The paper introduces CAVE, a structured credit assignment method that enhances Vision‑Language Models by evaluating intermediate reasoning steps through belief updates, evidence acquisition, and adaptive focus control. CAVE is built on a GRPO framework and guides models to optimize each action, improving reliability in fragmented visual reasoning. The authors also present TRACER‑Bench, a new benchmark covering nonlocal and semantically confusable reasoning dimensions, and show that CAVE boosts performance on both public datasets and this benchmark while maintaining strong general multimodal results.
By Tengda Guo, Jie Leng, Hanlei Li, Yaoyuan Liang, Qingyue Zhang, Dian Yang, Mingyu Zhang, Yuhua Fu, Shao-Lun Huang
TIDE (Tree-Induced Diagnostic Evaluation) transforms diagnostic evaluation into a behavior-driven decision-tree induction process, enabling fine-grained capability diagnosis of large language models. By recursively building a hierarchical capability tree from benchmark samples and model scores, TIDE identifies nuanced contrasts between models that aggregate metrics miss—such as Claude Opus 4.6 outperforming GPT-5.4 on microeconomic market reasoning while falling behind on contract law reasoning. The method leverages LLM judges to propose splits and uses behavior-based information gain to select them, demonstrating its effectiveness across five models and four benchmarks.
By Jiayi Chen, Hui Li, Yujiong Shen, Qiyuan Peng, Yanxin Li, Chenhao Huang, Shaofan Liu, Guoqiang Zhang, Junjie Ye, Shihan Dou, Ming Zhang, Maxm Pan, Tao Gui, Qi Zhang, Xuanjing Huang
ASTEROID is a student–teacher supervised learning framework that learns efficient test‑time exploration in partially observable sequential decision problems. By training a history‑conditioned student on on‑policy contexts while a privileged expert provides action labels conditioned on that history, the student learns to search while uncertain and exploit once the hidden state is revealed. Across eleven diverse environments, ASTEROID achieves roughly twice the returns under fixed exploration budgets, uses about 100× fewer interactions than reinforcement learning baselines, and transfers well to vision‑denied real‑world robot manipulation, improving success by +40% over standard distillation.
By Sriyash Poddar, Yanda Bao, Jacob Krantz, Matthew Chang, Xavier Puig, Roozbeh Mottaghi, Natasha Jaques, Abhishek Gupta
The paper reviews and categorizes large language models that combine step‑by‑step reasoning with interactive search, treating retrieval as an action within an iterative agentic loop. It analyzes these systems along four architectural axes—retrieval coupling, controller design, knowledge source, and grounding strictness—while also considering the learning signal as a cross‑cutting factor. The authors conduct controlled ablations and benchmark analyses to clarify how design choices affect search, stopping, and grounding behavior, aiming to provide a unified perspective for future research.
By Biswas Poudel, Nilson Chapagain, Amit Kumar, Xianshun Jiang
Residual-Based Causal Discovery via Pretrained In-Context Quantile Regression proposes Loscar, an extension of the RESIT framework to location-scale noise models (LSNMs). Loscar estimates conditional mean and scale using quantile regression, allowing pretrained tabular foundation models to serve as plug‑in estimators. Experiments on bivariate and multivariate causal‑discovery benchmarks show that in‑context quantile regression outperforms a feasible generalized least squares (FGLS) variant, especially when causes are multivariate.
By Daniel Klippert, Quang Nhat Huynh, Alexander Marx
The paper identifies a phenomenon called principle-level search collapse in feedback-driven large language model (LLM) optimizers, where the search narrows to a limited set of design principles despite ongoing candidate generation and minor improvements. Using a diagnostic benchmark across molecular optimization, network sentinel-node selection, and GPU-kernel optimization, the authors find that 78.3% of runs exhibit collapse, with 55.3% showing strong support. They evaluate common remedies such as diversity injection and reference-principle interventions, noting limited and inconsistent benefits, and emphasize the need for online collapse detection and improved mechanisms for generating and maintaining alternative principles.
By Nian Ran, Tianhe Jia, Zhengyang Mu, Yue Wang, Sebastian Brulin, Markus Olhofer, Mauricio A. Alvarez, Yang Liu, Richard Allmendinger
SANE is an agent‑based framework that automatically generates scorer‑specific perturbation recipes and expected score relations for large‑language‑model judges, using behavioral requirements encoded as reusable agent instructions. It validates each recipe against the scorer’s intent and the instantiated test case, producing a growing inventory of aligned recipes as more scorers are added. In experiments across five domains, SANE yields 3.7× more validated test families than the Judge Reliability Harness, detects deficiencies in all domains—including professionalism and toxicity—and achieves strong human alignment in judge ranking on faithfulness (Spearman ρ = 0.852).
By Ethan Ludwick, Yu-Chung Hsiao
The paper introduces OACap, a caption-guided active observation system for detailed video captioning that requests targeted observations to fill missing information and iteratively refines its captions. OACap consistently improves performance on general video-understanding benchmarks while using only a text‑only reader. The authors also present FSCap‑Bench, a new evaluation suite with Core‑Support Factuality Score and Cross‑Modal Semantic Round‑Trip Score to assess factual and semantic preservation in captions.
By Hao Shi, Hong Yan, Xiaoyang Zhang, Qing Cui, Chao Huang
MINER is a model-agnostic framework designed to improve clothes-changing person re-identification by suppressing garment bias. It works by contrasting hard examples and adaptively balancing intra- and inter-person variance in the dataset. Experiments demonstrate that MINER achieves state-of-the-art performance on four CC-ReID benchmarks and ranks among the leaders in all three tracks of the open TVRID RGB-Depth ReID challenge.
By Oron Nir, Eliyahu Strugo, Ariel Shamir
The paper introduces a contextual bandit algorithm that incorporates concavity constraints into a Bayesian Gaussian Process posterior. By combining a concavity-preserving regression spline basis with a constrained GP, the authors create a tractable shape-constrained estimator and a UCB algorithm. The resulting method achieves regret bounds that are at least as good as standard GP-UCB and can be tighter when concavity information is informative, as shown by experiments on benchmarks and a Warfarin dosing application.
By Kevin Li, Eric Laber
The paper introduces a trace-level divergence framework for evaluating multilingual large language models, breaking down agreement into structural similarity, order consistency, and holistic similarity. Experiments across three reasoning benchmarks, seven languages, and six leading LLMs show that conventional holistic similarity metrics overstate agreement, with up to 67.3% of top-scoring examples lacking structural alignment. Human judgments align more closely with structural similarity, highlighting language-specific divergence and the need for trace-level evaluation in multilingual deployments.
By Rui Sun, Peizhao Mei, Xiewei Cheng, Kaizhuo Chen, Taoran Li, Zhiyuan Yu, Varun Chandrasekaran
ClusterFusion is a hybrid text‑clustering framework that uses a large language model (LLM) as the core clustering engine, guided by lightweight embedding methods. The process involves embedding‑guided subset partitioning, LLM‑driven topic summarization, and LLM‑based topic assignment, allowing direct incorporation of domain knowledge and user preferences. Experiments on three public benchmarks and two new domain‑specific datasets show that ClusterFusion achieves state‑of‑the‑art performance on standard tasks and significant gains in specialized domains, with the authors releasing a new dataset and benchmark results.
By Yiming Xu, Yuan Yuan, Vijay Viswanathan, Graham Neubig
CanLegalRAGBench is a Canadian legal QA benchmark that uses realistic queries and expert‑annotated answers grounded in case law to evaluate retrieval‑augmented generation (RAG) systems. The study finds that retrieval performance varies with design choices and that open‑source embedding models can compete with closed‑source ones. It also highlights limitations of automatic evaluations, noting that RAG systems often produce hallucinations or overly detailed responses, with 8‑29% of claims unsupported by retrieved documents.
By Ethan Zhao, Maksym Taranukhin, Wei Cui, Moira Aikenhead, Vered Shwartz
MARS is a scalable knowledge‑graph question‑answering system that integrates large language models with knowledge graphs without requiring model fine‑tuning. It uses a structured retrieval process that links question entities to the graph, iteratively fetching next‑hop information and deciding when to stop and generate a SPARQL query. Evaluated on three KGQA benchmarks—including multilingual settings—MARS matches state‑of‑the‑art performance while remaining efficient and scalable.
By Nikit Srivastava, Daniel Vollmers, Ren\'e Speck, Nikolaos Karalis, Hamada M. Zahera, Axel-Cyrille Ngonga Ngomo
The paper introduces ALIAS, a method to assess memorization in large language models (LLMs) when generating TLA+ specifications. By applying consistent identifier renaming (CIR) and using a TLA+ model checker as an oracle, the authors show that renaming significantly reduces the correctness scores of models like Claude Opus 4.8 and GPT‑5, revealing the extent of memorization. The study also demonstrates that grading based on exact names inflates scores, and that fine‑tuned models lose a substantial portion of their training advantage when identifiers are renamed.
By Arslan Bisharat, Mujtaba Nazari, Khushboo Bhadauria, Brian Ortiz, Eric Spencer, Beatriz Santos, TaiNing Wang, George K. Thiruvathukal, Konstantin Laufer, Mohammed Abuhamad
LitePT introduces a lighter yet stronger 3D point cloud backbone that strategically uses convolutional layers in early, high‑resolution stages and switches to attention in deeper, low‑resolution layers. By adding a novel, parameter‑free 3D positional encoding called PointROPE, LitePT preserves spatial layout while reducing redundancy. The resulting model has 3.6× fewer parameters, runs twice as fast, and uses half the memory of the state‑of‑the‑art Point Transformer V3, yet matches or surpasses its performance across multiple tasks and datasets.
By Yuanwen Yue, Damien Robert, Jianyuan Wang, Sunghwan Hong, Jan Dirk Wegner, Christian Rupprecht, Konrad Schindler
The paper argues that AI‑generated image detection has focused too much on model design and not enough on how supervision is structured. By replacing blended supervision with orthogonal, disentangled supervision—semantic signals for a vision‑language model and artifact signals for a vision expert—the authors show a 9.5% improvement on five in‑the‑wild benchmarks. They also introduce the AIGI‑Now benchmark to evaluate detectors on eight modern generators with separate semantic and pixel‑level tests.
By Ruoxin Chen, Jiahui Gao, Kaiqing Lin, Keyue Zhang, Yandan Zhao, Isabel Guan, Taiping Yao, Shouhong Ding
AutoExpert proposes a new benchmark for automating 3D LiDAR annotation using expert‑crafted guidelines, redefining evaluation protocols and re‑annotating nuScenes and PandaSet datasets. The pipeline combines 2D detection and segmentation in RGB images, lifts these detections into 3D with known sensor poses, and generates 3D cuboids, improving 3D detection mAP from 12.1 to 25.4 on the AutoExpert‑nuScenes benchmark.
By Yechi Ma, Wei Hua, Shu Kong
ColonAdapter is a self‑supervised fine‑tuning framework that adapts geometric foundation models for 3D geometry estimation from monocular colonoscopy images. It incorporates a Detail Restoration Module and geometry consistency loss to improve low‑texture region performance and scale consistency, and uses a confidence‑weighted photometric loss for training stability. Experiments on synthetic and real datasets show state‑of‑the‑art results in camera pose estimation, monocular depth prediction, and dense 3D point map reconstruction without needing ground‑truth intrinsic parameters.
By Zhiyi Jiang, Yifu Wang, Xuelian Cheng, Zongyuan Ge
The paper introduces Delay‑Guided Attention for Multi‑stage Forecasting (DAMF), a method that explicitly models inter‑stage lag distributions in multistage manufacturing to improve quality prediction. DAMF uses a lag predictor to estimate the probability and magnitude of delays between adjacent stages, incorporates these soft distributions into cross‑stage attention, and employs gated residual fusion for information flow. The approach is evaluated on the LiquidSugar dataset and two controlled‑delay benchmarks, showing better quality prediction and more accurate lag recovery compared to baselines.
By Jiarui Zhang, Zhaoyang Guan, Zhixuan Xia, Naichen Shi