In briefTransformer-based surgical phase recognition for gallbladder removal surgery, with YOLO instrument detection
SummaryThe paper proposes a framework combining a Swin Transformer, a Temporal Transformer Encoder, and a parallel YOLOv11s network to recognize surgical phases and localize instruments in laparoscopic cholecystectomy. Causal temporal smoothing is applied to reduce phase flickering. On the Cholec80 dataset, the method reports 92% phase recognition accuracy, a macro F1-score of 0.86, and mAP@0.5 of 0.863 for instrument detection.
Sheekar Banerjee, Md. Shahabuddin, Md. Mahbub Hasan Akash et al.
In briefDUET4D: A self-correcting method for reconstructing 4D scenes from few camera views
SummaryThe paper tackles 4D scene reconstruction when only a few cameras are available, a setting that is highly under-constrained. It proposes DUET4D, a framework in which a 4D Gaussian splatting representation guides a generative prior so that its outputs are consistent across space and time. The method builds sequential alignments from current Gaussian evidence to steer latent restoration, then uses Sequential Fidelity Resampling to select primitives that are repeatedly supported. The authors report significant improvements in rendering quality and rendering efficiency.
In briefMore Memory, Better Videos? A Controlled Study of Visual Memory in Long-Video Generation
SummaryThis paper examines whether long-video generators that retrieve past visual frames to keep identities consistent actually improve the story. The authors introduce VMem-Bench, which evaluates reference selection on 92 real films and compares four retrieval policies against no-memory and gold-informed controls on 30 generated stories. Retrieval raises entity recall but lowers precision and forbidden-entity avoidance, and it can reinforce appearances that the story has since changed.
In briefReinforcement learning for constrained continuous-time stochastic LQ control: a study of how exploration affects the optimal mean policy when controls must be nonnegative.
SummaryThe paper studies continuous-time stochastic linear-quadratic control with a scalar state and nonnegative control constraints. It introduces an entropy-free auxiliary problem and uses moment reduction and state separation to derive two regional Riccati equations that characterize the finite-horizon value function and optimal conditional mean. For time-invariant coefficients, it characterizes the undiscounted infinite-horizon optimum and proves mean square asymptotic stability under stated assumptions. For unknown dynamics, it proposes model-free actor-critic schemes, and numerical experiments compare learned policies with analytical targets.
In briefWhy LLMs Are Confidently Wrong: Fixing Overconfident Errors by Intervening on Internal Attention Heads
SummaryThis paper addresses the problem of large language models assigning high confidence to incorrect predictions. It proposes CoCHI, a method that identifies and intervenes on internal model components responsible for different confidence behaviours, reducing overconfident errors while preserving correct predictions. Experiments across multiple LLMs on five multiple-choice question answering benchmarks show improved calibration metrics (ECE, NLL, Brier Score) with accuracy maintained.
In briefReason or Recite: Complementary Specialists for Traffic Anomaly Reasoning
SummaryThis paper addresses understanding road anomalies such as collisions and violations in video, which requires verifying events, describing scenes, and explaining causes. The authors propose Reason-or-Recite (RoR), which combines a 9B vision-language model tuned on rewritten structured reasoning traces with a quantized 27B model fine-tuned on answers to match annotation style. A training-free LLM judge selects between the two candidates for each question. RoR ranks first on the in-domain leaderboard of the 2026 AI City Challenge Traffic Anomaly Reasoning track with a mean score of 0.6788.
In briefDiagnosing and Mitigating Long-Session Degradation in Full-Duplex Speech Models
SummaryThis paper studies how full-duplex spoken dialogue models, which handle interruptions and overlapping speech, degrade over long conversations. The authors introduce a closed-loop evaluation framework and find declines in responsiveness, spoken content quality, and voice consistency over time. They propose self-distillation-based LoRA adaptation and a two-resolution context architecture that compresses older history, reducing KV-cache growth by more than 90% across three models.
In briefAskGLP1: A Knowledge Graph Agent Answering Questions About GLP-1 Drug Experiences From Online Patient Posts
SummaryThis paper presents AskGLP1, a system that turns online patient narratives about GLP-1 medications from Reddit and WebMD into a knowledge graph and answers questions about these experiences. An LLM-assisted pipeline grounded in MedDRA terminology distilled 163,135 Reddit posts and 2,506 WebMD reviews into 385,790 drug-side-effect and drug-purpose relations. The authors report that online narratives differ from FAERS reports, emphasizing everyday burdens like fatigue, and that experiences vary by drug and source.
Zhaoqian Xue, Anay Somani, Siddhartha Singh et al.
In briefAdaptive Memory Index Evolution (AMIE): A Plug-in Layer That Updates How Agents Access Stored Memories
SummaryThis paper addresses a problem in long-running AI agents: a stored memory is not always retrievable when needed. It proposes Memory Index Evolution and introduces AMIE, a lightweight plug-in that updates per-memory access state based on retrieval activity, without changing stored content or retraining the host memory system. Across five retrieval-based hosts and three benchmarks, trained AMIE improved downstream F1 or accuracy in all 15 comparisons and retrieval in all 10 comparisons.
In briefRobust camera-based 3D object detection via sparse-recovery model soup
SummaryThe paper addresses the weak robustness of camera-based bird's-eye view detectors such as Fast-BEV to real-world corruptions. It proposes SR-BEVS, which prunes weights by layer-wise magnitude, retrains the resulting diverse sparse models with random learning rates, and averages them into one dense model. Experiments on nuScenes report up to 33.8% better corruption robustness and 1.8% higher clean accuracy than Fast-BEV, with no added inference cost.
In briefSVD-Attention: A Tool for Efficiently Analyzing Attention Blocks in Transformers
SummaryThis paper proposes SVD-Attention, a post-hoc analysis method that decomposes each attention head's query-key interaction matrix and scores its singular modes by spectral magnitude and activation variance on a small calibration set. Across seven pretrained models and five task types, the redundant layers vary by task. Truncating the layers it flags costs at most 1.75 F1, while truncating the same number of layers elsewhere in the same model costs up to 50 F1. Mainly an analysis tool, it reduces inference FLOPs by 0.8-1.5% on its own and by 8-11% when combined with weight pruning, with 0-4% change in task quality.
Debopam Sanyal, Hongjie Chen, Riya Elizabeth Kuriyan et al.
In briefGuarded agent-in-the-loop residual adaptation for changing wind power weather conditions
SummaryThis paper addresses wind power forecasting models whose errors vary across weather regimes such as cold waves, heat-and-calm periods, typhoons, and storms. It proposes a framework that keeps the temporal backbone fixed and adapts only a small set of scene-conditioned residual experts, with an agent identifying the regime and applying corrections only when recent observed errors support them. Across nine backbones and five lead times, the guarded framework lowered both MAE and RMSE in all 45 backbone-horizon combinations. Long-term unattended deployment was not evaluated.
In briefGraph-informed Kriging for sample-efficient Bayesian optimization of wind farm layouts
SummaryThis paper proposes a Graph-informed Kriging framework for optimizing wind farm layouts, which become harder to optimize as turbine numbers grow and high-fidelity aerodynamic models are costly. It represents the farm as a directed aerodynamic graph and builds a farm-scale mean function from pairwise turbine interactions, so the Gaussian process models only the residual structure. Numerical experiments report more accurate annual energy production (AEP) prediction than ordinary and generic-trend Kriging under sparse data, and competitive solutions on the IEA Wind Task 37 benchmark with about 280 AEP evaluations.
In briefBeyond Feature Attribution: Recovering Hidden Interaction Structures in ML Models with Localized CBFI
SummaryThis paper introduces localized case-based feature importance (CBFI), a method for explaining individual predictions on tabular data. Unlike additive methods such as SHAP and LIME, which fold nonlinear interactions into single-feature scores, it splits predictions into structurally distinct parts, separating main effects from pure interactions. Experiments on wine quality classification and medical insurance regression show it recovers interactions hidden by additive models. Fidelity tests and synthetic benchmarks indicate these interactions match nonlinear dependencies in gradient boosting and deep neural networks.
In briefC-SafeGen: Safer LLM generation with claim-level streaming guardrails and certified risk bounds.
SummaryThis paper addresses the problem that large language models can still produce unsafe content, while existing guardrails mainly detect rather than guarantee safe decoding. It proposes Claim-based Stream Decoding (CSD), which uses a streaming guardrail model to check sequential claims and backtracks to revise those flagged as high risk. Using conformal analysis, the authors derive a high-probability upper bound on safety risk, and empirical results compare CSD with state-of-the-art safety decoding methods.
In briefGUIFlow: Faster GUI Agent Inference by Reusing Computation Across Steps
SummaryGUI agents built on vision-language models repeatedly recompute overlapping observations and histories, making multi-step interaction costly. This paper introduces GUIFlow, which reuses and selectively updates computation across interaction steps, using position-aligned historical KV states, change signals, and speculative drafts verified by the target model. The method requires no additional training, no separate draft model, and no changes to model inputs or the action interface. Reported experiments show up to 3.39x geometric-mean speedup on offline multi-image benchmarks and up to 3.49x in online closed-loop interaction on OSWorld, with task performance broadly comparable to the dense baseline.
In briefDecision Model Factory: Building Specialized Self-Controlled Decision Models
SummaryThis paper presents the Decision Model Factory, which uses a short recipe to train, calibrate, verify, and evaluate a chosen backbone model, producing a specialized decision model that the operator controls and deploys. On Japanese decision tasks, compact deployment, and long-context passage ranking, five factory-built models outperform their backbones, and three exceed the best results in a fixed public Japanese benchmark comparison. A dedicated runner computes only permitted-answer probabilities and reuses shared context, more than doubling serving speed on multi-question workloads with nearly unchanged decision scores.
In briefDimFlow: Bidirectional Grid-Expert Learning for Smoother Dynamic 3D Scene Reconstruction
SummaryThe paper introduces DimFlow, a framework that learns compact features for dynamic scenes through bidirectional interaction between a multi-resolution grid and a coordinate-based expert branch. In an aggregate comparison across D-NeRF, HyperNeRF, and DAVIS, it reaches 39.73 dB PSNR, 1.92 dB above 4D-GS, while matching that method's reported 85 FPS rendering speed. The model uses 18.9 MB, compared with 18.3 MB for 4D-GS.
In briefRoAR: Rewarding Agreement Among Correct Reasoning Rollouts to Improve GRPO
SummaryThe paper finds that correct-answer rollouts with lower agreement contain more spurious reasoning than those with higher agreement. It proposes RoAR, which adds a reasoning-level agreement signal to GRPO rewards, computed from existing rollout groups without extra annotation or sampling. On five math benchmarks, the 1.5B GRPO-RoAR model reports 55.5% mean pass@1, and a blind audit compares spurious-reasoning rates across four models.
WhyProposes RoAR, adding rollout agreement to GRPO rewards, supported by blind paired analysis and audits showing gains over several baselines in RLVR training.
In briefMINT: Attention-Based Message Passing to Detect Illicit Accounts in Financial Transaction Networks
SummaryThis paper proposes MINT, a graph neural network framework for detecting illicit accounts in financial transaction networks. It uses attentive message passing to reduce structural inconsistency caused by deceptive connections from malicious actors, and applies Positive-Unlabeled learning to address the scarcity of labeled data. Experiments on three financial transaction benchmarks, two public and one in-house, reportedly show better performance than several state-of-the-art baselines.
In briefCorrelation-aware exploration method for long-trace reasoning
SummaryThis paper addresses policy collapse in reasoning models trained with GRPO on long chains of thought. It proposes Info-GRPO, which augments prompts with latent seeds so the model explores diverse policies correlated with a latent prior, carrying diversity across whole traces rather than per-token randomness. Experiments report that it outperforms vanilla GRPO and entropy-regularized GRPO across several reasoning benchmarks.
In briefLocating Multiple Leaks in Water Networks Using Graph Neural Networks
SummaryLocating several leaks at once in water distribution networks is hard, especially when few pressure sensors are installed. This paper proposes a graph neural network that localizes a variable number of simultaneous leaks without fixing the leak count or enumerating every leak combination for training. The model is tested on the 31-node Hanoi and 92-node NET3 networks under varying demand, sensor coverage, and leak sizes. It outperforms multilayer perceptron, SVM, KNN, and random forest classifiers by up to 12% and 20% in accuracy, respectively.
In briefRelations as Parallel Transport: What Per-Entity Frames Imply for Attention
SummaryThe paper treats the freedom to rotate each entity's coordinate frame as a symmetry and studies what it forces in attention. Under orthogonal transports, invariant scores depend only on norms and the transported inner product, so relations must be supplied explicitly. It also analyzes the minimum real representation dimension needed to compose finite groups exactly, and reports experiments on CLUTRR and StepGame.
In briefReasoning patterns in AI coding models, studied to improve efficiency
SummaryThis study examines the reasoning traces that large reasoning models produce across four representative coding tasks. The authors identify four general reasoning patterns and positive and negative patterns linked to valid and invalid reasoning. They evaluate three applications: task identification, pattern-guided prompting, and pattern-aware early stopping. Early stopping substantially reduces API-reported token usage with only a modest drop in pass@1, and the applications generalize to unseen models.
In briefStop or Clarify? Experience-Conditioned Active Fact Verification
SummaryThis paper examines the intermediate decision in interactive fact verification of whether available evidence is sufficient to stop or whether further clarification is needed. It proposes SEAC, which wraps a fixed verifier with a lightweight evidence-acquisition policy that combines stop risk with a clarification-utility estimate from earlier trajectories. On SciFact, LIAR, and FEVER, SEAC improves mean accuracy and macro-F1 over the compared baselines, though the paired improvement on the small SciFact split is inconclusive.
In briefFed-MAPTrack: Federated Prompt Tuning for Multimodal RGB-X Object Tracking
SummaryThis paper proposes Fed-MAPTrack, a federated learning framework for RGB-X tracking, which combines RGB video with thermal, depth, or event data while keeping raw videos on local devices. The pretrained tracking backbone and prediction head are frozen, and clients only optimize and share prompt-related parameters, with the server aggregating RGB parameters globally and keeping separate banks for each sensor modality. Across LasHeR, VisEvent, and DepthTrack, the method improves mean FedAvg scores by 2.35 AUC, 3.45 SR, and 2.11 F-score points, respectively.
In briefHow Coding-Agent Harnesses Self-Evolve: A Study of Retrospective Adaptation
SummaryThis paper examines whether coding agents can improve their run harnesses (software, configuration, and instructions) using experience from earlier tasks. The authors ran Claude Code on 30 SWE-bench Verified issues, analyzed 125 harness change proposals from a review skill called \retro, and implemented 24 of them. Evaluating 12 changes on later issues, they found that only some changes eliminated or reduced the targeted problem, while others showed no reduction and some added work on unrelated issues.
In briefA Hierarchical Method for Recognizing Environmental Sounds in Complex Scenes: a lightweight two-stage model that sorts sounds into five coarse and 25 fine classes for urban noise monitoring.
SummaryThe paper addresses class confusion in single-stage environmental sound classifiers under imbalanced multiclass conditions. It proposes a lightweight hierarchical model combining a Frequency-Time Adaptive Network (FTANet) with a two-stage classification mechanism: sounds are first assigned to five coarse categories, then routed to category-specific classifiers for 25 fine-grained classes. On a self-built dataset, coarse accuracy is 95.45% with 0.2004M parameters, and on UrbanSound8K the ten-fold mean accuracy is 79.24%. Field tests on the Tianjin Online Acoustic Environment Monitoring Platform report 87.18% overall accuracy for the complete model.
In briefTraining Streaming Video Models to Keep Track of Events, Order, and Counts
SummaryThe paper argues that streaming video models often give the same answer to histories that differ in event order, repetition count, or relations, which it calls state under-specification. It proposes Visible-State Supervision (ViSS), which trains on probes of event identity, order, accumulation, relation change, and answer sufficiency, and Test-Time Causal View Search (TVS), which expands evidence access from recent frames to sparse event and history views while the model stays fixed. Together they reach the best reported accuracy among the compared published systems on StreamingBench and OVO-Bench using far fewer processed frames than exhaustive selection.
In briefRethinking Data Imbalance in Class Incremental Surgical Instrument Segmentation
SummaryThis paper studies how segmentation models for surgical instruments forget earlier classes when new instrument types are added, and how data imbalance biases them. The authors propose SurgCSS, a plug-and-play framework that inpaints realistic surgical backgrounds and blends instrument foregrounds in a class-aware way to balance data. It also adds a Class Desensitization Loss based on contrastive learning and dynamically fuses old and new model weights. Experiments on a newly constructed benchmark built from integrated datasets report significant gains over existing baselines.
In briefPhysEvolve: a self-evolving framework where vision-language models learn physical reasoning from simulator-verified executable worlds
SummaryVision-language models often answer questions about physical dynamics and causality unreliably, partly because counterfactual training data is scarce. This paper proposes PhysEvolve, in which a model proposes executable physics world programs that a deterministic engine simulates into matched baseline and counterfactual branches, producing answer-bound visual tasks. Verifier feedback from solving these tasks updates the shared policy via GRPO and an adaptive curriculum. Starting from Qwen3-VL-8B-Instruct, the authors report gains on PhysBench, CausalPhys, and ContPhy without human-intervention data, with transfer to unseen dynamics and real-world video.
In briefDiffAuction: An Auction-Driven Framework for Generative Image Advertising
SummaryThis paper proposes DiffAuction, a framework that lets multiple advertisers compete for exposure within a single diffusion-generated image. It combines a training-free method for controlling each advertiser's share of the image, an evaluator that uses a vision-language model and a large language model to score click-through intent, and a learnable affine maximizer auction for allocation and pricing. Experiments report that the generation method strengthens the link between target shares and click intent, and the learned auction yields higher platform revenue.
In briefEnd-to-end MXFP8 reinforcement learning on Blackwell: when matching helps, when it's faster, and how it fails
SummaryThis paper implements end-to-end MXFP8 reinforcement learning in verl and tests it on B200 GPUs with Qwen3-8B and Qwen3-30B-A3B. Matching the sampler's quantization grid to the learner's gives lower train-rollout KL divergence than a higher-precision BF16 sampler, but BF16 end to end stays more consistent without a quantized KV cache. MXFP8 only pays off in the dense setting above about 4.5k tokens per micro-batch, and the authors list seven weight-synchronization conditions and seven observed hardware failures. All runs are single-seed and at most 60 steps, so no claim is made about final model quality.
In briefWaypoint: An Analytical Model for Choosing Optimal GEMM Tile Sizes on CPUs
SummaryThe paper presents Waypoint, an analytical model that automatically identifies the optimal tile size for a given hardware and input matrix shape. It estimates performance by accounting for memory hierarchy, instruction-level parallelism, and multi-threading, then searches tile sizes using these estimates. The authors also built OPAM, a CPU GEMM library based on Waypoint, which they report matches state-of-the-art libraries and speeds up LLM inference by 1.36 to 1.98 times.
Srinidhi N, Sunanda Somwase, Hemant Kumar Chandrakant Chandavar et al.
In briefAccuracy and Efficiency of Small Pretrained Models for Flood Surface-Water Mapping
SummaryThis paper evaluates Prithvi-EO-2.0-Tiny for mapping surface water using 446 hand-labeled Sentinel-2 chips from 11 flood events in Sen1Floods11. It compares the model with a U-Net, a larger Prithvi model, and spectral-index thresholds. Fine-tuned Prithvi-Tiny reaches accuracy close to U-Net but does not surpass it, while running faster on a GPU and a laptop CPU. The authors conclude that compact pretrained models offer lower computational demand with competitive accuracy rather than consistent superiority.}
In briefLE-GDA: Training-free few-shot CLIP adaptation combining intermediate and final layer representations
SummaryThis paper proposes LE-GDA, a training-free method for adapting the vision-language model CLIP to few-shot tasks. It fits regularized shared-covariance discriminants to both intermediate and final CLIP representations, standardizes class scores per query, and combines them with the zero-shot classifier without gradient-based optimization. Across 11 datasets, four CLIP backbones, and five shot counts, LE-GDA achieves the highest mean accuracy on every backbone. The authors conclude that intermediate representations carry information beyond the final embedding.
WhyStandardized ensembling of intermediate-layer discriminants is a fresh combination, validated across 11 datasets, four backbones, and ablation controls.
Pavodi Maniamfu, Vogel Kiketa, Zezina Kuzulu Kuetu et al.
In briefSGDB: Method for Detecting Logical Anomalies via Separated Semantics, Geometry, and Evidence
SummaryThis paper proposes SGDB, a method for detecting logical anomalies, where scene components are present but their configuration violates task constraints. It separates object identity recognition from geometric measurement and binds geometric evidence to identities through partial one-to-one association, producing executable checks and traceable alarm explanations. The method reports 98.19% AUROC and 94.52% fixed-threshold F1 on five MVTec LOCO logical categories, and 93.56% AUROC and 92.76% F1 across all 50 VID-AD tasks.
In briefWhat Outcome RL Buys: Finishing, Reliability, and Rare New Reach
SummaryThis paper compares 20 released reinforcement-learning models (1.5B to 32B parameters) with the thinking models they started from, measuring pass@1 on physics and math problems under fixed token budgets. It splits each pass@1 change into finishing (answering within the budget), reliability (being right once finished), and reach (solving problems the base never solves). The authors find that under tight budgets most of the change is finishing, with more tokens reliability drives the gain, and RL rarely solves problems its base cannot, adding at most 2.1 percentage points.
In briefMMGNet: A Multi-Modal Network for 6-DoF Robot Grasping in Cluttered Scenes
SummaryThis paper presents MMGNet, a network that synthesizes 6-DoF robot grasps in cluttered scenes by combining point clouds with RGB images. It uses hierarchical image-to-3D feature injection and rotational graspness modeling, which predicts approach direction and in-plane rotation together. Ablation studies show the two modules provide complementary gains. MMGNet reaches a mean AP of 60.34 on GraspNet-1Billion with RealSense and improves real-world grasp success over AnyGrasp by 8.46 percentage points.
In briefCheckpoint reload explains most of a measured state gap in two reproductions
SummaryThis paper examines whether states reconstructed at inference in World Machine's State Discovery method match the stored training-time state table. The authors analyze two paired-checkpoint runs on Toy1D and compare rollouts under reloaded checkpoints with the weights that wrote the table. Matching rollouts to the table-writing weights removes most of the mean per-dimension Wasserstein-1 distance, and the authors conclude checkpoint reload accounts for most of the measured gap in these runs.
In briefAIPC: Evidence-guided agent workflows for deploying PyTorch projects on NPUs
SummaryThis paper presents AIPC, an AI agent method for deploying PyTorch inference projects on dedicated on-device NPUs. It combines structured skills with toolchain, runtime, and profiling evidence to diagnose and repair problems across layers, covering model adaptation, artifact building, on-target execution, accuracy repair, and performance optimization. Evaluated on Qualcomm's QAIRT NPU path across vision, speech, LLM, and diffusion models, it reports better accuracy and speedups than control runs in two cases.
Zhanwei Wu (1), Jianhao Su (1), Brian Zhang (1) et al.
In briefFATE-5: Continuous Big Five Personality Control for Dialogue Models
SummaryThis paper presents FATE-5, a two-stage framework for controlling Big Five personality traits in large language models along a continuous scale. It first learns soft prompt representations of personality profiles using dialogue data and a survey objective, then applies preference alignment based on dialogue quality and personality agreement to improve naturalness. The authors report that across three model backbones it keeps strong trait expression while producing more natural dialogue than other trait-focused methods.
In briefNumTrace: A Benchmark Auditing Where Financial Analyst Agents' Numbers Come From
SummaryThis paper introduces NumTrace, a benchmark that audits the source and financial meaning of numerical claims made by financial analyst agents. It contains 540 questions over 2,163 SEC EDGAR documents and gives each claim separate value, binding and support verdicts based on accounting rules. The authors report that most failing claims are right-number, wrong-account errors, and on 13F co-filing questions about 91% pass the value check while fewer than half pass the binding check.
In briefThinkBit: A two-bit compression method that searches codewords, recalibrates scales, and selects checkpoints by reasoning performance
SummaryThis paper presents ThinkBit, a two-bit quantization framework for reasoning language models that aims to cut memory use while preserving reasoning ability. It assigns bitwidths per layer, searches codewords block by block with Gumbel-softmax, recalibrates scales with distillation while codewords stay frozen, and runs brief quantization-aware training. On Qwen3-1.7B, the recalibrated checkpoint scores 46.7 on MATH-500 versus 30.0 for EfficientQAT, and a 14B model at 3.74 GB scores 94.8 against 97.5 for the 29.5 GB bfloat16 model.
In briefMedAdaptAgent: Automatically Adapting Medical Imaging Models to New Hospitals
SummaryPretrained medical imaging models can perform poorly at new hospitals because imaging protocols and preprocessing differ. This paper presents MedAdaptAgent, an automatic harness in which an LLM-driven planner selects compatible adaptation experiments and keeps the selected model only when it improves on the pretrained model, delivering it as a runnable package. Across 26 segmentation and classification transfers, mean segmentation Dice rose from 0.483 to 0.774 with 20 local annotations, and mean classification AUROC rose from 0.810 to 0.837.
WhyProposes an LLM-driven automated harness for medical imaging model adaptation with evidence gating; 26 transfers with comparisons show gains, impact limited to medical imaging transfer.
In briefLayoutMamba: A Hybrid Mamba-Transformer Method for Efficient Document Layout Analysis
SummaryThis paper proposes LayoutMamba, a document layout analysis method that combines Transformers for local feature extraction with Mamba for modeling long-range global dependencies. It also introduces a Geometric Overlap Penalty Loss (GOPLoss) to reduce redundant bounding box overlaps. The authors report that it surpasses state-of-the-art AP on PubLayNet with 8x fewer parameters and ranks top-3 among published methods on D4LA.
In briefClinVerse: a realistic multi-agent environment for training and testing medical diagnosis agents
SummaryThis paper presents ClinVerse, an interactive multi-agent environment built on real electronic health record data for training and evaluating medical diagnostic agents. It is based on ClinPatient, a dataset of 150k diagnostic cases, and involves six specialized clinical roles, where a doctor agent gathers evidence through interviews and exam requests before diagnosing. A rule-based scoring framework measures information gathering, exam ordering, and diagnostic accuracy. Experiments report that nine self-evolution methods consistently improve agents' diagnostic performance within the environment.
In briefCoDeRNet: Selective Cross-Task Routing for Change Detection and Captioning
SummaryThis paper addresses change detection and captioning from bi-temporal remote sensing images, where detection and captioning have different supervision granularity. It proposes CoDeRNet, which keeps task-specific representations while selectively routing information between the two branches using a Confidence-Decomposed Routing mechanism. According to the authors, experiments on LEVIR-MCI and WHU-CDC show strong overall performance, especially better caption generation, with competitive localization.
In briefRubriX: A Fast, Fine-Tunable Non-Autoregressive Rubric Judge
SummaryThis paper presents RubriX, a model that judges which level of a rubric a response meets, without generating text. It uses a LoRA-adapted encoder to embed the input and each level description, then applies a softmax over cosine similarities to give a probability for each level. On held-out rubrics it reports 75.4% exact accuracy, above 57.1% for Jev and 49.8% for a Qwen3-8B judge, with latency staying near 28 ms as level descriptions grow.
In briefAdaptive Retrieval, Selective Answers: Conformal Factuality in Agentic RAG
SummaryThis paper evaluates conformal factuality in agentic retrieval-augmented generation, where a language model adaptively issues follow-up retrieval queries or stops retrieving before answering. The authors test six model-dataset configurations using post-generation conformal filtering. Higher target levels yield higher full-support rates but lower answer availability, with the reliability-informativeness trade-off characterized explicitly.
In briefDocTalk: Fine-tuned Language Models for Reliable Medical Question Answering from Discharge Summaries
SummaryThis paper presents DocTalk, a question-answering system that helps doctors by giving concise answers grounded in de-identified discharge summaries. The authors built 500 clinical question-answer pairs and fine-tuned six open-source language models, three general and three biomedical, using LoRA. Instruction tuning improved lexical, semantic, and clinical metrics, with Llama-3-Med42-8B performing best. On the external PubMedQA benchmark, Llama-3.1-8B-Instruct's accuracy rose from 0.931 to 0.977.
Asad Gohar, Md Ariful Islam Mozumder, Moon-Il Joo et al.
In briefGADAR: Action-Ranking Policy That Transfers Across Classical Planning Domains
SummaryThe paper introduces GADAR, a learned policy for classical planning that takes the domain description as input and constructs grounded actions without search, so it can run on domains unseen during training. Symbols are represented by their structural roles in the domain rather than by names, using three layers. In leave-one-domain-out zero-shot evaluation over eight benchmark domains, a single policy trained on seven domains solves 58.3% of test instances, close to the 59.3% of dedicated single-domain models.
WhyProposes a zero-shot planning policy reading domain descriptions and symbol roles, with leave-one-domain-out results near single-domain models; impact limited to classical planning.
In briefUniGFM-RAG: Jointly Training Graph Foundation Model Retrieval and LLM Reasoning End-to-End
SummaryLarge language models struggle with complex reasoning over scattered knowledge, and GraphRAG organizes that knowledge as graphs to help. This paper proposes UniGFM-RAG, which couples a graph foundation model retriever with an LLM generator through a differentiable graph bridge, and adds a semantic-guided distillation module to reduce reliance on costly annotations. The authors report that it consistently outperforms strong baselines on three multi-hop QA datasets and three out-of-domain GraphRAG benchmarks.
In briefPriority-preserving aggregation of advantages for optimizing multiple rewards in policy training
SummaryThe paper addresses the problem that adding auxiliary reward signals to a primary objective in LLM post-training can overturn the primary objective's preferences, a failure it calls advantage priority inversion. It proposes PAPO, which asymmetrically combines primary and auxiliary advantages: auxiliary supervision is kept when objectives agree and smoothly attenuated when they conflict, preserving the primary sign. Experiments across multiple domains and models with GRPO and PPO show PAPO achieves the best primary-objective performance while keeping auxiliary performance best or competitive.
In briefAgriDefend: A low-cost offline device using multimodal AI to detect crop stress before visible symptoms
SummaryThis paper presents AgriDefend, a roughly $205 offline edge device that combines canopy camera images with soil probe readings. A MobileNetV2 network analyses the images, a Random Forest processes soil trend features, and a logistic-regression model fuses them into a five-class crop-health prediction. In a greenhouse drought trial, the fused model reached 87% accuracy, and soil signals appeared about four days before visible wilting.
In briefSemantic Atoms for Scalable Text Analysis in LLM-Agent Data Analytics
SummaryThis paper addresses analyzing text columns in spreadsheet data, where useful information is often hidden in free text. It proposes two query-driven semantic atoms: Tagging, which returns one string per text item, and Categorization, which returns category objects over many items. Results are exposed as a table column or category table so SQL or Python can compute exact answers. On the 40-request TABench benchmark, the TADA-Agent system reportedly matches baseline semantic quality while using over ten times fewer semantic-model requests.
In briefDecoyCount: Testing Whether Vision Models Reason Like Humans When Counting
SummaryThe paper introduces DecoyCount, a benchmark of 1,660 real-world images with designed decoys, paired with neutral, cue-aligned, and multi-hop scenario prompts. It finds that models can count correctly when given a cue but often fail to choose the right counting process unaided, a gap of 11.8 to 22.1 percentage points. More capable models, larger thinking budgets, and agentic processes often showed larger gaps, suggesting stronger general reasoning does not transfer to aligned visual reasoning.
In briefLearning smooth zoom between dual cameras by synthesizing decoupled geometric and photometric training data
SummaryThe paper addresses the abrupt change in perspective and color when smartphones switch between dual cameras during zoom, and introduces a dual-camera smooth zoom task that synthesizes intermediate frames. It proposes ZoomGS v2, a data synthesis framework that separates geometric and color transformations from 3D scene modeling so training data can be generated from easily captured single-camera scenes, along with ZoomFI, an interpolation network using bidirectional warping to fuse dual-camera information. Experiments on data captured with two phones show that training on the constructed data improves various video frame interpolation models, and ZoomFI achieves state-of-the-art quantitative and visual results.
In briefRestoring Black-Box Point Cloud Models Under Rotation Shifts Without Gradients
SummaryThis paper addresses point cloud models that are accessible only through score outputs, whose performance drops when test-time rotations fall outside the training distribution. It proposes CanDO, a test-time canonicalization method that uses only forward queries and derivative-free optimization to find input transformations that maximize model confidence. Across several architectures and benchmarks, CanDO recovers 76% of the oracle gap under azimuthal misorientation and 51% under arbitrary SO(3) rotations for part segmentation.
In briefSeeing Before Naming: Context-Grounded Learning of Novel Visual Concepts
SummaryThis paper presents SynerNet, a framework that learns concept-specific name embeddings for unfamiliar visual categories and exposes them to linguistic contexts while conditioning them on visual evidence. On four VISTA-Beyond domains, it reaches a mean 16-shot accuracy of 74.35%, versus 72.45% for few-shot name learning. In the name-masked transfer setting across 14 domains, mean accuracy rises from 45.17% to 46.74%.
In briefPosePlaner: Refining Camera Pose Predictions Using Pairwise Planar Geometry
SummaryThis paper proposes PosePlaner, which uses flow matching to refine camera poses from feed-forward multi-view models by treating refinement as conditional generation from coarse geometric matches and an initial pose. The refined estimate fuses generated pose hypotheses, acting as a data-driven analog of RANSAC voting. It is trained on two-view data but can periodically refine streaming n-view pose estimators without retraining, and the authors report it corrects predictor failures and reduces drift in long sequences.
In briefSmall Samples, Smarter Priors: Using LLM Semantics to Improve Bayesian Factor Analysis
SummaryThe paper introduces SEUF, a framework that turns LLM paraphrases and embeddings, aligned with hypothesized item-factor assignments, into prespecified priors for Bayesian confirmatory factor analysis. It was tested on a 38-item survey of 617 industrial-park stakeholders and on two public surveys, FEVS and RIASEC. With 50 respondents, SEUF reached loading precision comparable to conventional CFA with 70 to 82 respondents, and showed lower loading errors in both public surveys.
Shaohuai Feng, Mohd Wira Mohd Shafiei, Mohd Sukri Shafie et al.
In briefBeyond Behaviour: Stimulus-response safety evaluation of vision-language-action AI models
SummaryInspired by humans' biological sensitivity to safety-critical stimuli that triggers defensive action, the paper examines whether vision-language-action models show similar responses to visible hazards and harmful consequences. The authors test three models on two video benchmarks, analysing their internal representations and action tendencies. Their findings suggest that task execution may remain the default response even when the danger is internally registered.
In briefCollective Emergent Alignment: How Agent Interaction Improves Reasoning Alignment in LLMs
SummaryThis paper studies whether interaction among multiple large language model agents produces reasoning alignment beyond what individual self-refinement achieves. It introduces REMAP, a framework that measures this effect using Semantic Coverage and interaction-specific alignment gains. Experiments across several model families and two multi-hop reasoning benchmarks show broadly positive gains, especially when reasoning requires combining evidence spread across agents. The authors also propose compact fact-level communication, which improves alignment and task performance while reducing communication overhead.
Boxuan Wang, Zhuoyun Li, Sarvapali D. Ramchurn et al.
In briefWhy predicting clean images can beat predicting noise in diffusion networks
SummaryThis paper compares two diffusion model targets: predicting the clean image (x-prediction) versus predicting the noise. It shows that noise prediction must preserve information needed to cancel input corruption, and under Gaussian corruption a linear input bottleneck sets a denoising-error floor. Experiments on ImageNet with JiT show measured denoising losses match the predicted bottleneck floor and its reduction or removal through preconditioning.
In briefStreaming Recurrent Memory: a recurrent memory that learns online without storing past experience
SummaryThis paper studies how a learner that receives a continuous stream of data, without storing raw experience, can keep a recurrent state useful for future prediction and control under delayed feedback. The authors introduce Streaming Recurrent Memory (SRM), which recalls past information by content or time, tracks elapsed steps, and is trained with one-step backpropagation through time. SRM outperforms recurrent baselines on associative recall and online language modeling and achieves near-perfect success on an associative T-Maze with delayed rewards.
In briefLogic Hijacking of RAG: Stealthy Single-Document Poisoning via Offline Optimization
SummaryThis paper studies single-document poisoning of Retrieval-Augmented Generation (RAG) systems under black-box access, where an attacker can only modify one document in the corpus. The proposed MONOLOGUE framework avoids explicit refutation: it treats the true fact as a historical baseline and presents the counterfactual as a dated, record-correction update. Payloads are refined offline through local multi-dimensional optimization combining text quality audits and surrogate behavior checks. The authors report that across three benchmarks, three retrieval backends, and four open-weight generators, the method achieves the strongest overall attack efficacy among the single-document attacks evaluated.
In briefTransporting Quantiles to Codebook: Scalable Vector Quantization That Avoids Codebook Collapse
SummaryThis paper proposes FlowVQ, a vector quantization method that addresses codebook collapse and poor code utilization in VQ-VAE. A lightweight flow matching model is trained alongside the autoencoder to learn a probability flow ODE that transports a source distribution. Codebook entries are then obtained by transporting even-quantile partitioned prior codes and computing Voronoi centroids under the induced distribution. The method is evaluated on image tokenization and generation, and on two high-precision tokenization tasks, where it is reported to scale and to use its codebook fully.
In briefTrustJev: Keeping Policy Authority Separate from Untrusted Evidence in LLM Decisions
SummaryThe paper introduces TrustJev, a framework that lets software choose among predefined actions from natural-language input while preventing instructions hidden in external evidence from overriding the application's policy. A single Transformer computes protected policy scores and a correction budget, limiting how much the evidence can adjust the decision. Across three datasets and 14 baselines, TrustJev achieves the highest mean Balance, and in the authorization benchmark it answers all 32 attacks per source correctly for 95.77% of sources.
In briefInterpretable Brain Age Prediction from Fiber Orientation Data with Rotation-Equivariant Deep Learning
SummaryThe paper presents a brain age prediction framework that combines white matter fiber orientation distributions with an SE(3) equivariant CNN operating on spherical harmonic coefficients. Evaluated on 692 participants aged 36 to 90 from the Human Connectome Project Aging dataset, it outperformed the DTI-based models and conventional CNNs tested, with a mean absolute error of 3.54 years and correlation of 0.958. Occlusion analysis showed model sensitivity to different tracts varies across age groups.
Chenyang Fan, Robert D. Sanders, Elie Matar et al.
In briefSPARTA: A sparse ternary LLM accelerator for efficient long-context inference on edge devices
SummaryThis paper presents SPARTA, a hardware accelerator for running large language models with long contexts on edge devices. It uses sparse dataflows tailored to the prefill and decode phases, a unified compute fabric for ternary matrix operations and attention, and an adaptive two-stage attention engine to cut KV-cache traffic. The authors report 4.1 to 4.8 times less KV traffic than full attention, with task-dependent quality loss, and higher throughput and energy efficiency than several baseline systems.
In briefMonocular Precession Estimation with Selective Output: A Synthetic Study
SummaryThis paper studies estimating the momentum axis and mean precession rate of elongated tumbling objects from a calibrated stationary camera. The proposed ORBIT-SCOUT method returns an estimate only when numerical, geometric, temporal and competing-hypothesis checks pass. On 96 preregistered synthetic scenes, it gave correct estimates for 25 of 64 supported scenes and no incorrect ones, but did not meet its registered coverage targets.
In briefSilent but costly: detecting prompt-caching misuse in LLM agents
SummaryThis paper presents the first empirical study of 120 prompt-caching-related pull requests across 10 popular open-source agent systems, finding that misuse occurs in both request configuration and usage read-side counting, and mostly fails silently. The authors introduce CACHELINT, which uses CodeQL to locate caching call sites and checks required preconditions against a provider capability oracle or an LLM guard. On labeled data it reaches 0.71 precision and 0.78 recall, and it found 32 misuses across 571 real projects at 0.86 precision. The findings received 21 developer confirmations and led to 18 upstream fixes.
WhyFirst empirical study of prompt-caching misuse plus a CodeQL-and-LLM detector, validated on 120 labeled cases and 571 projects, yielding 18 upstream fixes.
In briefCan fast decision models replace dedicated rerankers in RAG? A systems evaluation
SummaryRetrieval-augmented generation systems often use a dedicated reranker to select retrieved candidates, but this adds latency. The paper tests whether System-1-style decision models, which make direct structured predictions, can match reranking quality with less inference overhead. It screens 21 open-weight models on 11 benchmarks and compares a promising one with a dedicated reranker, finding that limiting candidate length restores reliability and makes some configurations competitive. However, a tuned reranker remains faster and uses far less GPU memory, so replacement requires weighing quality, latency and memory together.
In briefPrecedence-aware resource allocation: extending AUTOC to multi-level treatments
SummaryThis paper studies how to allocate limited resources across individuals under a fixed budget when higher-intensity treatments require lower-level treatments first. The authors introduce Multi-Level AUTOC (ML-AUTOC), an evaluation metric generalizing AUTOC to multi-level treatments with precedence. They also propose precedence-aware policies for assigning treatments under budget constraints. The results indicate that maximizing realized treatment effects may require reallocating treatments across individuals.
In briefPosition Matters More Than Channel: How Advice Can Push a Language Model Off a Correct Answer
SummaryThe paper studies how advice placed in a prompt can talk a language model out of an answer it already gets right. Testing Qwen3.5-27B on four-option MMLU-Pro items, the authors find that where the advice appears matters far more than the API channel used. Attributing the note to an authority or phrasing it assertively also raises error, while a replayed correct worked answer protects the model's answer.
WhyIsolates position, attributed authority, and prior output as separate factors with controlled comparisons, giving fairly solid evidence relevant to prompt safety.
In briefWhen Does Data Value Reduce to Class Balance? Coverage View of Per-Point Valuation
SummaryThis paper argues that selecting training data by taking the top-k points under a fixed per-point score, such as Data Shapley or influence, is a poor rule where selection matters most. A point's value depends on which points are already chosen, so a single scalar score cannot capture redundancy. In an equal latent-mode model, the authors derive a closed-form excess risk for fixed top-k selection relative to adaptive coverage. They recommend class balance when available labels identify the learner's modes, and explicit coverage otherwise, with the advantage limited to cases where the modes are expressed in the learner's representation.
In briefCentralized critics for N-agent ad hoc teamwork, comparing estimators that use only controlled agents' histories with those that also include uncontrolled teammates' histories
SummaryThe paper studies which privileged information a centralized critic should use during training when learning agents must coordinate with unknown, uncontrolled teammates in N-agent ad hoc teamwork. It compares controlled history estimators, which condition only on the learner's own agents, with joint history estimators, which also condition on uncontrolled teammates' histories. The authors prove both estimators are unbiased and identify when the joint history estimator has strictly lower variance. Experiments show joint history critics match or outperform controlled history critics across several tasks.
Andrew McAlinden, Caroline Wang, Steven Shi et al.
In briefTimeWorld: Action-Conditioned Multimodal World Models for Time Series Forecasting
SummaryThe paper introduces TSWM-Bench, a cross-domain benchmark of seven real-world datasets with text interventions grounded in recorded action-outcome data. It finds that conventional multimodal forecasters underperform history-only time series foundation models, while simple text-prefix adaptations improve results across backbones. The authors propose TimeWorld, a plug-in framework that adds action-response modeling to pretrained foundation models using prototype-guided latent modulation and action contrastive learning.
In briefSpatio-temporal intelligence for cities via twin and agent systems
SummaryThis survey examines how urban twins and agentic AI can be combined into an urban Spatio-Temporal Intelligence (STI) system that monitors city conditions, analyzes problems, plans and executes interventions, and adapts through feedback. It reviews how the twin builds a persistent, information-rich urban representation and how the agent interprets, reasons over, and acts on that twin. The paper proposes a two-loop blueprint, an inner agent loop for grounding, reasoning, and verification and an outer system loop linking data, twin, action, feedback, and updates, and outlines a research agenda with a shared prototype.
In briefText as the Bottleneck: A Small-Scale Replication of LLM Play in Riichi Mahjong
SummaryThis paper replicates a riichi-mahjong study of large language models on a much smaller budget, using models of at most 2B parameters on one consumer GPU. With identical data and labels, a 2.6M-parameter board-based encoder matches the strong AI Mortal more closely than any text-prompt LLM. The authors also find that small differences in decision agreement do not reliably translate into differences in table placement.
In briefBehaviorMem: Building and Evaluating User Memory from Real E-Commerce Behavior Logs
SummaryThis paper studies how to build user memory from real e-commerce interaction logs, so systems can both remember concrete behavioral evidence and infer underlying needs and preferences. The authors introduce BehaviorMemBench, a benchmark built from logs of 1,194 users and 17,462 behavioral batches over more than three years, and propose BehaviorMem, a memory design. Experiments report that BehaviorMem leads the evaluated agentic memory methods on both remembering and inference.
In briefNatural Jammers: when non-robust features mislead standard models on clean images
SummaryThis paper shows that non-robust features in deep neural networks can be antagonistic, acting as "Natural Jammers" that mislead standard models even on clean, unmodified images. Using disagreement with robust models, the authors identify a Jammed Set and find that jamming occurs only when high-frequency components align with specific spatial structures. Disrupting this alignment via spectral filtering or small geometric shifts restores correct predictions, while the overall prevalence of jamming stays roughly stable and is redistributed rather than eliminated.
In briefToDoPro: Faster, Better Reference-Based Image Generation via Early Token Reduction
SummaryThis paper addresses the high computational cost of reference-based image generation models, where combining multiple reference images and text into one token sequence causes quadratic complexity growth. The authors propose ToDoPro, which runs a few denoising steps on 2x downsampled reference images to let the model itself estimate each reference token's importance. Based on these scores, it discards low-importance tokens, replaces medium-importance tokens with downsampled versions, and keeps high-importance tokens before the main denoising stage. Experiments on three base models and two benchmarks report faster inference, lower peak GPU memory, and improved generation quality from filtering out noisy tokens.
In briefTuring Test Signal: Indistinguishability as a Learning Reward
SummaryThe paper proposes a teacher-Turing-Test training method in which a stronger teacher model both serves as the imitation target and acts as the distinguisher, while a weaker student is trained to make the teacher attribute its responses to itself. The student uses this self-attribution probability as a reinforcement learning reward, without ground-truth answers. Experiments on CommonsenseQA, GSM8K, KodCode, and Spider, with Qwen2.5-1.5B as student and DeepSeek V4 Flash as teacher, report improvements on independent held-out evaluations.
Emanuele Rimoldi, Riccardo Neumarker, Tomaso Poggio et al.
In briefArya Forge: Auditing Validation and Release Approval for Financial Instruction Data
SummaryThe paper presents Arya Forge, a system for financial instruction-data pipelines that combines typed records, current-content checks, and approval coverage tied to a dataset manifest, so that content validity is separated from permission to release a dataset. The authors evaluate it in a financial-data systems case study, including an integrity audit of 100,000 synthetic records and 240 comparison scenarios. The findings apply only to the tested implementation and workloads and do not establish general financial correctness, authenticated approval, or downstream model improvement.
In briefHow personalized AI reply suggestions affect creators' direct messages
SummaryThis paper studies how social media content creators handle large volumes of direct messages. Interviews with 15 creators revealed fatigue from writing suitable replies while keeping their media personas. The authors built a system that generates replies tailored to each creator's writing style and compared it with 13 creators. Creators valued softened tone and expected higher response rates, but had to verify AI outputs, and the suggestions could shift their attention away from engaging with followers.
Eun Jeong Kang, Jingruo Chen, Jadon Geathers et al.
In briefDynEdit: Entropy-Guided Dynamic Method for Long-Sequence Knowledge Editing in Language Models
SummaryThis paper addresses updating outdated or incorrect knowledge in large language models without full retraining. It proposes DynEdit, which uses token-level entropy to guide non-uniform sliding windows that cover difficult, high-entropy regions, and applies entropy-guided optimization to information-dense tokens within each window. Experiments on three benchmarks and two backbone models show DynEdit consistently outperforms competitive baselines across diverse long-form editing settings.
In briefThe Right Fact, Wrong Decision: A Benchmark for Scientific Agent Memory Validity
SummaryThe paper argues that scientific agents must judge whether a remembered fact is valid for a given decision, not just retrieve it. It introduces SciValidityBench, with 64 tasks across five validity modules and six scientific domains, and the SciStateForge framework, which converts distributed experimental histories into explicit decision-governing state. Across four model families, compiled scientific state improves valid decision-making by 9.2 to 18.8 percentage points over a generic Event Graph.
In briefFactorized Gradients for Scalable, Highly Expressive Parametric Diffeomorphisms (FG-CPAB)
SummaryThis paper introduces FG-CPAB, a factorized-gradient method for scalable CPAB transformations that separates trajectory integration from parameter-space projection. It reduces gradient computation complexity from O(dTN) to O(TN+dC), reporting speedups of up to four orders of magnitude in 2D and about 500x lower peak GPU memory use. This makes fine-tessellation CPAB practical in 2D and 3D for the first time.
In briefEfficient multi-modal video fusion with dynamic region-guided temporal interaction
SummaryThis paper proposes an efficient network that fuses videos from different sensors, such as infrared and visible cameras, to produce temporally consistent videos with salient targets and structural detail. It models temporal dependencies with linear complexity using a test-time training paradigm, applying temporal interaction only to dynamic regions, and reduces flicker with a feature-level constraint built on pretrained MAE features. The authors report better fusion quality and temporal consistency than existing methods on multiple benchmarks, with high efficiency on edge devices.
In briefThe Physical Cost of Context: Resource Accounting for Coding Agents Under Quality Constraints
SummaryCoding agents resend growing context on every model call, yet context policies are usually judged only by prompt tokens saved. This paper proposes a quality-gated protocol that splits executed prefill into logical prompt minus reused prefix and measures KV residency, time, and energy. On 450 localization cases derived from SWE-rebench V2, preloading retrieved blocks and anchored history compaction both pass a non-inferiority gate. In a separate study, preloading cut model calls and a GPU-energy proxy relative to on-demand search, while compaction saved nothing further. The authors also find that deleting history can turn a token saving into extra work.
In briefFrom Belief Inference to Task Action: Testing AI Theory of Mind
SummaryThis paper examines whether language models can understand what others know and believe, and act on that understanding in collaborative tasks. The authors extend existing verifiable household and calendar scheduling tasks with asymmetric information and assess world understanding, information access, beliefs, decisions, and execution separately. Results show these abilities diverge: models can recover others' beliefs correctly yet fail to use them in completing tasks.
In briefMechanistic analysis of LoRA, and a method that adapts only the layers each task needs
SummaryThis paper uses mechanistic interpretability to study how LoRA changes the internals of large language models, finding that its effects are uneven across layers and components and that important layers differ by task. It proposes Adaptive-LoRA, which picks task-specific layers automatically using frozen-base zero-ablation calibration. On Qwen2.5-1.5B across 11 reasoning tasks, it outperforms full-layer LoRA using 30% of layers and a 3.5× reduction in adapter parameters.
In briefBabysitting AI Agents: A Taxonomy and Study of Delegation Failures
SummaryThis paper examines cases where AI agents carry out delegated, multi-step tasks in ways users did not expect, which the authors call agentic delegation failures. Using 163,487 Reddit posts about AI agent use, they built MANDATE, a taxonomy of 23 failure categories in seven groups. A follow-up study with 73 agent users found that failures rated more severe were harder to notice, and prior experience with a failure increased noticeability. The authors also assess research maturity on each failure and identify common ones that have received little attention, such as agents running without making progress.
Shravika Mittal, Sunnie S. Y. Kim, Scott Counts et al.
In briefQuery-driven learning that learns only what a query needs for lifted reasoning
SummaryThe paper studies query-driven learning, where an agent learns only the knowledge needed to verify a small set of queries instead of learning a full knowledge base first. It extends prior work to the first-order setting and identifies a necessary and sufficient condition called bounded concealment. The authors propose a quasi-polynomial-time algorithm at the lifted level. They prove soundness and completeness relative to space-bounded lifted proofs, with quasi-polynomial sample and runtime bounds.
In briefDecisionTraceBench: A Benchmark for Tracking Key Decision Information in Multi-Speaker Conversations
SummaryThis paper introduces DecisionTraceBench, a benchmark for evaluating how well systems track decision-critical information across multi-party spoken conversations. It assesses decision event sequences, speaker attribution, scope, and temporal evidence using trajectory-level F1. The dataset consists of Chinese financial dialogues with synthetic and human-performed audio. Results show that seven evaluated systems score between 5.73% and 21.35%, indicating that overall transcription error rates do not adequately capture full trajectory recovery.
In briefSceneAgent: an agent pipeline that turns 3D and 4D captures into simulation environments
SummarySceneAgent is an agentic pipeline that converts visual captures such as 3D Gaussian Splatting, photogrammetry, LiDAR, and 4D videos into interactive simulation environments. It handles object segmentation, physical parameter estimation, part decomposition, articulation, and deformable simulation, and corrects errors via simulator-based visual inspection. On a cluttered GraspClutter6D bin scene, it reports higher F1 scores than SimFoundry, and it reports improved dynamic IoU on three 4DCodeBench scenes. The authors also show diverse simulated interactions and preliminary policy transfer to a real robot.
In briefSentinel-2 to 1-m Imagery: A Model for Cross-Satellite 10× Image Sharpening
SummaryThis paper tackles reconstructing 1-meter satellite images from Sentinel-2 imagery, which come from different satellites and require a tenfold magnification. The authors introduce the YREB-SR benchmark and CrossVAR, a visual autoregressive model with a diffusion refiner. On 3,015 test scenes, CrossVAR reports the best scores among compared methods across PSNR, SSIM, LPIPS, and FID. The source code is publicly available.
WhyProposes CrossVAR, a visual autoregressive model for real-world ×10 cross-sensor SR, achieving best metrics over 3,015 test scenes; novel and well validated, with solid value for remote sensing SR.
In briefDiscovering topics from linear autoencoders: a way to extract interpretable themes from recommender models without extra training.
SummaryThis paper examines how linear recommenders and topic models are related under a latent-factor model of user-item interactions. It shows that latent topics can be read directly from the weight matrix of a closed-form linear autoencoder, providing interpretations of recommendations without additional training. The authors also propose item and user normalizations to keep popular items and very active users from dominating the topics. Across six public datasets, the method yields higher topic quality than dedicated topic models while remaining competitive in recommendation performance.
