so convolutional neural networks can learn from them. Using full-transcriptome data with a fixed network and training procedure, the authors find a trade-off between spatial smoothness and gene-level fidelity. Assigning multiple genes to the same pixel obscures expression differences and can reduce prediction, while one-to-one gene-to-pixel assignment improves prediction across the tested backbones even as smoothness decreases.
In briefLeanPRM: Using the Lean Proof Assistant to Label Steps in Natural-Language Proofs for Training Verifiers
SummaryThe paper introduces LeanPRM, a framework that uses the Lean proof assistant to automatically label each step of natural-language mathematical proofs as correct or incorrect, without human annotation. It contains 10,634 proofs with 152K labeled steps. A Qwen3-4B verifier trained on LeanPRM raises exact step-label accuracy on a held-out benchmark from 48.2 to 74.7, compared with 60.3 when trained on PRM800K.
WhyUses Lean to auto-label natural-language proof steps without humans; verifiers beat PRM800K-trained ones on several benchmarks, with solid evidence.
Rodney Lafuente-Mercado, Salman Rahman, Jingyan Shen et al.
In briefzPocket: Improved Ligand Pocket Prediction Using Frozen Co-Folding Features
SummaryThis paper introduces PocketWatch, a curated benchmark of unseen ligand binding pockets stratified by sequence, structural similarity, and crypticity. It also presents zPocket, which trains pocket-prediction heads on frozen features from co-folding models. zPocket outperforms baselines across all stratifications, with the best model reaching an overall per-target AUPRC of 0.745. The authors further show gains in centroid recall and explore applications in fragment screening and proteome-scale scans.
Tally Portnoi, Allison M Keys, Benjamin R DiFrancesco et al.
In briefHelpMeOut: A Neuropsychology-Based Benchmark for Spotting Errors in Household Tasks
SummaryHelpMeOut is an embodied AI benchmark for monitoring household task execution and detecting errors as they occur. Drawing on neuropsychology research on everyday action errors, it provides 20,000 paired activity episodes spanning nine error categories. The authors report that learned-monitor, symbolic, change-point, and frontier LLM prompting baselines perform poorly, often missing when execution goes wrong.
In briefSWORD: A scene-grounded distance for judging how close predicted 3D scene graphs are to the true ones
SummaryStandard 3D scene graph training penalizes only exact object and relation label mismatches, so it does not measure how semantically close a wrong prediction is to the ground truth. The paper proposes SWORD, which builds object and relation distance metrics from training annotations and lifts them to Wasserstein distances over node and null-augmented edge marginals. The authors prove SWORD is a strict metric on aligned marginal probabilistic scene graphs and that it upper-bounds decoded semantic error.
In briefMedEvo: Improving Medical AI Reasoning by Repairing Failed Reasoning Paths
SummaryMedEvo is a method that uses medical knowledge, including a tag taxonomy and error-type diagnosis, to repair failed reasoning trajectories of medical multimodal models and to select training data. A gold-blind auditor finds the first consequential error and gives a minimal hint, so the model keeps its valid prefix and regenerates the faulty suffix. With supervised fine-tuning alone, Qwen3.5-9B's mean accuracy across seven benchmarks rises from 65.57% to 69.12%, above teacher-distillation SFT (66.93%).
In briefStable token-level scheduling for serverless LLM inference, to keep response generation smooth and fast
SummaryThe paper notes that existing serverless LLM inference methods mainly optimize time to first token and overlook inter-token latency during response generation. It proposes Token Window Pacing (TWP), a closed-loop scheduler that uses runtime feedback on inter-token latency, queue depth, and workload to control request admission at the token level, with chunk-based prefill to reduce interference. Experiments on realistic serverless workloads report up to 45% higher throughput than ServerlessLLM and up to 58% higher than conventional schedulers.
Hardi Ashishbhai Desai, Balaji N K, Sourav Kanti Addya
In briefStoring Context in LLM Weights: Making Encoded Information Usable as Context
SummaryThis paper studies encoding context as temporary weight updates in large language models to avoid the high inference cost of long contexts. Because encoded information mixed with general knowledge may not be used as context, the authors attach an identifier to each encoded context so the model can recite it. They also add a prompting framework that guides the model to use the recalled information as context. The method outperformed baselines using off-the-shelf LLMs across controlled tasks and practical benchmarks.
In briefLatentRiskBench: Measuring State-Dependent Risks in Enterprise Agents
SummaryThis paper addresses safety risks in enterprise agents that depend on workflow state, where the same action can be safe in one context and violate policy in another. It introduces LatentRiskBench, which pairs risky trajectories with benign ones sharing the same terminal action, along with residual risk (RR) metrics. Experiments report that action-level evaluation detects only 0.09 of state-dependent violations, while state-level evaluation reaches 0.96.
In briefCurriculum-Based Surgical Video Generation for Robotic Thyroidectomy
SummaryThis paper proposes a framework for generating surgical videos that combines hierarchical annotation of procedures (phase, task, action) with curriculum transfer learning on a pretrained text-to-video diffusion model. The method is applied to Bilateral Axillo-Breast Approach robotic thyroidectomy, a procedure without an existing curated benchmark. The action-level model reduces FVD by 20.1% and FID by 15.7% relative to the task-level model, and a matched-budget ablation attributes the gains to the curriculum design. A user study with six clinicians rated visual realism highly but noted remaining gaps in anatomical accuracy and clinical reliability.
In briefSTRATA-S12: A Stratified Method for Collecting an Open Sentinel-1/Sentinel-2 Dataset for Self-Supervised Learning
SummaryThe paper introduces Deficit-Driven Stratified Acquisition (DDSA), an online algorithm that orders satellite image acquisitions by how far each land-cover, season, and region stratum is from its quota. Replayed over nearly 2.9 million real attempts, it fills all 192 strata far more evenly than uniform or random ordering. The authors release STRATA-S12, a co-registered Sentinel-1/Sentinel-2 corpus with STAC and PDS4-aligned metadata, and a linear probe confirms it suits self-supervised pre-training.
Christian Sanchez-Leon, Dongming Chen, Dongqi Wang
In briefDynamic K-Head: Adaptive Reasoning-Step Budgets for Continuous Latent Reasoning
SummaryThis paper introduces Dynamic K-Head, which chooses how many latent updates a continuous reasoning model should use for each question before generation, trained with difficulty scores derived from the question alone. On Llama 3.2 1B with GSM8K, it reaches 54.44% exact-match accuracy at 12.03 average tokens, versus 52.31% at 15.21 tokens for fixed-K=6 CoDI, a 2.12-point gain with 20.9% lower token cost. Across three out-of-distribution arithmetic datasets, it reduces cost by 19.3–25.7% while improving accuracy.
In briefRALI@iKAT 2026: Retrieval-aware user profile denoising through personalized conversational dense retrieval
SummaryThis paper addresses the problem that user profiles built from conversations become noisy over time, which hurts personalized retrieval. It proposes Personalized Conversational Dense Retrieval (PCDR), which jointly embeds conversation history and user profile into a latent space designed for retrieval. The method needs only a few personalized relevance annotations for training and is evaluated on the iKAT 2026 tasks as the core retrieval component of a conversational AI agent.
In briefGenerating City-Scale Human Mobility Through Encounters with LLMs
SummaryThis paper introduces EncounterMob, a large language model framework for generating city-scale human mobility data. Unlike prior methods that focus on individual trajectories, it preserves both personal mobility and the interactions that occur when residents meet at the same place and time. Experiments on Japan and NYC datasets report that it outperforms all external baselines on seven encounter metrics and six personal mobility metrics.
In briefCombining 3D posture neutralization and statistical shape modeling for head-neck-shoulder ergonomic design
SummaryThis paper presents a framework that combines posture neutralization with statistical shape modeling for the head-neck-shoulder region. The authors scanned 200 Chinese adults once each in a natural posture and fit a parametric head-neck model to separate cervical pose from anatomy, resetting the pose to a common neutral configuration. The resulting shape space provides a standardized template and population variation modes as baselines for parametric ergonomic design.
In briefCowork-Gym: A framework for synthesizing interactive collaborative work environments at scale
SummaryThe paper introduces Cowork-Gym, which generates collaborative work environments through a build-execute-revise loop that validates and repairs them using execution evidence and review feedback. A stateful user simulator responds to the agent through dialogue and direct edits to a shared workspace. The authors synthesized 64.9K environments across 14 application scenarios and used them to post-train Qwen models of 9B, 27B, and 35B-A3B. The models improved over their base checkpoints by 8.7, 6.4, and 4.1 points on average across six agentic benchmarks.
In briefPrT-Flow: Flow matching that can compute action probabilities for robot reinforcement learning
SummaryPrT-Flow is a flow matching method that computes probability densities over robot actions while keeping multimodal, multi-step generation. It uses an invertible Jacobian-Tractable Transformer and Conditional Euler-step Training. The paper reports 96% success on Robomimic Lift imitation learning and 96% on OGBench Cube-Triple after 600k online steps.
In briefDeltaTTG: Training Multi-Turn Agents with a Token-to-Go Progress Meter
SummaryThis paper addresses the problem that multi-turn agents trained with reinforcement learning receive only sparse terminal rewards, which say little about individual steps. The authors propose DeltaTTG, which learns a reusable progress meter from recorded trajectories of diverse policies, using the number of remaining tokens until task completion (Token-to-Go) as its supervision target. During actor training, changes in the frozen meter's estimates redistribute advantages across steps without altering the reward. The authors report that DeltaTTG improves average scores over GRPO and outperforms compared RL baselines on ALFWorld, PaperSearchQA, and BFCLv3.
WhyProposes a reusable Token-to-Go progress meter redistributing advantages without task rules; evaluated against several RL baselines on three benchmarks with ablation-style analysis.
In briefRoTA-Bench: A benchmark testing whether vision-language models can understand 3D shapes and rotate objects to match a target 2D outline.
SummaryThis paper introduces RoTA-Bench, a benchmark that tests whether vision-language models can build 3D understanding from 2D views and use it for precise control. Models must rotate an object within a limited action budget to match a target silhouette. GPT-5.6-Sol succeeds on only 17% of tasks, while undergraduate participants achieve a 100% success rate, and fine-grained control depends heavily on interaction history.
In briefFixing Final Answers in Multimodal Reasoning Using the Image
SummaryChain-of-thought reasoning on images can drift from the visual content, and the final answer cannot be corrected once the reasoning trace is fixed. The authors observe that the correct option often retains stronger image support in middle-to-deep layers even when a wrong option is committed earlier. They propose TVER (Text-to-Visual Evidence Reallocation), a training-free method that moves bounded attention from answer-related text to question-matched image patches without rewriting the reasoning. The method improves answers across benchmarks and backbones in both multiple-choice and open-answer evaluations.
In briefCan AI agents truly forget? A study of machine unlearning for LLM agents
SummaryLarge language model agents may learn private information during training and reveal it later. The paper argues that removing knowledge from model parameters alone is insufficient, since information can persist in memory or be reacquired through tool calls. The authors introduce AUBench, a controlled benchmark that extends the TOFU fictitious-author setting to an agent environment with persistent memory, an offline web snapshot, and tool-accessible documents. Evaluating existing unlearning methods, it reveals a gap between model-level forgetting and system-level agent unlearning.
In briefDo Generative Models Really Help? Evaluating Them for Rhetorical Discourse Parsing
SummaryRhetorical Structure Theory (RST) parsing has long been dominated by discriminative models built on pretrained encoder-only language models. This paper compares generative language models with traditional discriminative ones on RST parsing. At matched parameter counts, discriminative parsers perform much better, though much larger generative models sometimes outperform them. Generative models have a particular advantage in elementary discourse unit (EDU) segmentation, and the authors report new state-of-the-art results on the RST-DT benchmark for end-to-end parsing and segmentation.
In briefOpen-o3-Video-v2: Video Reasoning Backed by Checkable Spatio-Temporal Evidence
SummaryThis paper presents Open-o3-Video-v2, a video reasoning model that actively acquires missing visual information and records the timestamps and object bounding boxes supporting its answers. The authors build the STGR-v2-CoT135K and STGR-v2-RL-13K datasets and propose the CS-GRPO reinforcement learning method. On the V-STaR benchmark, the 4B model outperforms its Qwen3-VL-4B backbone by 8.5% and 12.6% on two metrics.
WhyCombines active evidence acquisition with verifiable grounding via CS-GRPO, yielding clear V-STaR gains over the Qwen3-VL-4B backbone.
In briefWhy textGrain watermarks are harder to detect, and whether sum tests remain optimal
SummaryThis paper studies when and how efficiently the textGrain watermark for LLM text can be detected after human edits. The watermark's strength is calibrated against sampling entropy, which the authors show shrinks its minimax detectable region compared with Gumbel-max and maximal-coupling schemes. They find the sum detector is optimal only in some regimes and is suboptimal for the optimal error exponent. They also propose a finite-sample weighted multiscale binomial scan that needs no knowledge of next-token distributions, entropy budget, or edit fraction.summary
WhyProves textGrain's minimax detection boundary is strictly smaller than Gumbel-max's, derives optimal exponents, and proposes a weighted multiscale scan detector; impact is mainly within watermark dete
In briefDecouple to Forget: Targeted Unlearning of Sensitive Facts in Large Language Models
SummaryThis paper proposes D2F (Decouple to Forget), a targeted unlearning method for large language models that separates sensitive facts from non-sensitive meaning and surface wording. It extracts a fact-level mediator from each answer, removes only the sensitive parts, and generates counterfactual answers that keep the benign content. Preference-based contrastive optimization is used to reduce information leakage. The authors report that D2F balances forgetting and retention better than strong baselines across various benchmarks.
In briefCode2Worlds-2: A Physics-Grounded Agent Harness for Simulating Embodied Worlds
SummaryThe paper extends Code2Worlds, which generates dynamic 4D scenes from language, so the scenes become executable in a physics engine (MuJoCo). It adds a scene compiler that assigns collision meshes and physical properties, and a simulator-in-the-loop refiner that checks and corrects scenes using simulator feedback. The authors also introduce SimCode4D, a benchmark with five task templates. They report that the resulting scenes are physically plausible, simulator-compatible, and robot-interactable.
In briefDisaggKD: Knowledge distillation that places teacher and student on separate nodes to save GPU memory
SummaryThis paper proposes DisaggKD, which splits the teacher and student models of knowledge distillation across different nodes so that per-GPU memory no longer limits teacher size. Teacher replicas stream hidden states to the student, and a fused KD loss equals the full-vocabulary loss in exact arithmetic without forming full logits. On the Qwen3.5-35B-A3B to 9B pair, the split layout matches the colocated LM loss and runs faster, while on larger pairs colocated runs still run out of memory where the split trains.
In briefShould enterprise policy decisions be learned or reasoned? A controlled comparison on a 27B Korean model
SummaryThe paper compares how an enterprise agent should output its action under a written policy, either as a generated JSON object or a single constrained token, and whether it should reason first. The two interfaces showed no detectable accuracy difference on in-distribution items but differed markedly in restart stability. On a sealed benchmark with labels from a hidden executable policy, reasoning improved exact-action accuracy, while cross-language transfer results were unstable. The model and the re-rendered data with gold labels are released.
In briefBeyond Bias Scores: Answerability Confounds in Activation Steering and a New Direction Design Method
SummaryThe paper argues that the ambiguous-context score in the BBQ bias benchmark mixes whether a model answers with its actual preference. Across 11 categories and five models, contrastive steering often changes answer rates, with many shifts toward the unknown option. The authors propose Gradient-Constrained Direction Design (GCD), which reduces answer-rate distortion and yields larger preference changes under a matched Qwen-7B protocol. They recommend reporting answerability alongside bias scores.
In briefMemAct: A framework for streaming spatial understanding using memory and active exploration.
SummaryMemAct is a framework that lets an embodied agent process incoming observations as they arrive, writing structured memory, retrieving past records and frames when needed, and moving or turning to gather new views. A shared vision-language model drives this loop, so the agent can keep exploring or answer within one reasoning process. The authors built a training set of about 1M examples and MemAct-bench with 20,000 main QA pairs and 2,000 timing questions. Experiments reportedly outperform existing spatial and streaming baselines, approach frontier multimodal models on the benchmark, and keep streaming inference efficient.
In briefFrameLite: A Compact Vision-Language Model for Aesthetic Photo Cropping
SummaryFrameLite is a small photo-cropping model built on Qwen3.5-0.8B that suggests better-composed crops by balancing subjects, context, and image edges. It is trained in three stages: spatial and composition training, Quality-Guided Online Supervision, and Greedy-Referenced Optimization. At inference it returns a single coordinate answer without search or an external scorer. The authors report that it outperforms compared large vision-language croppers on public benchmarks, and plan to release code and weights.
In briefScaling laws for mixed-precision quantized LLM pretraining
SummaryThis paper proposes a scaling law for quantized pretraining that handles mixed precision across forward and backward passes, grounded in optimization theory and validated empirically. It separates quantization's effects on model capacity and optimization efficiency, and explicitly models the persistent optimization error from forward and backward quantization. Experiments on Qwen3 under selected 4-bit regimes, together with published Llama-style results, show up to 11 times lower prediction MSE than existing laws. The law is also extended to layer-wise mixed precision to identify the optimal recipe under a hardware-aware compute budget.
Jinuk Kim, Rajarshi Saha, Muhammed Fatih Balin et al.
In briefTokens Strike Back: Safety-Aware Text Diffusion Denoising
SummaryThis paper proposes SADDER, a training-free method that improves safety guidance for text diffusion models. It uses a per-token safety signal in the embedding space to gate updates from the existing Safety-Aware Denoiser (SAD), so tokens flagged as unsafe are more likely to be revised. The authors derive position-wise target distributions under which the modified denoiser remains exact. Tested on masked and uniform-state diffusion language models in unsafe continuation and jailbreak settings, SADDER is reported to balance safety, generation quality, and inference cost better than existing methods.
In briefCalibrated disambiguation (PML-CD) for learning from partially multi-labeled data
SummaryThe paper addresses partial multi-label learning, where training samples are over-annotated with both relevant and irrelevant labels, as often happens with crowd-sourced data. The authors derive an optimal curriculum weighting function that is equivalent to calibrating the model's predicted confidences. They design a transferable calibrator that takes a histogram of positive samples' confidences and outputs training weights, together with a prototype alignment regularization. Experiments on Pascal VOC, MS-COCO, NUS-WIDE and CUB reportedly show the method outperforming existing state-of-the-art PML approaches.
In briefUsing Executable Prefix Values to Pinpoint Agent Failure Causes
SummaryThis paper examines whether the "first mistake" labels that human or LLM judges assign to failed LLM agent trajectories identify actual causes. The authors restart agents from intermediate states of failed runs and estimate each step's success probability, along with a point of no return. On 150 SWE-bench Verified tasks, they find that judge labels often do not match drops in success probability, and restarting at measured points outperforms restarting at judge labels.
WhyCompares judge-labeled failure steps against executable prefix value curves, showing labels often mark symptoms; real restart experiments on 150 SWE-bench tasks support it.
In briefFrom Pixels to Executable Physics: Behavior-Grounded World Models for Predicting Robot Action Outcomes
SummaryThis paper presents DR-WM, a world model that predicts how a quadruped robot's actions play out on outdoor terrain, where the same command can produce different motion depending on ground contact. The method reconstructs scene geometry from an initial RGB image and adds terrain contact priors calibrated with real-world action-response data, then simulates motion with spatially varying friction. The simulated results guide a video world model's training. Experiments report better performance than geometry-only and uniform-physics baselines on key physical metrics.
In brief4LC: A 4D Latent Cache for Efficient Conditioning in Dynamic World Modeling
SummaryThis paper introduces 4LC, a 4D latent cache that reuses visual representations of static and dynamic scene content as conditions for video generation, avoiding repeated RGB rendering and re-encoding. It combines representations from multiple encoders such as VAE, DINO, V-JEPA, and VGGT, bound to shared geometry and motion, and offers a spatiotemporal interface for reading the cache at specified times and viewpoints. The authors build 4LC-Video on this design, which handles camera control, temporal control, novel-view synthesis, scene revisits, and long-term generation with a single model and no task-specific fine-tuning. Experiments report more efficient condition acquisition, effective spatiotemporal controllability, and scene consistency across tasks.
In briefRELAY: Selective self-distillation with adaptive hand-off for rubric-based RL
SummaryThis paper addresses rubric-based reinforcement learning, where scoring a response against many criteria into one reward loses detail and can give GRPO no gradient when responses tie. RELAY combines selective distillation from a frozen teacher that sees the rubric with GRPO, then uses a plateau detector to remove the teacher and continue with GRPO alone. The authors report the highest in-domain scores among compared methods in all four tested settings.
In briefFedLoRD: Personalized federated learning using the low-rank structure of parameter differences
SummaryThis paper proposes FedLoRD, a personalized federated learning method that exploits the observation that differences between global and personalized model parameters are approximately low-rank in each layer. It formulates local training with nuclear norm regularization on these layer-wise differences and rewrites it in a factorized form so plain stochastic gradient descent can solve it efficiently. Image classification experiments on three benchmark datasets show FedLoRD outperforms several existing personalized federated learning methods.
Tsutahiro Fukuhara, Junya Hara, Hiroshi Higashi et al.
In briefSmartphone-based AI system for inspecting power distribution equipment in the field
SummaryThis paper presents a smartphone-based AI inspection system that assesses distribution infrastructure from photos taken by field personnel, covering pole lean, crossarm tilt, voltage level, vegetation-clearance risk, and transformer ground clearance. It was pilot deployed across five feeders at a power utility in Andhra Pradesh, India, where 1,222 defects were identified and 84 were recorded as rectified. The available data do not show that the system improved reliability. The authors also publicly release the PD-Defect dataset.
Sampath Balaji Masabathula (1, 2), Srija Vanapalli (1 et al.
In briefWhen ASR Errors Matter: Task-Role Diagnostics for Tool-Using Voice Agents
SummaryThe paper introduces SpeechEvalX, a paired diagnostic protocol that assesses how ASR errors affecting different task roles change structured tool calls. Across 594 source utterances, joint call accuracy drops from 82.3% with reference transcripts to 71.9–74.9% across three ASR front ends. Number and negation edits cause 21.9–25.7 point losses, far more than control edits, and detailed task-role labels raise source-grouped ROC-AUC from 0.73 to 0.78 over binary criticality labels.
In briefStateHarness: Keeping Track of Evolving Project State for Long-Horizon Coding Agents
SummaryThis paper presents StateHarness, a runtime framework that helps coding agents maintain an evolving record of a project without updating model parameters. The agent logs its current problem, findings, related code, and unfinished work, and the system surfaces possibly affected callers, tests, and documentation. Across three repository-level evaluations, the framework improves task completion and several change-consistency measures, though pass-to-pass regressions increase on ChainSWE.
In briefBackground2Background: a self-supervised method to remove background from high-dimensional microscopy images
SummaryThis paper presents Background2Background (B2B), a self-supervised framework that removes background in microscopy by exploiting the stronger spatial correlation of background compared with signal. Local pixel shuffling breaks signal correlations while keeping background structure, turning the signal into effective noise for Noise2Noise-based removal. Tested on fluorescence, stimulated Raman scattering and mid-infrared photothermal imaging, it reports over 40-fold background suppression and reveals hidden features.
In briefMAS-Gate: A node-centric method for quantifying multi-agent AI systems across their full lifecycle
SummaryThis paper studies multi-agent systems built on large language models, which split tasks among several agents. It proposes MAS-Gate, a single node-centric quantification method that is applied before execution for workflow design, during execution for runtime editing, and after execution for error attribution. Each run is modeled as a directed acyclic graph, and the method scores nodes using within-node utility and inter-node strength. Experiments report improvements over phase-specific baselines in workflow construction, targeted editing, and error localization.
In briefOpenDepth: Reusing Image-Editing Priors for Robust Open-World Depth Estimation from a Single Image
SummaryThe paper notes that depth estimation methods built on text-to-image models lose accuracy under open-world conditions such as lighting changes and bad weather. The authors propose OpenDepth, which uses image-to-image editing priors to map differently styled images with the same geometry into a unified latent space. It aligns shallow Diffusion Transformer features and adds a joint RGB latent reconstruction task to separate style interference from geometry. The authors report new state-of-the-art zero-shot results on standard and extreme-weather benchmarks.
In briefCo-Evolving Verifiers and Harnesses for Training-Free Self-Improvement on Open-Ended Tasks
SummaryThis paper introduces UCoE, a training-free method that evolves an executable verifier program together with an agent's harness state while model parameters stay fixed. Candidate verifiers are selected by how well they discriminate on response contrasts, and the chosen verifier diagnoses drafts, guides repairs, and decides which response is served. The authors report higher preference rates than direct generation on six open-ended benchmarks and a higher mean score on three verifiable benchmarks.
In briefFacts Matter Less Than Behavior: Why Long-CoT Distillation Works
SummaryThis paper studies why long chain-of-thought (Long-CoT) distillation improves language model reasoning. The authors find that training only on flawed reasoning traces or a restricted prompt set performs nearly as well as standard pipelines. They propose the Behavioral Hypothesis, arguing that distillation mainly aligns models with structured reasoning patterns rather than memorizing facts. They also compare loss functions, finding that a probability-based loss works well for simple patterns while standard cross-entropy loss is essential for complex chains.
WhyProposes a behavioral hypothesis showing flawed or restricted traces match full distillation, challenging rejection-sampling assumptions, though experimental breadth is limited.
In briefPLEDGE: Progressive error-bounded compression for 3D Gaussian Splatting scenes
SummaryThis paper presents PLEDGE, a training-free, progressive codec for 3D Gaussian Splatting that lets users declare tolerances for screen-space position, footprint deformation, view-dependent color, and opacity. These tolerances are converted into per-primitive quantization bounds that hold across a declared camera domain, including unseen viewpoints. Experiments on thirteen scenes confirm the declared bounds and show competitive rate-distortion performance, with the four tolerances carrying different rate costs.
In briefMorphology-aware deep learning for predicting colorectal cancer survival from tissue slide images
SummaryThe paper proposes MorphContrast, a contrastive multiple-instance learning method that uses tissue morphology labels as structural supervision to predict five-year survival from H&E whole-slide images. It was evaluated on TCGA, Alliance/CALGB 89803, and SurGen cohorts, with external SurGen AUCs of 0.669 for the single model and 0.678 for the ensemble. The authors conclude that tissue morphology can guide outcome-aware representation learning in colorectal cancer prognosis.
In briefTDR: A Dataset for Restoring Tagalog Diacritics
SummaryTagalog homographs share spelling but differ in pronunciation and meaning, and diacritical marks (tuldik) that distinguish them are rarely kept in modern text. This paper introduces TDR, an annotated news-based corpus of 40 frequent Tagalog homographs, and a collocation-driven retrieval method to expand the data and reduce class imbalance among minor variants. The authors report that pretrained models lack this capability, but small inexpensive models and fine-tuned LLMs can learn it.
Ivan Yuri De Leon, Lanz Railey Fermin, Zachary Aldridge Africa et al.
In briefTanawpamana: Testing Vision-Language Models on Philippine Heritage Sites
SummaryThis paper introduces the Tanawpamana Dataset, a collection of Philippine heritage site images with metadata drawn from the Talapamana registry of the National Commission for Culture and the Arts. It also presents the Tanawpamana Benchmark, which contains 4,046 tasks for assessing how well models recognize and know these sites. Testing seven vision-language models, the authors found they struggle more with fine-grained distinctions between similar sites than with coarse recognition, and their geographic knowledge is weaker when inferred from images than from text.
In briefTraining General Navigation Robots Without Human Labels via Action Self-Distillation
SummaryThis paper proposes Action Self-Distillation (ASD), a method for teaching robots language-conditioned navigation without human-annotated language or reasoning labels. A single open-source vision-language model serves as both teacher and student: the teacher uses privileged trajectory feedback to infer the navigation goal and reasoning, while the student learns from onboard observations alone to predict future trajectories and describe the scene and plan. Tested on 13 real-world indoor and outdoor navigation tasks, the method improves success rate over the strongest baseline by 30% indoors and on multi-stage tasks, and reaches 95% success outdoors.
In briefFast material parameter inference from multi-view deformation videos using physical fields
SummaryThe paper proposes a single-pass framework that predicts dense physical fields, including force/contact masks, motion-flow fields, and stress-response maps, from multi-view RGB deformation videos. These fields condition the classification of constitutive type and regression of continuous material parameters. On the authors' DeformFieldBench dataset of 5,000 simulations, the method reports the best median errors on all four continuous parameters and about 30 seconds of inference, versus 15 to 120 minutes for per-scene optimization. The authors note remaining domain and identifiability gaps in cross-simulator and real-world tests.
In briefRe-examining the 16x pretraining compute gap between diffusion and autoregressive language models
SummaryA widely cited study estimated that diffusion language models need about 16 times the training compute of autoregressive models to match their likelihood. This paper argues that the original method no longer reflects current diffusion model practice and re-evaluates scaling laws across about 1,700 models using exact likelihood, block diffusion training, and an encoder-decoder architecture. Under this new setup, the compute multiple falls to 12.8x with exact likelihood, and further to between 3.0x and 7.8x depending on block size. The authors conclude that pretraining diffusion language models from scratch is far more viable than previously assumed.
WhyRe-measuring the diffusion versus autoregressive compute gap across about 1,700 models shows the 16x figure substantially overstates it.
In briefControlling softmax in Muon via a unified Hilbert-distance framework for optimizer design
SummaryThe paper argues that Muon's RMS-based control of linear-layer outputs misrepresents changes in softmax output distributions. It proposes Hilbert distance as a metric that bounds cross-entropy changes and probability redistribution. The authors show that this framework unifies RowNorm for the LM head and the query-key steepest descent underlying Compositional Muon. They suggest Hilbert distance as a guiding principle for optimizer design in softmax modules.
Aditya Somasundaram, Charles Guille-Escuret, Alexander Moreno et al.
In briefRAN-JEPA: An Action-Conditioned World Model for Energy-Efficient Mobile Network Control
SummaryThis paper presents RAN-JEPA, an action-conditioned world model for controlling mobile radio access networks (RAN) with energy efficiency in mind. It first learns a predictive latent representation from observation-only network data, then learns action-conditioned latent dynamics from controlled sleep/wake trajectories while keeping the encoder fixed. Evaluated on public operator data and a real-equipment four-cell 5G testbed, it achieves lower action-conditioned prediction error than a GPT-2 world-model baseline and a mechanistic rule. In closed-loop control, it supports higher sleep ratios than both baselines at comparable congestion cost.
In briefTracking label-shifting nodes: a change-aware framework for dynamic graph classification
SummaryEntities in dynamic graphs sometimes switch classes between time snapshots, and these nodes, called shifters, are harder to classify than stable nodes. The authors find that existing dynamic graph neural networks systematically misclassify shifters, keeping them tied to their old class because neighborhood aggregation preserves outdated signals. They propose CHASE, a model-agnostic framework that detects label shifts and overrides stale neighborhood information. CHASE improves shifter performance across all tested models and datasets while preserving performance on stable nodes, and the authors also release three new benchmarks.
WhyProposes a model-agnostic framework for label-shifting nodes in dynamic graphs, with structural analysis and experiments across multiple models and datasets.
In briefRisk-averse diffusion model alignment at inference time, with finite-sample guarantees
SummaryThe paper proposes a risk-averse method for aligning diffusion models at inference time, balancing expected cost against Conditional Value-at-Risk under trajectory-level KL regularization. The alignment problem reduces to a one-dimensional optimization over the CVaR threshold, with conditions for convexity and quadratic regularization enabling derivative-free ternary search. The authors provide finite-sample guarantees and a terminal total-variation bound. Experiments on synthetic data and Stable Diffusion image generation show improved tail performance over risk-neutral baselines with competitive average cost.
In briefSeparating answerability from extraction in LLM radiology report question answering
SummaryThis paper argues that radiology report question answering involves two separate decisions, whether an answer exists and which span gives it, yet RadQA reports them as one F1 score. The authors propose an evaluation that reports false-answer and over-abstention rates alongside F1 and splits F1 changes into extraction and answerability parts. Applied to 13 LLMs on the RadQA test set, models with similar F1 differed more than twofold in false-answer rate.
In briefBenchmarking self-supervised 3D medical imaging models and ranking them before fine-tuning
SummaryThe paper compares 23 self-supervised 3D medical vision foundation models on the same 14 segmentation, classification and regression tasks across several anatomies and imaging modalities. It evaluates transferability estimators and introduces InsightRank, which ranks candidates before fine-tuning using frozen-feature statistics. No model ranked in the top three across all tasks of a type, and a three-model shortlist reduced recorded segmentation job time by 85.8%.
In briefBudget-Limited Distributed Semantic Exploration for Multiple Robots
SummaryThis paper presents a distributed semantic exploration framework for multi-robot teams operating under communication and mobility constraints. It combines Dirichlet mapping, budget-constrained trajectory selection, and Resource-Aware distributed Greedy (RAG) coordination, with each robot choosing paths that maximize expected reduction in Dirichlet entropy. In gridworld ablations, the method improves coverage, semantic mIoU, and map certainty over shortest-distance frontier selection. In photorealistic full-system simulation, it is the only compared method that respects the per-trajectory mobility budget while achieving the highest mean coverage and semantic accuracy.
Changxuan Yao, Sandilya Sai Garimella, Lu Gan et al.
In briefEdgeWise: Running long-task AI agents on local devices by selectively calling cloud models
SummaryEdgeWise is an edge-cloud framework in which a small model on a local device handles most steps of long agent tasks. It delegates only critical steps to a cloud large language model when stronger capability is needed. The authors use step-level reinforcement learning to jointly optimize task solving and delegation decisions, and study when and how to delegate. Experiments report performance matching or exceeding pure cloud-LLM execution, with inference cost reduced by up to 62% on SWE-Rebench and 34% on Terminal-Bench 2.1.
In briefFlashVSR+: Faster, Memory-Efficient Video Upscaling Tuned to Human Preferences
SummaryThis paper presents FlashVSR+, a video super-resolution model that aims to be more efficient and better aligned with human preferences. Its hybrid attention design reportedly achieves up to 2x DiT speedup and 45% lower peak memory, enabling untiled 4K processing on a single 24GB RTX 4090. The authors also propose Reward-Yield Distribution Matching Distillation (Ry-DMD), which incorporates black-box rewards into one-step distillation without backpropagating through the reward model, and show it extends to text-to-image generation.
In briefHow Reference Wording Affects Retrieving Robot Task Instructions
SummaryThis paper studies how the choice of reference descriptions affects identifying a robot task from a short instruction. The authors extract 296 descriptions from 62 RLBench task files and compare five training-free lexical ranking methods. Character n-gram TF-IDF reaches 61.3% mean top-1 accuracy, above BM25 at 55.6% and word TF-IDF at 54.9%. Results vary notably depending on which descriptions are indexed.
In briefMany Steps into One: Limits of Safe Memory Reuse for LLM Agents
SummaryThis paper studies how LLM agents compress many reasoning steps into reusable thought templates, skill libraries, memories, and caches, a practice the authors call reasoning contraction. The authors argue that common success-and-failure counters are poorly tuned change detectors and can keep returning stale answers after facts change. In experiments on multi-hop question answering and code tasks, these counters served many more stale answers than the certified memories the authors propose. The paper also presents a recipe that weights failures by likelihood ratio, certifies memory utility against live reasoning, and restarts certification when facts change.
WhyReframes memory reuse as reasoning contraction, proves counter-based change detectors are poorly tuned, and validates with multi-LLM experiments and a scaling law.
In briefConvex-analytic view of query selection for verifying Markov decision process safety
SummaryThis paper studies safety verification of Markov decision processes, checking whether failure-state reachability probabilities jointly stay in a convex safe region under every scheduler. The authors propose a unified framework that measures violations in an explicitly chosen error unit, scoring directions by a support-function gap. They show the maximal score equals the smallest enlargement of the safe region, in units of that measure, needed to contain the outer approximation. Experiments evaluate how the measurement rule affects verification performance.
In briefTacitVLA: Robot vision-language-action model that reasons without adding delay
SummaryVision-language-action models that reason in text before acting are slowed by the time needed to generate that text. TacitVLA's action expert never attends to the reasoning it generates for the current call, so reasoning supervision improves the backbone representation without delaying actions. Across four benchmarks covering manipulation and autonomous driving, it matches explicit-reasoning counterparts at up to 3.4x lower latency.
In briefEveryone Was Right: Word Representations in Language Models Are Both Abstract and Holistic
SummaryThis paper studies how language models organize lexical representations. The authors split contextual word representations into item, class, and word-class interaction components, analyzing six pretrained transformers across two linguistic categories. They find the interaction component is consistently present, grows across layers, and encodes class information beyond the shared class effect. The interaction also appears in artificial-language learners even when the input contains only additive structure.
Zachary Nicholas Houghton, Yu Zhou, Dan Pluth et al.
In briefCrossPortrait: Animating Portraits Across Species via Adversarial Motion Alignment
SummaryThis paper presents CrossPortrait, a framework that animates a portrait of one species using a driving video of another, including human-to-animal, animal-to-human, and cat-to-dog cases. It splits portrait motion into head poses and expression keypoints, using a species-routed expression encoder with species-specific LoRAs. An expression-level adversarial domain adaptation (Ex-ADA) reduces leakage of driving-species information. The authors report that experiments show expressive, identity-preserving animation that outperforms other state-of-the-art methods.
In briefToward Green AI: How Quantization Affects Energy Use in LLM Inference
SummaryThis paper provides the first unified evaluation of weight-only quantization for large language model inference, measuring energy use alongside accuracy and runtime across different input and output lengths and several GPUs. It tests 10 post-training quantization methods on five models and four benchmarks. Fused-kernel methods such as EETQ int8 and Bitsandbytes int8 achieved the largest energy savings, up to 4x versus FP32 on short text generation, with negligible accuracy loss. The authors also find energy usage closely tracks runtime, suggesting latency optimization can serve as a reliable proxy for reducing environmental impact.
In briefTwinPrune: Density-aware two-phase token pruning for vision-language models
SummaryThe paper proposes a unified spatial-density score to decide which tokens to keep or merge in both the prefill and decode phases of vision-language model inference. The authors argue that decode-pruning claims validated only on prefill-focused benchmarks are misleading, and they evaluate on full-pipeline benchmarks such as TextVQA and MM-Vet. They report a three-way Pareto frontier across compute, memory and accuracy, and state that the code is publicly available.
WhyUnifies a spatial-density score for prefill merging and decode KV eviction, flags an evaluation pitfall, and compares against H2O and SnapKV across several models and benchmarks.
In briefATTUNE: An AI Agent That Automatically Speeds Up LLM Training Setups
SummaryThis paper presents ATTUNE, a coding agent that tunes the configuration of large language model training jobs, such as how models are split across GPUs and how memory and compute are traded off. It works unattended within a GPU-hour budget and uses an external correctness gate to block settings known to change the training objective. The authors report an average 1.7x throughput gain on H100, H200 and B200 clusters and better results than existing automated tuners.
WhyCombines a coding agent with search skills and a correctness gate for recipe tuning, beating Bayesian and random baselines across hardware in thorough comparative experiments.
In briefBox-WAM: Oriented Bounding Box Trajectories as Compact State Representations for World Action Models
SummaryWorld action models (WAMs) are policy models that predict both actions and future world states, and they outperform action-only models on various robotic manipulation tasks. This paper proposes Box-WAM, which replaces dense future video frames with oriented bounding box trajectories of task-relevant objects as a lightweight prediction target. The authors report that box trajectory prediction substantially reduces computational cost while improving action prediction performance.
In briefYUBI-Real2Sim: 3DGS Scene Checks for Trustworthy Robot Policy Testing
SummaryThe paper presents YUBI-Real2Sim, a framework that uses 3D Gaussian Splatting to rebuild real scenes in simulation for evaluating VLA robot policies with dual fisheye wrist cameras. Its pipeline combines 200-degree fisheye capture, 3DGS training, 3D foundation models, and MuJoCo physics, and it reproduces the real-world policy ordering in eight of nine scene-task comparisons. The authors also test whether novel-view quality scores predict ranking reliability, finding the results task-dependent and not yet a general scene-acceptance rule.
Hiroki Nishizawa, Takehiko Ohkawa, Nicola Scianca et al.
In briefGaussian Splatting for Real-Time Scene Reconstruction: Methods, Systems, and Trade-offs
SummaryThis survey reviews research on 3D Gaussian Splatting (3DGS) for real-time scene reconstruction. It groups existing work into five families: single-shot generation, dynamic representations, efficiency and compression, streaming and immersive systems, and Gaussian Splatting SLAM. The authors separate render-only throughput from the costs of generating, updating, tracking, mapping, and optimizing a scene, and compare quality, latency, memory, and update cost trade-offs using sparse-view benchmarks.
In briefC{\AE}NL: Collapse-Based Active Neural Learning with Objective-Aligned Representation Control
SummaryThis paper proposes C{\AE}NL, a framework that separates sample acquisition, Dynamic Collapse Regularization (DCR), and DCR's coefficient policy. DCR penalizes the part of the normalized within-class to between-class scatter residual that exceeds a frozen reference. On ImageNet-100 at 60% annotations, entropy acquisition combined with DCR and MACC-Lite outperforms entropy alone by 1.742 points on ResNet-50 and 1.834 points on ResNet-18, and both gains survive Holm correction. The authors also find that the benefit of geometry-informed regularization depends on the task, and that clean geometry alone cannot guarantee input robustness.
In briefMERIT: Measuring How Much Each Signal Contributes to Multimodal Hearing-Focus Decoding
SummaryThis paper presents MERIT, a multimodal decoder and evaluation method for auditory attention decoding, which aims to identify the talker a listener is attending to so hearing devices can amplify that voice. It shuffles each input stream and uses Shapley values to split the accuracy drop across EEG, eye tracking, head motion, and scene video. On a public dataset, MERIT reaches 0.82 four-way accuracy within listeners and 0.74 on unseen listeners, against a chance level of 0.25. The authors find that adding behavioral streams raises overall accuracy but almost halves the brain's share of it.
In briefCoupled echo state networks show when interaction makes groups smarter than independent members
SummaryThis paper uses a reservoir computing model to study how interaction among group members can improve accuracy and efficiency. Echo state networks were trained independently to identify Japanese speakers from vowel recordings, then tested jointly with and without coupled feedback. Pairs with pooled feedback reproduced the dyadic advantage reported by Bahrami et al. (2010), while simply aggregating independent outputs did not. At equal total computation, larger groups of smaller interacting models outperformed smaller groups of larger models.
Polyphony Bruna, Rick Dale, Michael J. Spivey et al.
In briefA foundation model that predicts next events across different event logs
SummaryThis paper presents the Foundation Event Predictor (FEP), a model that predicts the next event in logs from diverse domains such as business processes, manufacturing, and healthcare without parameter adaptation. It uses attention-based in-context nonparametric inference, with permutation-robust embeddings and quantile transforms to improve generalization across systems. The model was pretrained on 696 real-world event datasets and tested on 50 unseen datasets. Results show it often approaches models trained separately on each dataset and substantially outperforms existing foundation-model baselines.
In briefGiving Language Models a Digital Sense of Proprioception
SummaryThis paper proposes a digital analog of proprioception, the human sense of one's own body position, letting language models access their internal states. The authors train models with SFT and on-policy training to understand subtypes such as routing decision awareness, token entropy, decision depth, and logit lens information. They report that these abilities do not affect model capabilities. The models are then evaluated for situational awareness, legibility, and epiphenomenality to assess how this access changes their behavior.
In briefGrouping Online Learners by Participation Patterns Despite Unequal Observation Periods
SummaryThe paper clusters learning management system records from 16,448 learner-course sequences at a South African public university to identify participation profiles. It finds that most participation measures remain related to observation-window length even as rates, and that course-grouped adjustment reduces this link. The primary consensus yields four profiles, which explain only 2.87% of duration variation, with a mean assignment support of 0.739.
In briefVeriPIE: Checkable Query Plans and Evidence-Guided Selection for Text-to-SQL
SummaryVeriPIE is a neurosymbolic framework that pairs language-model generation with symbolic checks to turn natural-language questions into SQL. It uses QuerySketch, a structured plan language checked by a Lean-based backend, and selects candidates using execution agreement and searches for distinguishing databases. On the Spider development set, it reaches 72.05% test-suite accuracy, above DeepEye's 65.67%.
In briefTime-Synchronous Prequential Evaluation for Predicting Next-Week Student Inactivity from Moodle Logs
SummaryThe study builds a calendar-time synchronised prequential framework that predicts whether students become inactive the following week, using Moodle learning management system logs. It compares SGD logistic regression, OSELM and an echo state network under different threshold policies and feature sets. On 49,329 held-out predictions, the models showed similar discrimination, with each leading on different metrics such as recall, precision or MCC.
In briefTracing the Origin of Unsafe Behavior in Self-Evolving Coding Agents
SummaryThis paper studies how to find which agent and program version first introduced an observed unsafe behavior in a system where coding agents evolve their programs and reuse each other's code, or whether it arose locally. The authors propose combining behavioral fingerprint matching, historical localization, and recursive upstream tracing. On 158 queries across 85 controlled scenarios, the method reaches 96.8% exact attribution accuracy, versus 77.2% for exhaustive fingerprint matching without recursion.
WhyIntroduces origin attribution for unsafe behavior via behavioral fingerprints and recursive tracing, reaching 96.8% accuracy versus 77.2% without recursion across 158 queries.
In briefTorchSpin: A PyTorch Tool for Differentiable Electron Paramagnetic Resonance Simulation
SummaryThis paper presents TorchSpin, an open-source PyTorch framework that ports EasySpin's electron paramagnetic resonance (EPR) calculations into a differentiable tensor-based environment. It covers continuous-wave EPR, ENDOR, pulse EPR, spin dynamics, and related calculations. In matched benchmarks, CPU execution was faster than EasySpin for six of nine workloads. In a Cu(II) powder-spectrum demonstration, feedback through TorchSpin improved a convolutional neural network without parameter labels during refinement.
In briefStructure-preserving fidelity diagnostics for PDE surrogate models
SummaryMachine learning surrogates of partial differential equations are usually trained on numerical solver data, so they may learn discretization error rather than true physics. The paper proposes simple diagnostics that check whether a trained model preserves a known physical property, such as a conserved quantity, symmetry, or Hamiltonian structure, without needing training data or the governing equation. Across five systems, the diagnostics correlate with true dynamical error at up to 0.94 rank correlation, while test loss is uncorrelated or anti-correlated. For Burgers, KdV, and Navier-Stokes equations, the diagnostics distinguish numerical schemes that test loss ranks as equivalent.
In briefREA: From Correct LLM Judgments to Safe Decision Delivery
SummaryThis paper introduces REA (Resolve–Express–Act), a framework that traces how a safety guard's judgment is turned into application actions through parsing and fallback handling. Experiments show that correct predictions can appear in outputs that applications cannot parse, so fallback logic may allow harmful requests or block safe ones. Matching the parser to the output format and fixing formatting reduced the share of harmful requests allowed despite correct guard judgments from 95.5% to 0% in the authors' fail-open replay.
In briefScaling Coding-Agent Teams for Better Inference Compute Use
SummaryThis paper studies how to allocate inference compute across teams of coding agents, treating team design as a compute allocation problem. On ProgramBench, a well-allocated heterogeneous 17-agent team comes within 1.4 points of a strong homogeneous 5-agent team at 70% lower cost. The authors find that delegating planning to workers is most valuable, and that cheap models work well as planners but not as implementers.
Keenan Samway, Daniel D'souza, Omar Darwiche Domingues et al.
