In briefComparing an analytical model and discrete-event simulation for robotic picking in warehouse aisles
SummaryThis paper presents an analytical versus discrete-event simulation (DES) benchmark for intra-aisle robotic picking in automated warehouses, covering shuttle travel, tote handling and item picking. Simulation results are compared with the analytical formulation of Lerher et al. across 180 scenarios. Across the 174 principal scenarios, the mean absolute relative cycle-time discrepancy is 0.8655% and the maximum is 7.0370%. The results show close but not exact agreement under the tested conditions, and no validation against an operating warehouse was performed.
In briefBehavioral probes that reveal information flow in LLM swarms
SummaryThis paper studies how information moves inside groups of large language models, rather than only measuring final task performance. The authors adapt two hallucination-analysis scores, External Context Score (ECS) and Parametric Knowledge Score (PKS), as probes of each node's dependence on external context and on its own knowledge. Experiments show ECS rises with input relevance while PKS falls as useful context accumulates. Probe-guided pruning reduced redundant communication and token use, and probe-guided path search converged faster than utility-only baselines.
In briefToken-set choice confounds POPE: auditing yes/no answer extraction in VLM hallucination tests
SummaryThe authors tried to reproduce a hallucination-mitigation result on the POPE benchmark and found that the answer-readout method alone shifts F1 scores. The model never produced the two vocabulary IDs the evaluation looked for, but instead emitted space-prefixed forms, and reading those forms raised LLaVA-1.5-7B's adversarial F1 by 6.13 points. Most surveyed papers do not state their readout method, and the reference run used 4-bit NF4 quantization, so the paper says it cannot tell how much of any reported gain reflects evaluation choices.
In briefABACUS: Program-guided multistep reasoning with frozen judgment models
SummaryThis paper introduces ABACUS, a framework that jointly builds judgment schemas and composition programs for frozen judgment models. A generative teacher designs the procedure offline, the judgment model answers local semantic questions online, and a saved program composes these decisions into an answer. On three BIG-Bench Hard tasks, the programs reach near-perfect 99.67% macro accuracy, compared with lower accuracy for direct Kev and final-answer-only Qwen3.5-4B, while keeping parameters fixed.
In briefPaceVLA: Adaptive Time-Step Compression for Faster Robot Action Models
SummaryThis paper addresses the slow execution of Vision-Language-Action (VLA) robot models, which often produce many fine-grained low-level actions. It proposes PaceVLA, which uses mode-specific action decoders at different temporal granularities and a router that chooses the execution mode at each step based on whether coarse execution is appropriate. According to the paper, experiments on three multi-task simulation benchmarks and six real-world tasks show improved execution efficiency while maintaining task success rates across multiple robot embodiments.
In briefUsing activation steering to make ASR systems more accurate and fairer across speakers
SummaryThis paper studies activation steering as a way to improve both overall accuracy and demographic fairness of automatic speech recognition (ASR) systems. The authors compare paired activation subtraction, which proves ineffective, with supervised optimization, which yields steering vectors that improve performance and fairness across gender, age, and dialect. They also find that vectors trained on acoustically similar speakers perform as well as or better than those trained on the same speaker or demographic group, enabling few-shot or label-free zero-shot adaptation.
Felix Herron, Alex Colagrande, Solange Rossato et al.
In briefHORIA: Reconstructing Humans and Objects in 4D from Monocular Video
SummaryThis paper introduces HORIA, a method for jointly reconstructing a person and an object in 4D from a single-view video. It splits the video into temporally grounded interaction phases using motion-aware captioning, generating a text description for each phase. It then re-renders the initial reconstruction from novel viewpoints, refined with a text-to-video diffusion prior to compensate for limited single-view information. The authors report temporally coherent, accurate reconstructions across diverse interactions and state-of-the-art performance.
Hyeongjin Nam, Taeryung Lee, Daniel Sungho Jung et al.
In briefEmbodiedWorld-200K: Embodied World Agents via a Planning Dataset and Interactive World Models
SummaryThis paper studies Embodied World Agents that understand diverse visual environments and act in open-world settings. It releases EmbodiedWorld-200K, a dataset of over 200K samples across diverse scenes, with multi-level annotations from camera trajectories to semantic instructions, plus unified evaluation splits and metrics. The authors also propose reinforcement-learning post-training inside a video world model used as a low-cost environment proxy. They report that pretraining on the dataset improves open-world planning and that world-model RL yields clear gains on out-of-distribution scenes.
In briefProductive state-space exploration for multi-turn RL in LLM agents
SummaryThis paper studies reinforcement learning for LLM agents in multi-turn, long-horizon environments, where agents often repeat invalid actions or loop without progress. The authors attribute this to a mismatch between the agent's internal state representation and the environment's dynamics. They propose a conditional state-space exploration framework with an intrinsic reward that penalizes actions whose observations duplicate prior state memory. The authors report that the approach substantially improves policy learning and reduces invalid actions and repetitive loops.
Mohammed Haroon Dupty, Yong Liang Goh, Song Guo et al.
In briefCXBench: Testing whether AI customer-service agents follow required procedures
SummaryThis paper introduces CXBench, a benchmark for evaluating tool-using language model agents in six regulated customer-service domains. It scores final outcomes and process compliance separately, checking required disclosures, consent, and procedures against trace evidence. Across seven models and 120 tasks, 55.5% of 840 retained executions were false-passes, reaching correct outcomes without compliant processes, and only 13.8% passed both checks.
In briefAdaExploreBench: Testing LLM agents on proactive exploration and adaptation in changing environments
SummaryThis paper introduces AdaExploreBench, a benchmark that models environment dynamics as hidden rules and tests whether LLM agents can actively gather evidence to infer them. It includes implicit rule drift, where rules change without notice, so agents must notice degraded performance and re-explore. Experiments on eight leading LLMs show the task remains hard, with the best model reaching 61.9% calibrated accuracy, and the authors find agents reason well given evidence but struggle to acquire it and detect outdated knowledge.
In briefOrderGPT: Learning the Generation Order for Discrete Data Models
SummaryThis paper studies the order in which discrete data is generated: autoregressive models go left to right, while masked diffusion models use any order, and the authors treat these as two extremes of one axis. They propose OrderGPT, which learns a task-level generation order within a single training run without canonical-order supervision. According to the paper, the method improves downstream modeling metrics across language, human pose, and tabular data.
WhyProposes learning task-level generation order within GPT, validated on language, pose, and tabular data, though experimental breadth and comparisons appear limited.
In briefClaim Laundering: Why Agent Models Trust Logs Over Actual Outcomes
SummaryThis paper studies how AI agents decide their next action when their own log says a step succeeded but the environment shows it failed. The authors test four decision models across Android, ALFWorld, and WebShop, finding that logged records often outweigh observations and lead to continued actions after silent failures. They propose EVIGATE, an intermediate-layer veto that uses internal signals to improve task success in injected-failure experiments.
Tina Khezresmaeilzadeh, Seyedarmin Azizi, Erfan Baghaei Potraghloo et al.
In briefSearching for multi-agent pathfinding policies directly in code
SummaryThe paper proposes using autonomous coding agents to search for coordination policies directly in code space, rather than relying on hand-designed algorithms or opaque reinforcement learning. The agents iteratively propose, implement, evaluate, and refine algorithms guided by simulation feedback and correctness checks. The authors report that this approach outperformed manually engineered solutions in two case studies.[SEP]
WhyProposes code-space policy search, beating a proprietary MAPF solution by 58.9% in simulated warehouses, with a finite-time completeness proof.
Matthias Eder, Pantelis Emmanouil, Vivek Farias et al.
In briefTRACE: A Framework for Testing Text-Vision Embedding Models' Robustness and Hard Tasks
SummaryThis paper introduces TRACE, an evaluation framework that groups 168 tasks into General, Adaptability, and Challenge categories to show where text-vision embedding models' performance comes from. Controlled experiments on two models show that combining general and adaptation data raises Adaptability scores by 7.67 and 7.09 points while largely preserving General performance. However, Challenge scores stay below 5.0, and multilingual and visual-document tasks are identified as major weaknesses. The authors conclude that data adaptation alone is insufficient to solve these problems.
In briefCoASFL: Coupled Consistency Optimization for Asynchronous Split Federated Learning
SummaryThis paper proposes CoASFL, a framework for asynchronous split federated learning that trains large deep models on resource-constrained edge devices. It uses dynamic activation merging to reduce inconsistency across clients and a consistency-aware aggregation mechanism that adjusts weights by participation frequency and progress gaps to reduce inconsistency between clients and the server. Experiments on an 80-device heterogeneous edge testbed report training speedups of 1.61 to 3.74 times over state-of-the-art methods and improved final accuracy under heterogeneous conditions.
In briefSpeech LLMs That Breathe: Modeling Breaths and Non-Verbal Sounds in Dialogue
SummaryThis paper addresses the lack of non-verbal vocalizations, such as breathing, affective sounds and discourse cues, in current speech interaction models. The authors cover 13 major NVV types and build a natural dialogue dataset through an automated synthesis pipeline. Using multi-stage fine-tuning and NVV-aware GRPO, they report improved expressiveness of VocalNet and VITA-Qinyu in empathy and role-play dialogues.
In briefGlobal probabilistic air pollution forecasting with a generative model at kilometre scale
SummaryThe paper introduces Foehn, a generative model trained on global kilometre-scale PM2.5 estimates that also computes where the air originates. In a year of retrospective forecasts, it outperformed the operational forecast at three stations in four. Its calibrated ensemble mean detected local pollution extremes four days ahead, matching the operational forecast's one-day lead.
WhyProposes generative Foehn for kilometre-scale probabilistic PM2.5 forecasts, beating operational forecasts at most stations in retrospective tests.
In briefGMoE-FTI: A Universal Face Template Inversion Method Using a Genuine Mixture-of-Experts
SummaryFace recognition systems often store facial templates instead of raw images to protect privacy, but template inversion can reconstruct faces from them. Existing methods only work with a single template type. This paper proposes GMoE-FTI, which uses a genuine mixture-of-experts mechanism trained in three stages to produce latent codes compatible with pre-trained StyleGAN3. On the LFW benchmark and seven unseen template types, it achieves an average 40% TAR improvement, with FAPD and FAPC each reduced by 2%.
In briefMITRA-lexicon: Sanskrit dictionaries with attested translations into English, Tibetan and Chinese
SummaryStandard Sanskrit dictionaries cite attestations only as abbreviated edition references, without links to digitized texts or translations. This paper presents MITRA-lexicon, six directional lexica with about 2.18 million headwords and 10.4 million attestations, each paired with aligned sentences and character offsets. The alignments are produced by a multi-step pipeline using grammatical annotation models, fine-tuned language models and a teacher-student setup, and are evaluated against reference dictionaries and gold alignments. The lexica, the span annotation model and a lookup web interface are released.
Sebastian Nehrdich, Kengo Harimoto, Sven Sellmer et al.
In briefKV-Surgeon: A Compensation-Aware Method for Offline KV-Cache Compression
SummaryThis paper presents KV-Surgeon, a method for offline compaction of a language model's key-value cache, which builds a smaller cache once so it can be reused when querying the same context repeatedly. It uses calibration data to account for interactions among cache entries and derives a quadratic surrogate with closed-form expressions for jointly choosing which entries to remove and how to compensate the remaining ones. Evaluated on QuALITY, LongHealth, and Long Code Arena with models from 4B to 30B parameters, it reports improved cost-accuracy trade-offs over the attention matching baseline.
In briefPreserving diverse reasoning paths in diffusion LLM decoding with lineage-preserving SMC
SummaryThis paper proposes lineage-preserving sequential Monte Carlo (LP-SMC) for inference-time decoding of masked diffusion language models. It combines Chopthin resampling with an input weight floor to avoid losing lineages through multinomial resampling, which can discard candidates that may still reach a correct solution. Across mathematical-reasoning and code-generation benchmarks on two base models, LP-SMC raises oracle coverage in every setting and improves end-to-end accuracy in most settings.
Minoo Ahmadi, Seyedarmin Azizi, Mehdi Kamal et al.
In briefEvoSkill v2: a system where agents rewrite their own task instructions to improve results
SummaryEvoSkill v2 is a three-agent system that edits the instructions, or skills, given to an agent without changing its model or harness. On SpreadsheetBench v2 its skills raised held-out pass rate 7.7 percentage points above the stronger of SkillOpt and SkillBoost, and 5.4 on τ³-bench banking. It also let a less capable model, Claude Haiku 5.5, gain up to 21.5 percentage points on the banking task.
In briefTESS: LLM-guided search for planning robot-style tasks step by step
SummaryThis paper addresses embodied task planning, where an agent breaks a complex task into atomic actions. The authors propose an LLM-guided search method that uses commonsense knowledge to turn instructions into goal conditions, aid exploration in partially observable environments, build a rule-based world simulator, and prune candidate actions. Experiments on VirtualHome and ALFWorld are reported to show superior performance.
In briefOmniGround: Improving AI Video Timestamp Location via Attention Consistency Training
SummaryThis paper studies why omni-modal models that handle video, audio, and text struggle to locate moments in time. It finds that the model's internal attention identifies target moments more accurately than the timestamps it generates, calling this a temporal generation gap. The authors propose TACG, a post-training method that aligns internal attention with predictions and ground-truth annotations, and train OmniGround on a 580K-sample dataset. The paper reports 37.5 mIoU on LongVALE and 35.7 mIoU on UnAV-100, with competitive general audio-visual and vision-only grounding performance.
In briefCondenseVLA: Learnable History Compression for Efficient Multi-Frame Robot Control
SummaryThis paper proposes CondenseVLA, a multi-frame vision-language-action policy for robot manipulation. It adds a learnable module on the action-decoding side that uses learnable queries and cross-attention to compress recent per-frame features into a fixed set of slot tokens, with auxiliary regularization to reduce redundancy among slots. The method raises Simpler-WidowX success from 58.7% to 69.8%, reaches 97.2% on LIBERO and 52.8% on real-world tasks, and achieves 21.6% on memory-dependent RMBench tasks.
In briefMIRA: A Multimodal Illustrated Research Agent That Writes Controllable Reports With Text and Images
SummaryThis paper presents MIRA, a deep research agent that searches the web and writes illustrated reports combining text and images under user-specified requirements. Its harness moves through four stages: evidence collection, report submission, feedback-driven revision, and final delivery, and the foundation model is post-trained on constructed multi-hop multimodal research trajectories. Using Qwen3.6-35B-A3B, the authors report state-of-the-art results among open agents on 7 multimodal-search and deep-research benchmarks, with traceable textual and visual evidence in the reports.
In briefUtilityBench: Measuring Language Model Utility When Errors Have Different Costs
SummaryThis paper proposes a utility-based evaluation approach, arguing that language model errors carry different real-world costs, such as serious patient-safety risks in medicine, which standard protocols ignore by treating all errors equally. The authors build UtilityBench, a benchmark of 15,000 instances with verifiable answers, and use Best-Worst Scaling to estimate error severity. They also introduce UtilityAUC, a metric based on the area under the utility curve. Evaluating 20 state-of-the-art models, they find that higher accuracy or more advanced architectures do not reliably yield higher UtilityAUC.
In briefSAPO: Self-Supervised Attention Policy Optimization to Improve Temporal Video Grounding in Vision-Language Models
SummaryVideo vision-language models understand content well but often locate events in time imprecisely. The authors found that the models' internal attention intervals beat their predicted timestamps on about 66% of valid examples, which they call the temporal generation gap. They propose SAPO, which uses this internal attention as self-supervision with unlabeled video-query pairs to improve temporal grounding. The method improves results across several models and datasets, with relative mIoU gains up to 33.7%, while preserving general video understanding.
In briefMaestro: Coordinated Agentic Lean Proof Development
SummaryMaestro is a Lean proof harness that handles failures in LLM-driven formal proof writing through global coordination. Provers report why they failed, and a coordinator with a view of the whole proof decides whether to guide a prover, correct a false lemma, or retract and replan a flawed subtree. The authors evaluate it with four frontier LLMs on seven mathematics and three software verification benchmarks, reporting more proofs than each model's native coding agent. It also reports outperforming published prover systems on shared benchmarks.
WhyProposes global coordination of failure recovery in Lean proving, with broad benchmark comparisons against native agents and prior provers showing clear gains.
In briefLEAPS: Testing how long-context language models learn patterns and track state
SummaryThis paper introduces LEAPS, a benchmark of 12 new datasets covering agentic workflows, behavioral modeling, and market and world event forecasting, for evaluating long-context language models. It tests 30 API and open models with inputs from 8K to 512K tokens. Models that nearly perfectly handle recall tasks differ widely in state tracking and pattern learning, and these abilities degrade sharply as inputs grow longer. Thinking models often underperform their base counterparts.
In briefLEAP: Faster Diffusion Sampling via Optimal Linear Leaps
SummaryThis paper proposes LEAP, a method that speeds up diffusion model sampling by making optimal linear jumps along the sampling trajectory, without extra model evaluations. It solves a supervised least-squares problem in the local trajectory subspace, based on the low dimensionality and cross-trajectory consistency of sampling paths. Experiments report better results than prior methods on text-to-image (Stable Diffusion 3.5, FLUX.1-dev) and image-to-3D (TRELLIS.2) generation. LEAP can also be added to existing solver-based methods as a plug-and-play enhancement.
WhyProposes local-subspace least-squares linear leaps for diffusion sampling, a substantive improvement over ODE solvers, validated across several text-to-image and image-to-3D models.
In briefFisherEdit: Fisher-information-guided knowledge editing for large language models
SummaryEditing facts in large language models one after another can degrade their general abilities as parameter changes accumulate. The authors revisit AlphaEdit and find its robustness comes mainly from update regularization rather than the null-space projection it was designed around. They propose FisherEdit, which uses Fisher geometry to apply importance-aware regularization in parameter space, penalizing changes to sensitive parameters more strongly, with an efficient row-wise closed-form solution. Experiments on Llama-3, Qwen2.5, and Olmo-3 across several benchmarks report state-of-the-art results with better preserved stability and downstream accuracy.
In brief3D-MoE: Hardware-software co-design for efficient MoE decoding on 3D-stacked memory
SummaryThis paper addresses the memory-bound decoding of Mixture-of-Experts language models, where expert loads are skewed and change over time. It proposes 3D-MoE, which exploits complementary expert-load fluctuations to balance work across groups offline, handles remaining hotspots with a collaborative memory network, and remaps experts online. The authors report an average 4.63× MoE-layer speedup over expert parallelism and 1.58× to 1.86× end-to-end throughput gains over the strongest near-memory baselines.
In briefCALM: a training-free decoding method that masks transition words to curb overthinking in reasoning models
SummaryThis paper presents CALM, a training-free decoding method that reduces overthinking in large reasoning models. It monitors trial answers at step boundaries and applies logit masking to transition tokens such as "Wait", "Alternatively", and "Hmm" to suppress redundant reasoning. Across four models and five benchmarks, the authors report 18.6%–25.0% fewer tokens than full chain-of-thought reasoning, with accuracy maintained or slightly improved.
In briefCo-evolving text and spatio-temporal graph structure for edge classification
SummaryThis paper studies edge classification on dynamic text-attributed graphs, such as review networks and news-derived knowledge graphs, by letting text representations co-evolve with network topology. The authors use a lightweight Transformer adapter updated through spatio-temporal gradients from a contrastive pretraining objective, with a language model maintenance objective to prevent semantic distortion. Evaluated on the Dynamic Text-Attributed Graph Benchmark, the approach improves over frozen-text baselines across several Web-related domains, while avoiding the high overhead of end-to-end joint training.
In briefMulti-scale fusion of fast vision Mamba and foundation model features for Gleason grading of prostate pathology images
SummaryThis paper proposes a multi-scale classifier based on state-space models for Gleason grading of prostate whole slide images. It fuses features from a frozen pathology foundation model with those from a lightweight FastVim network, distilled from the foundation model and pre-trained on ImageNet, using Mamba encoders and a jointly trained fusion head. On the PANDA dataset under 5-fold cross-validation, it reports macro F1 of 0.894 and Gleason 5 F1 of 0.802, a 12% and 21% relative improvement over the best single-scale model.
In briefDetecting Unfamiliar Inputs While Learning Tasks in Sequence Without Stored Data
SummaryThis paper studies out-of-distribution (OOD) detection in continual learning, where a model learns tasks one after another and must recognize unfamiliar samples. The authors propose a rehearsal-free framework inspired by Kolmogorov-Arnold Networks (KAN), along with a task-adaptive classifier, TA-KAN, that reduces recency bias and improves separation between known and unknown samples. They also introduce a geometry-guided OOD score that works without past in-distribution data. According to the paper, the method significantly improves OOD detection while keeping strong classification performance and achieves state-of-the-art results in this setting.
Nimeshika Udayangani, Flora D. Salim, Christopher Leckie et al.
In briefSWE-Residual: A benchmark testing whether coding agents stay reliable when they inherit partly finished work
SummaryThis paper introduces SWE-Residual, a benchmark that tests whether coding agents keep their performance when the starting state already contains earlier work on the task. It includes 198 residual states from DeepSWE and Terminal-Bench, each left by a predecessor agent that reported completion although hidden tests still failed. Across five frontier agents, success dropped by 20.6 to 53.5 percentage points on DeepSWE and by 20.6 points or more on Terminal-Bench compared with a clean start.
WhyIntroduces starting-state robustness; 198 residual states versus clean starts show acceptance bias, with solid evidence and broad impact on coding-agent evaluation.
In briefWhen AI Should Speak: A Benchmark for Timing Proactive Agent Contributions in Group Chat
SummaryThis paper introduces PITCH, a benchmark built from human judgments on when a proactive AI agent should contribute to multi-party chats. Participants in 35 unscripted conversations placed 384 candidate contributions and marked ideal turns, acceptable windows, or times they should not appear. Evaluating 15 models across 28 configurations, the authors find that models place contributions well with full conversation history but chime in too early when deciding live, with best-turn accuracy dropping from 57.8% to 19.0%.
WhyFirst large-scale human-judged benchmark for proactive agent timing reveals systematic premature interventions by models; novel design with broad relevance mainly to multi-party chat agents.
In briefDiagnosing multiple motor faults at once with an attention-based cross-talk network
SummaryIndustrial motors can suffer several faults at the same time, so this paper addresses compound fault diagnosis, which identifies multiple fault components and their severities. The authors propose RNDR-TFCA, which keeps task-specific feature extractors while selectively exchanging information across tasks, using time-frequency and channel attention. On a rotor-bearing system dataset, it achieved a macro F1 score of 0.898, outperforming the baselines, with 713 K parameters.
In briefParameter-efficient adaptation for speaker verification using constrained residual paths and per-frame layer weighting
SummaryFull fine-tuning of large self-supervised speech models for speaker verification is costly. The paper proposes a Manifold-Constrained Residual Adapter (MCRA) that mixes grouped residual features through a Sinkhorn-normalized matrix with a learnable gate, and Frame-Adaptive Layer Aggregation (FALA), which computes layer weights per frame using a shared learned query. Compared with a frozen WavLM baseline using a fixed layer-weighted sum, the method reports relative EER reductions of 22.4% on VoxCeleb1-O and 2.9% on CN-Celeb. The trainable adaptation parameters amount to only 0.0009% and 0.0016% of WavLM-Large's parameters, respectively.
In briefRethinking sparse-view 3D Gaussian splatting from a few-shot learning perspective to curb overfitting
SummaryThis paper addresses overfitting and geometry degradation in 3D Gaussian Splatting (3DGS) when only a few viewpoints are available. It treats each Gaussian primitive as a neuron and frames the problem as few-shot learning, using optical flow-guided pseudo-view synthesis with joint camera-scene optimization, plus motion-aware dropout and early stopping. The authors report improvements over optimization-based and feed-forward baselines on the SIGGRAPH Asia 2025 volumetric video challenge sparse-view dataset, and say the method placed first in that track.
In briefLearning the Interface, Not the Implementation: Output Format Matters More Than Training in LLM Hardware Code Generation
SummaryThis paper studies how large language models generate register-transfer-level (RTL) hardware code, using SECDED ECC memory wrappers as a bounded test task with formal equivalence checking. It compares seven training methods and four output representations while holding the task and checker fixed. Raw Verilog output reaches at most 0.17 pass rate, while outputting a compact configuration reaches 0.83 without training, and all training methods then score 0.89 to 1.00. The results show output representation matters far more than training regime.
Siddharth Patel, Vikram Singh, Siddharth Savyasachi Malu et al.
In briefAsynchronous On-Policy Self-Distillation (AsyncOPSD): a faster way to train one model as both student and teacher
SummaryThis paper proposes AsyncOPSD, a framework in which one language model acts as both student and teacher, with the teacher given privileged information. It separates rollout generation from learner updates to raise training throughput, and uses staleness-aware policy control and token-level entropy weighting to handle outdated trajectories. Across three model sizes and math reasoning benchmarks, it reports 1.96 to 2.25 times higher throughput than synchronous OPSD, along with higher average accuracy.
In briefDeltaTrace: Signed Attribution Method Explaining How Input Words Affect Reasoning Language Models
SummaryThis paper presents DeltaTrace, an attribution framework that explains how context tokens support or oppose a language model's fixed response. It measures how the response's log-probability changes when the explained tokens are masked, then distributes that difference backward through the computation graph. On the evaluated attention model, DeltaTrace achieves the lowest RISE and MAS deletion-faithfulness scores on most tasks and the highest needle-in-a-haystack recall, exceeding FlashTrace by 5.67 to 24.33 points.
In briefLearning to Reason Efficiently with Interventional Step Credit
SummaryLarge reasoning models often produce long chain-of-thought traces that mix useful and unnecessary steps. The paper proposes interventional step credit (ISC), which uses causal interventions on model parameters to measure how much each step contributes to the reasoning. It also introduces IGPO, which uses ISC to guide fine-grained advantage rescaling in reinforcement learning to cut redundant steps. Across two model scales and nine benchmarks, IGPO reduces average tokens by up to 62% and performs better overall than compared methods.
In briefVision-Language-Delta-Action: Compact Change Variables as Interfaces for Robot Manipulation
SummaryThis paper proposes VLDA, which replaces full future-state prediction with a compact Delta variable capturing latent world-state changes, serving as an interface between vision-language understanding and action generation. A Local Transition Filter focuses on manipulation-relevant changes while suppressing static reconstruction bias, and a Global Transition Anchor preserves the overall direction of state changes. Delta-Driven Action Refinement uses Delta as a structured seed for action denoising. Experiments report improved long-horizon consistency, robustness, and generalization.
In briefResidual-guided adaptive Tucker rank allocation for efficient MCI-to-AD conversion prediction
SummaryThis paper proposes a method for parameter-efficient fine-tuning of pretrained 3D networks that allocates Tucker ranks across convolutional layers under a fixed adapter-parameter budget. Rank choices are guided by layerwise reconstruction residuals and selected via dynamic programming to minimize their normalized sum. Evaluated on ADNI data for predicting mild cognitive impairment conversion to Alzheimer's disease within 36 months, it reaches a mean AUC of 0.7706 across five seeds with about 0.618 million trainable parameters, 1.45% of the full model.
In briefCODA: A cohort- and drift-aware foundation model for multimodal clinical medical reasoning
SummaryThis paper proposes CODA, a foundation model for multimodal clinical reasoning that accounts for both patient cohorts and individual variation. It uses cohort-aware population stratification and drift-aware individual adaptation, with reinforcement learning used to discover clinically meaningful cohorts. Each patient's deviation from cohort prototypes is modeled as a structured drift signal. Experiments on real-world electronic health record benchmarks report consistent effectiveness and generalizability across closed-QA, open-QA, and report generation tasks.
In briefBenefit-targeted routing for LLM-based supplier evidence screening in manufacturing
SummaryThis paper studies how an evidence-review system should allocate extra model judgments and control their use, using 6,164 frozen manufacturing source-task pairs. On the original 300-case test, benefit-targeted routing achieved 202/231 agreements with two human references using 60 consultations, versus 189/218 for a logistic-risk router. On an additional 300 cases the differences did not establish an advantage, and the 200-pair identity audit reports 97.45 percent weighted merge precision, while prospective company-disjoint performance remains untested.
In briefMinimum-cost robust evidence certificates for Horn reasoning when sources fail together
SummaryThe paper studies the cheapest evidence certificates that keep a Horn-logic conclusion valid even if up to k source domains are withdrawn. It separates evidence records from asserted facts and uses transversals of minimal-support footprints to characterize robustness. An exact synthesis procedure is proven correct, terminating, and cost-optimal, with complexity results given. Experiments report up to 92% optimum coverage within practical resource bounds.
In briefActionRoPE: Treating Actions as World Coordinates for Consistent Control in Planar-Camera Game Models
SummaryThis paper addresses real-time interactive generation in planar-camera game worlds such as top-down and isometric games. It introduces ActionRoPE, a parameter-free interface that converts camera displacement caused by actions into world-coordinate rotary positions, so the same location keeps the same positional address over time. It also proposes Coordinate-Conditioned Noise Scheduling, which assigns denoising schedules based on whether regions were observed or are newly revealed. The authors report better adherence to commanded motion and a 66% reduction in revisit drift compared with the strongest of four baselines.
In briefAdaptive decoding for token-efficient reinforcement learning: widening candidate-token support during training to improve RLVR sampling efficiency.
SummaryThis paper studies whether the set of candidate tokens used in rollout decoding for reinforcement learning with verifiable rewards (RLVR) should stay fixed or widen during training. The main schedule moves from top-K=10 to K=20 and then to full support as the token budget grows. In three runs on a 4B code-RLVR setup, it outperformed a matched fixed-K=10 baseline by a mean of 1.80±0.42 points on interpolated 150M-token validation, and performed comparably to a separately trained fixed-K=20 baseline. The authors conclude that rollout support is an important control for sampling efficiency and call for further study of adaptive schedules.
In briefAre self-evolving AI agents able to keep learning new tasks without forgetting old ones?
SummaryThis paper examines whether large language model agents that improve by rewriting their own prompts, memory, and skills can learn continually as tasks change. The authors test three methods (ReasoningBank, GEPA, Meta-Harness) on four drifting task streams built from BIRD, BFCL, and ScienceWorld. They report that self-evolution often improves overall performance, but higher adaptability tends to reduce stability, so current self-evolving agents remain incomplete continual learners.
In briefStatistical learning theory for DPO: how misspecified preference models limit performance
SummaryThis paper studies how the Bradley-Terry modeling assumption in Direct Preference Optimization (DPO) limits its performance. It separates structural misspecification from model capacity and finite-sample estimation error, characterizing structural error via the triangle-curl of preference log-odds. The authors derive excess-risk bounds, approximation rates for ReLU score models, and finite-sample guarantees, and introduce contextual preference accuracy (CPA) as a related measure.
WhyCharacterizes DPO's structural misspecification via triangle-curl of preference log-odds, with matching upper and lower bounds and finite-sample guarantees.
In briefArabic Financial NLP: A guide from corpora to analysis
SummaryThis chapter examines how natural language processing can be used to study Arabic financial texts such as corporate reports, market disclosures, journalism and banking interactions. It follows the path from source collection and text processing to analysis of vocabulary, entities, causal attribution and multilingual reporting. The author also considers what financial summaries and retrieval-grounded answers preserve or lose, and reviews published Arabic resources. The chapter argues that linguistic context and documentary provenance should remain part of the evidence.
In briefREST: Reliability-guided spatial transcriptomics training to predict gene expression from H&E pathology images
SummaryThis paper proposes REST, a framework that uses spatial transcriptomics (ST) data to train models that predict molecular information from H&E-stained histopathology images. It builds a Biological Program Bank from gene signature, pathway, and cell-type sets, weights image-ST alignment by per-spot reliability, and uses learnable program queries to reconstruct gene expression from morphology. Trained on 209 samples and evaluated on 70 disjoint samples, REST achieved the highest Macro correlation among ST-supervised methods and outperformed existing pathology foundation models.
WhyProposes reliability-guided program-structured ST supervision, evaluated on large data and outperforming prior methods and pathology foundation models.
In briefMASIF: Implicit Failures as a Window into Multi-Agent Process Behavior
SummaryThis paper studies implicit failures (IFs), local execution errors in multi-agent systems that are corrected, bypassed, or carried forward without determining the final outcome. Analyzing 1,828 traces from nine benchmark configurations, the authors find IFs in 54.3% of successful and 75.2% of failed traces, and build taxonomies of error types and absorption mechanisms. They also introduce IF-GRPO, a reinforcement learning algorithm that uses IF costs to refine preferences among successful trajectories, reporting gains in accuracy and process metrics along with trade-offs between performance and IF count and cost.
In briefLearning Who Talks to Whom: Role-Aware Communication Scheduling for Efficient LLM Agent Teams
SummaryThis paper studies how LLM-driven multi-agent systems can decide dynamically which agents communicate with each other and when, to balance reasoning quality against token cost. The authors propose ACTS, a framework using centralized training with decentralized execution, with a role-conditioned relational attention module and a topological attention mixer. Across 9 benchmarks covering code generation, multi-step reasoning, general knowledge, and competitive mathematics, the paper reports accuracy gains of up to 6.19 percentage points and token reductions above 15%.
In briefBeyond Evaluation: Using Rubric Structure to Guide Reinforcement Learning
SummaryThis paper observes that rubrics used to grade open-ended answers are rarely used to guide training, even though they encode useful capability structure. The authors propose RSRL, a reinforcement learning framework that recovers capabilities from rubric criteria, uses them to strengthen underdeveloped skills, and preserves capability-specific gains during optimization. Reported experiments show gains of up to 7.43 over the initial policy and 7.60 over non-rubric structured RL baselines on scientific question answering.
In briefChecking whether open-vocabulary human-object interaction scores reflect real understanding
SummaryThis paper examines how open-vocabulary human-object interaction detectors are evaluated, since free-text descriptions make standard category scoring and semantic similarity scoring both imperfect. The authors propose a diagnostic protocol that scores the same predictions under five views, from official category evaluation to strict propositional equivalence. On HICO-DET across six systems, the leading system changes depending on the view, and semantic gains range from 1.6 to 20.3 mAP. The authors recommend reporting semantic scores alongside relational validity and paired uncertainty.
S M Asif Hossain, Ruksat Khan Shayoni, Md Kishor Morol
In briefACE-Trader: Adversarial Co-Evolution Framework for LLM Trader and Red-Team Agents
SummaryThe paper proposes ACE-Trader, a trading framework in which an LLM-based trader and a red-team agent that reviews its decisions learn together from shared trading records. Each agent uses a context-dependent memory selector trained on net-return differences with and without memory assistance. Experiments on U.S. stock trading report higher cumulative return than all 22 active baselines, and the best Sharpe ratio and maximum drawdown among active and passive methods.
In briefWhether language models can judge their own confidence, and how reliably we can measure it
SummaryThis paper examines whether large language models know what they know, comparing verbalized confidence, linear probes on hidden states, and agreement among sampled answers. The authors ran a pre-registered audit of five Qwen2.5 models on factual retrieval and multi-step reasoning tasks. They report that verbalized confidence is unstable across prompt templates and that Expected Calibration Error can be misleading, and two of five pre-registered hypotheses were falsified.
In briefSWIFT: Training stronger diffusion models by extrapolating targets from a frozen weak model
SummaryThe paper proposes SWIFT, a self-guided training method in which a frozen weak denoiser pushes the clean regression target away from its prediction, amplifying underfit components. It adds no inference overhead and keeps the original sampler. Experiments on CIFAR-10 and ImageNet-256 report improved FID across model scales, including one ImageNet-256 setting that surpasses the baseline's best FID using roughly half the total training compute.
In briefMeasuring How Seller Cues Sway AI Shopping Agents in Price Terms
SummaryThis paper examines how seller-controlled cues such as badges and endorsements influence the recommendations of large language models acting as shopping agents. The authors introduce price-equivalent auditing, which raises a cued product's price until it is chosen half the time, so different cues can be compared in monetary terms. Across 113,472 trials, seven models and two product categories, they find that choice-rate audits substantially under-resolve seller influence, and they release stimuli, logs, and code as a reproducible benchmark.
WhyProposes pricing cues to escape choice-rate saturation, validated over 113,472 trials across seven models and a new decision-native model.
In briefNot every rollout needs distillation: a correctness-filtered self-distillation method for cheaper training
SummaryThis paper examines whether every student-generated trajectory in on-policy self-distillation (OPSD) provides distinct supervision worth its cost. The authors propose C-OPSD, which generates one response per problem and distills only a subset of incorrect responses. On Qwen3-1.7B, 4B, and 8B across AIME24, AIME25, and HMMT25, C-OPSD reportedly matches full OPSD performance while training takes about 40% less time.
In briefOV-PEAR: Observation-verified pose and scale estimation for unknown industrial objects
SummaryThis paper presents OV-PEAR, a framework that reconstructs 3D models and estimates 6D pose and scale of unknown industrial objects so robots can manipulate them without CAD models or object-specific training. It works in three stages: segmenting RGB-D instances and generating 3D hypotheses, anchoring them to the robot frame, and verifying voxel occupancy with multi-view evidence to suppress hallucinations. On the BOP-Industrial benchmark it reports 0.657 average AR, and in real-robot connector assembly it reaches 99.67% grasp and 84.00% assembly success.
In briefWhere RL fails and how targeted pretraining repairs it
SummaryThis paper studies why reinforcement learning with verifiable rewards (RLVR) fails to improve language model reasoning. The authors find that failed rollouts are predicted by a small set of characteristic n-grams, and use them to retrieve text from the pretraining corpus for continued pretraining before RL. A 35M-token corpus outperforms continued pretraining on the same amount of randomly sampled text in 10 of 11 settings, and the retrieved text transfers across model families.
In briefDiffusion language models vary in generation speed by task and language.
SummaryThe paper analyzes two diffusion language models across 10 tasks and finds that forward passes per token depend heavily on the task. Copying or lightly editing input text is cheap, while open-ended writing is costly. Highly tokenized languages and wrong answers also need more passes per token, and weighting candidate answers by decoding speed outperforms majority voting.
Mohsen Fayyaz, Ali Modarressi, Daniel Israel et al.
In briefFrontierRefactor: A Benchmark for Performance-Focused Multi-Scale Code Refactoring
SummaryThis paper introduces FrontierRefactor, a benchmark of 60 engineering tasks across four scales, from computational kernels to full repositories, for testing coding agents on refactoring that preserves behavior while improving performance. Evaluating 1,080 execution runs, the authors find that functional correctness and measured speedup are largely decoupled, with 41.2% of submissions correct but lacking verified performance gains. Multi-turn execution-timing feedback raised strict success from 16.0% to 40.0%.
In briefHihaze: One-Step Diffusion for High-Resolution Real-World Image Dehazing
SummaryThis paper proposes Hihaze, a one-step diffusion framework that removes haze from high-resolution real-world images. It uses attentive skip compensation and a spatial degradation adapter to handle uneven haze, and a training-free global-guided patch fusion method to keep predictions consistent across patches. The authors report better visual quality and faster inference than prior methods on four real-world benchmarks, with PSNR gains of 3.62 dB on Dense-Haze and 3.69 dB on NH-HAZE over the strongest baselines.
WhyProposes one-step diffusion with training-free patch fusion, showing large PSNR gains over baselines on four real-world benchmarks.
In briefWhen Retrieval Partly Conflicts with Skills: Procedural Hallucination in Language Models
SummaryThis paper studies how language models, when following a skill's procedure while using retrieved task context, can adopt cues that violate the skill and produce answers that are supported by retrieval but invalid under the skill, which the authors call procedural hallucination. They introduce SourceClash, a benchmark of 1,271 bundles and 6,355 prompts across ten procedural domains, and report that across nine models valid-answer rates average 25.7% under conflicting retrieval versus 78.2% and 72.8% under two supportive settings. Mechanistic analysis of open-weight models attributes most of the repair to value vectors, and intervening on that pathway also improves final answer validity.
In briefOptimizing social-network marketing budgets via a generalized fractional incentives framework
SummaryThe paper studies how to allocate limited marketing budgets continuously across social network users who differ in responsiveness and influence. It introduces Budgeted Generalized Fractional Influence Maximization (BGFIM) and a scalable discrete greedy algorithm with provable approximation guarantees. Experiments on multiple large networks show continuous allocation can outperform standard heuristics.
Abhishek K. Umrawal, Akhil Bhimaraju, Eliot W. Robson et al.
In briefWhen to Stop Asking: Factor Coverage Bias in LLM Medical Determinability Judgments
SummaryThis paper studies whether large language models can correctly judge if the available information is sufficient for a clinical decision when information is incomplete. The authors built a controlled pairwise benchmark, adding decision-relevant or irrelevant text to determinable and undeterminable cases. Adding a relevant factor made models more likely to label undeterminable cases as determinable, and this effect persisted with increased reasoning effort.
Yusuke Watanabe, Yohei Kobashi, Taisei Tosaki et al.
In briefWhether language models can tell apart the meanings of modal verbs like "may" and "must" depending on context.
SummaryThis paper tests whether language models can distinguish the different senses of modal verbs such as "may" and "must," including root senses like permission and obligation and epistemic senses like knowledge and belief. The authors evaluate seven models differing in architecture, size, and training data across four datasets, using vector-based similarity and surprisal measures. Larger models trained on diverse text use additional context effectively and, without context, favor the sense dominant in corpus statistics. Smaller models, including those trained on child-directed language, show inconsistent sensitivity to modal senses and benefit less from context.
Ezgi Ba\c{s}ar, Yevgen Matusevych, Annemarie van Dooren et al.
In briefDynamic Dual-Feedback Conformal Inference for Time Series Forecasting
SummaryThis paper addresses uncertainty quantification in time series forecasting, where sequential data show temporal dependence and distribution shifts. It proposes Dynamic Dual-feedback Conformal Inference (DDCI), which combines actual feedback (whether the true value falls inside the interval) with estimated feedback from empirical conformal thresholds to limit excessive interval changes. According to the authors, DDCI yields more stable and narrower prediction intervals in sequential settings while keeping coverage near the nominal level in practice.
In briefIs It the RL Method or the Data? A Controlled Test of GRPO Versus Rejection-Sampling Fine-Tuning
SummaryThis paper compares GRPO and rejection-sampling fine-tuning (RFT) using the same prompts, verifier and token cap on a procedural counting task with Qwen3-4B-Base. GRPO outperformed RFT on every prompt set, by 10.9 points on held-out test problems with curated prompts. The authors note these are controlled measurements on one synthetic task family and one base model, with three seeds per arm.
In briefWhen preferred and stable argumentation semantics give the same results: a path-based method
SummaryThe paper studies structural conditions under which preferred and stable semantics agree in abstract argumentation frameworks. It introduces a general path-based criterion built from sets of paths in the attack graph and the binary relations they induce, and derives sufficient conditions stated via well-foundedness. These conditions strictly generalize Dung's limited controversiality and SCC-symmetry conditions, and they also yield conditions for the existence of stable extensions.
In briefCommunication-efficient federated domain generalization for edge networks using dynamic prototype regulation
SummaryThis paper addresses the challenges of deploying vision-language models over bandwidth-limited wireless edge networks, where large parameters and non-IID data hurt performance. It proposes DPGR-GCAM, which uses a dual parameter-efficient fine-tuning architecture and a channel-aware tiered transmission scheme to improve node survival under fading. A Gradient-Conflict Aware Modulation mechanism is introduced to reduce client drift within a low-dimensional feature manifold. The authors report improved cross-domain generalization and sustained convergence under network fading scenarios.
In briefCan Classical Gradients Replace Learned Targets for 2D Gaussian Initialization?
SummaryThe paper asks whether classical normalized Sobel gradient maps can replace the optimization-derived pseudo-PPM targets used by Instant GaussianImage. It proposes a training-free analytic initializer that produces initial parameters in about 0.32 seconds, and a neural initializer trained on DIV2K with online Sobel supervision that scales to 10,000 LSDIR images. Under domain shift, both approaches are reported to outperform Instant-GI on Manga109 and Kvasir-SEG at various budgets.
In briefSector-Ring Linear Reservoirs: Assigning Decay Timescales by Sector with State Reuse
SummaryThis paper proposes the Sector-Ring Linear Reservoir (SRLR), which fixes eigenvalue angles evenly around a circle and assigns decay timescales, chosen from a few values, to consecutive angular sectors. Input weights are scaled using the training input's power spectrum so that each mode's state scale is comparable across candidate settings. Mode states for each decay-timescale value are computed once and reused across candidates. On prediction and reconstruction benchmarks from chaotic systems, SRLR had lower mean test error than a single-timescale reference model in all eight cases and was competitive with echo state network baselines. The search computed 99.8% fewer mode trajectories than per-candidate recomputation.
In briefThinking before transcribing: context-guided reasoning for hotword-aware speech recognition
SummaryThis paper addresses contextual speech recognition, where candidate hotwords may be incomplete or wrong. The authors propose generating a short, audio-grounded caption before transcription so the model can judge candidate relevance first. The method combines supervised fine-tuning with targeted trajectory injection to strengthen group relative policy optimization on difficult utterances. Experiments across diverse context conditions report improved transcription quality and hotword recognition, with robustness to irrelevant candidates.
In briefAirMPA: A meteorology-to-pollution model for global air quality forecasts
SummaryThis paper proposes AirMPA, a decoupled autoregressive deep learning framework that separates meteorological and pollutant forecasting. It predicts 13 atmospheric pollutants at 0.4° spatial and 12-hour temporal resolution, using meteorological fields, emission priors, and historical concentrations. The authors report that it outperforms Aurora and CAMS in about 90% of pollutant-lead-time cases over a 5-day horizon, and can run on external meteorological forecasts.
In briefFair regression with controlled group prediction gaps: a method that bounds Wasserstein distances between groups' predictions to tune fairness against accuracy.
SummaryThis paper studies fairness in regression. Instead of requiring exact demographic parity, which forces identical prediction distributions across sensitive groups, it bounds the pairwise Wasserstein distances between group-specific prediction distributions by a chosen tolerance. On the real line, it characterizes solutions via transport and quantile optimization, and for squared loss with two weighted groups it derives soft-thresholding, shrinkage, and clipping forms. It also establishes rates for a known-cost plug-in estimator and describes an observed-loss proxy when the cost is unavailable. Experiments on synthetic data and California 2018 ACS data illustrate the tunable fairness-accuracy trade-off.
In briefSAR: Learning Global Item-Space Matching and Geometry-Aware Placement Ranking for Buffered Online 3D Bin Packing
SummaryThis paper addresses online 3D bin packing with a bounded buffer, where decisions must be made without knowing future item arrivals. It proposes SAR (Spatial Affordance Retrieval), a two-stage learned framework: a picker matches buffered items to residual container space using voxel-token encoding, and a loader ranks feasible placement candidates by geometry, then both stages are jointly optimized for long-horizon packing return. On the Random, JD, and SF datasets, the best SAR variants improve utilization over strong online baselines by 0.96 to 4.20 percentage points without a buffer, and reach up to 93.55%, 91.04%, and 90.41% utilization with a buffer of five items.｡
In briefCoMem: Learning to Predict Failures from Retrieved Past Experience
SummaryThis paper introduces CoMem (Compounding Memory), a method that makes retrieved past experience an input to an LLM agent's learned memory, both during training and at inference. The agent builds a bank of its own successful and failed reflections, learns to write failure hypotheses from retrieved cases, and at test time uses newly retrieved cases to guide each attempt. Across two backbones and five benchmarks in math, code, and multi-hop QA, the authors report strong overall performance. They also report generalization to out-of-distribution tasks and that smaller models can guide stronger solvers.
In briefCan LLM User Proxies Replace Humans in Measuring AI Reliance?
SummaryThis paper audits whether language-model user proxies can reproduce human reliance on AI answers, using the Rel-A.I. study as a test case. The authors replicate its 90 questions and 15 domain-by-marker reliance rates with open models up to 72B parameters, using a crossover design to separate question content, label averages, and their interaction. Matching the agent's specialty to the question domain raised reliance by 13.3 points on average, while the best A/B proxy showed large errors against the human data. The authors conclude that this single-study audit does not establish proxy validity for human effect sizes.
In briefUniJev: Separating accuracy from confidence calibration in three-modality decision models
SummaryThe paper presents UniJev, a framework that scores candidate answers with a shared head trained jointly on text, images, and video. Supervised decision models improve accuracy over their backbone across all three modalities, while existing reinforcement-based calibration methods change accuracy little and have inconsistent effects on visual calibration. The authors introduce Cora, which groups predictions by confidence and aligns each group's confidence with its empirical accuracy, lowering image and video calibration error at 2B and 9B scales while preserving accuracy.
In briefDetection Augmented Learning for Bandits in Changing Environments
SummaryThis paper studies non-stationary multi-armed bandit problems where environmental changes are unknown. The authors propose Detection Augmented Learning (DAL), a modular method combining a bandit algorithm, a change detector, and a forced exploration schedule. A single regret bound shows DAL transfers the regret of a stationary base algorithm to piecewise stationary settings and matches or improves known bounds. Experiments across nine bandit settings report that DAL consistently outperforms state-of-the-art methods.
Argyrios Gerogiannis, Yu-Han Huang, Subhonmesh Bose et al.
In briefPredicting Latent Futures to Improve Hyperspectral Image Classification
SummaryThis paper proposes Stochastic Latent World Modeling (LWM), a framework that improves hyperspectral image classification by modeling dependencies in the latent representation space. It encodes spectral and spatial information with separate transformer branches and predicts multiple future latent states through stochastic transitions. The predicted trajectory is combined with the base representation for classification. Across five benchmark datasets, LWM reports overall accuracies from 89.21% to 99.52% under limited-label settings.<br>
Muhammad Ahmad, Gemine Vivone, Ronny H\"ansch et al.
In briefOpenSaC: A Python Library for Search-as-Code in Deep Research Agents
SummaryDeep research agents answer complex questions by repeatedly searching, reading sources, and synthesizing evidence. This paper proposes Search as Code, in which agents write programs that combine search, source reading, reranking, model calls, and local processing. The authors release OpenSaC, an open-source Python library with a companion skill, and compare it with standard Codex built-in web search on BrowseComp and WideSearch.
In briefA study on giving AI fact-checking agents professional expert skills to improve evidence search and claim verification.
SummaryThe paper distills publicly available Agence France-Presse (AFP) Fact Check guidelines into a reusable Professional Fact-Checking Skill (PFC-Skill). It introduces SkillCheck, which lets LLM agents use this skill at inference time to guide evidence retrieval and claim verification. On the AVeriTeC benchmark, SkillCheck improves accuracy and Macro-F1 over skill-free DEFAME baselines across several LLM backbones.
In briefFrom Trace to Fix: Residual Task State for Recovering Long-Horizon Agents
SummaryThis paper addresses recovery from failures in long-horizon, stateful language agents, where the causing action, its symptom, and the state needing repair often differ. It proposes TRACE-FIX, a framework built on versioned residual task state that turns execution logs into a dependency-linked ledger and plans repairs under a budget. The authors also introduce LongH-Debug, a benchmark that separately evaluates root-cause attribution, verified-fault recovery, and end-to-end closed-loop execution.
In briefVHA: VQA-Based Hallucination Assessment for Generative Image Super-Resolution
SummaryGenerative image super-resolution can produce realistic high-resolution images but may add unsupported content or alter existing details, a problem called hallucination. The authors propose VHA, a bidirectional visual question answering method that uses a vision-language model to check whether a super-resolved image and its ground truth share the same semantic and low-level visual evidence. They also develop VHA+, a lightweight model that predicts VHA scores directly, and build a human-annotated benchmark, VHA-bench. Both VHA and VHA+ are reported to correlate highly with human judgments of hallucination.
In briefEvaluating a Fast Decision Model as a Reader of Cut-Short AI Reasoning Under a Token Budget
SummaryLarge language models often produce more reasoning tokens than necessary, and capping the budget cuts cost but can lower accuracy when forced answers come from incomplete reasoning. The paper evaluates Jev, a System-1 decision model that selects among options without generating reasoning, across seven reasoners and five multiple-choice benchmarks. Jev achieves higher mean accuracy than forced answering at budgets from 500 to 16,000 tokens, including reasoning and answering costs.
In briefEvalRx: Agentic Diagnosis and Verified Repair of Foundation Models
SummaryThis paper presents EvalRx, an open-source agentic framework that diagnoses and repairs AI models using statistical methods. It finds systematic failure patterns, forms testable hypotheses, checks them with frozen pre-specified tests, and applies tiered repairs. Repairs are accepted only if they improve the target behavior without unacceptable regression. The authors report gains in all 12 evaluated model family and modality combinations, ranging from +1.4 to +10.9 percentage points.
In briefTRACE: A Method That Tracks Multimodal Clinical Records to Predict Hospital Patient Risk
SummaryThis paper proposes TRACE, a framework that organizes hospital clinical events from health records, chest x-rays, radiology reports, and discharge summaries into daily trajectories. It encodes these trajectories with temporal attention to predict in-hospital mortality, length of stay, and readmission. Tested on nearly 180,000 stays from MIMIC-IV, eICU, and MIMIC-III, TRACE achieves the highest AUROC across all nine dataset-task settings. Trajectory-level attribution also identifies risk-associated hospital days and modalities.
In briefActTime: A Action-Conditioned World Model for Time Series Forecasting
SummaryThis paper introduces ActTime, a world model for time series that conditions predictions on actions that can be changed before execution. It updates latent states from observed history using an action-conditioned transition and rolls the same transition forward under candidate actions to generate future trajectories, trained only on factual data. Evaluated on simulator-grounded tasks in building energy, traffic, greenhouse climate, and inventory management, the authors find that factual accuracy, intervention fidelity, and decision quality are related but not interchangeable.
In briefReasoning at Each Step, Not Memory: How LLM Agents Track Changing States
SummaryThis paper studies how a language-model agent's actions depend on its reasoning in a file-system task where the correct action depends on changing lock states. Using Qwen3-8B and Qwen3.5-9B, the authors separate generating reasoning before each action from retaining earlier reasoning in the prompt. Most of the reduction in lock-violating commands comes from reasoning generated at the moment of action, and the content of that reasoning matters more than filler or an accurate state table.
