The paper titled "Lost in Motion: Vision Language Models Fail the Dynamic Gauges Test" examines how current Vision‑Language Models (VLMs) perform on real‑time gauge reading tasks in industrial settings. Using a new dataset of video sequences featuring circular, linear, and Vernier gauges under varied motion, the study tests state‑of‑the‑art models from GPT‑5 and Gemini 3 families. Results show that these VLMs struggle to accurately interpret needle trajectories and scale semantics, falling short of the traceability and reliability required for safety‑critical monitoring and thus not meeting IEEE and ISO standards for trustworthy synthetic instruments.
By Tairan Fu, Francisco Javier Santos-Mart\'in, Javier Conde, Elena Merino-G\'omez, Pedro Reviriego
TOPAS is a Task‑Oriented Prefix‑Aware Scheduler designed for multi‑agent large language model serving. It jointly decides which agent prefixes to retain in a shared key‑value cache and which requests to schedule, balancing the reduction of each task’s longest remaining service path against the benefit of downstream prefix reuse while accounting for movement and preemption costs. Experiments on synthetic DAGs and MetaGPT software‑development workflows show that TOPAS can reduce mean and p99 job completion times by up to 39.8%/49.4% and 22.0%/26.6% respectively compared to the best baselines.
By Hongqiu Ni, Han Tian, Chi Zhang, Guopeng Li, Haisheng Tan
VisDocAgentBench is a closed‑corpus benchmark that evaluates static versus agentic retrieval for visually rich documents, using 2,375 pages from 100 documents and 120 queries that span direct, one‑bridge, and two‑bridge evidence structures. The benchmark includes semantic, relational, and visual queries, full‑document review, and hard‑negative validation. Results show that a strong visual retriever performs well on direct items but poorly on two‑bridge items, while agents improve performance, especially when using visual retrieval and iterative search capabilities.
By Lexiang Hu, Yanzhao Zhang, Mingxin Li, Dingkun Long, Yikang Li, Fuwei Zhang, Yisen Wang, Zhouchen Lin
The paper introduces SalesLLM, a bilingual (Chinese/English) benchmark for evaluating large language models (LLMs) in realistic sales dialogues. It comprises 30,074 scripted configurations and 1,805 curated multi‑turn scenarios from Financial Services and Consumer Goods, with controllable difficulty and personas. An automatic evaluation pipeline uses an LLM judge for sales‑process progress and fine‑tuned BERT classifiers for end‑of‑dialogue buying intent, while a user model, CustomerLM, is trained to improve simulation fidelity. SalesLLM scores correlate strongly with human ratings (Pearson r = 0.86) and reveal that top Chinese LLMs match junior‑to‑intermediate human salespeople but not experts, with cross‑lingual consistency remaining poor.
By Xuanbo Su, Wenhao Hu, Le Zhan, Yuting Xie, Kailin Lyu, Kaijie Chen, Ziwei Li, Yeqiang Wang, Haibo Su, Yunzhang Chen, Ling Huang
The paper investigates how explicit reasoning in Large Reasoning Models (LRMs) affects their ability to persuade and be persuaded. Experiments on objective and subjective tasks reveal a Persuasion Duality: reasoning boosts an agent’s persuasive power by about 21 percentage points while also making it less susceptible to incorrect persuasion by up to 10 percentage points. However, the study finds that persuasiveness often relies on superficial cues like response length and repetition rather than logical validity, and that persuasion can amplify or attenuate non‑linearly across multi‑hop agent chains. The authors also propose an attention‑guided prompt‑level adversarial argument detection method that improves agent robustness.
By Haodong Zhao, Jidong Li, Zhaomin Wu, Tianjie Ju, Zhuosheng Zhang, Bingsheng He, Gongshen Liu
SAG (SQL‑Retrieval Augmented Generation) is a structured retrieval framework that indexes documents as event‑entity pairs, forming latent hyperedges that preserve n‑ary relations without building a global knowledge graph. At query time, shared entities act as join keys, dynamically creating a query‑scoped neighborhood of related events while keeping each evidence chunk intact. Experiments on HotpotQA, 2WikiMultiHopQA, and MuSiQue demonstrate that SAG outperforms existing dense‑retrieval baselines, achieving the highest recall and end‑to‑end QA performance, especially as reasoning‑chain complexity grows.
By Yuchao Wu, Junqin Li, Xingcheng Liang, Yongjie Chen, Yinghao Liang, Linyuan Mo, Guanxian Li
The paper introduces a scalable product‑linking system that uses a retrieve‑then‑match cascade. First, a lightweight text cross‑encoder auto‑resolves the majority of merchant‑catalog product pairs with high precision, while an agentic multimodal vision‑language model handles the remaining ambiguous cases by inspecting images and performing web searches. This approach balances computational cost and accuracy, improving overall link coverage from 68% to 77% without requiring fine‑tuning of the agent.
By Jian Wang, Steven Xu, Sanjyot Thete, Maryam Barouti, Tom Tang, Elaine Wu, Charu Sareen, Kyle MacDonald
The paper introduces the Hybrid Hierarchical Multi-Agent Framework (H$^{2}$MAF), which fuses decision-level outputs from EfficientNet-B3 and ConvNeXt-Tiny with semantic arbitration by multimodal large language models Gemma 4 E4B and Qwen3.5 4B to produce explainable plant disease diagnoses. Evaluated on 14,364 images from PlantDoc and two Cornell robotic field datasets, the framework achieves up to 99.3% accuracy, with Gemma improving PlantDoc accuracy from 63.9% to 68.5% and demonstrating low critical‑risk error. The results highlight the potential of MLLM arbitration for reliable, explainable agricultural AI under real‑world field conditions.
By Ranjan Sapkota, Konstantinos I. Roumeliotis, Pengyao Xie, Nikolaos D. Tselikas, Lirong Xiang, Manoj Karkee
The paper introduces KOPE, an experience‑driven framework that records hardware kernel optimization trajectories in an Experience Graph Memory and uses Active Context Management and Injection to retrieve relevant past decisions under a fixed token budget. KOPE preserves decision order, outcomes, and alternative branches, enabling evidence from completed runs to inform future optimization steps. In experiments, KOPE achieves a 1.54× speedup over the strongest baseline, raises pass rates from 60.0% to 84.6%, and reduces token consumption dramatically, demonstrating the benefits of continual learning from external experience while keeping the foundation model unchanged.
By Siyuan Chen, Runlin Hou, Shenxiu Wu, Yansong Sun, Junming Cao, Yiyu Zhang, Shudi Shao, Junhao Qiu, Zhichao Lu, Qingfu Zhang
The paper presents 3DGAA, a fabrication-first framework that generates view-consistent, geometry-preserving adversarial wraps for vehicles using 3D Gaussian splatting optimization. It ensures consistency across viewpoints, illumination, and occlusion while limiting changes to vehicle geometry, producing realistic print-only textures that significantly reduce detection confidence and average precision in simulations and physical tests. Ablation and efficiency studies analyze the impact of physical filtering, augmentation, and shape-consistency regularization, and the method demonstrates robustness against common preprocessing defenses and cross-detector transferability.
By Yixun Zhang, Lizhi Wang, Junjun Zhao, Wending Zhao, Feng Zhou, Yonghao Dang, Jianqin Yin
JIT‑Agent is a model that automatically generates task‑adaptive agent harnesses for any off‑the‑shelf LLM, replacing manual, task‑specific harness design. It learns to compose, repair, and evolve harnesses using a fixed four‑module protocol, and its use boosts performance on benchmarks such as DeepSearchQA and OdysseyBench, outperforming several mature agent runtimes. The approach demonstrates that harness intelligence can be trained, transferred, and compounded independently of model scaling.
By Guibin Zhang, Leo Lu, Fangzhou Xie, Kang Zhu, Junhao Wang, Zhifei Xie, Zhaochen Yu, Zihang Liu, Zhongxiang Sun, Qiankun Li, Yue Liao, Heng Chang, Xiaobin Hu, Qibing Ren, Wangchunshu Zhou, Shuicheng Yan
EduAgentQG is a multi‑agent framework for generating personalized mathematics questions that explicitly controls diversity and aligns with educational objectives. It operates through a closed‑loop cycle of planning, writing, evaluation, refinement, and checking, using fine‑grained evaluation to ensure logical correctness, solvability, and alignment with knowledge concepts, difficulty, grade level, and core competencies. The authors built a benchmark of 10,273 questions across Grades 1‑9 and demonstrated that EduAgentQG outperforms existing methods in diversity, objective consistency, and win rate.
By Rui Jia, Min Zhang, Fengrui Liu, Bo Jiang, Kun Kuang, Zhongxiang Dai
Latent-CoT-Drive (LCDrive) is a vision‑language‑action model for autonomous driving that replaces natural‑language chain‑of‑thought reasoning with a latent language capturing possible outcomes of driving actions. The model interleaves action‑proposal tokens, aligned with the model’s output actions, and world‑model tokens grounded in a learned latent world model to reason about future outcomes. After a supervised cold‑start using ground‑truth future rollouts, LCDrive is further refined with closed‑loop reinforcement learning, achieving faster inference, higher‑quality trajectories, and greater gains from interactive RL than both non‑reasoning and text‑reasoning baselines on a large‑scale end‑to‑end driving benchmark.
By Shuhan Tan, Kashyap Chitta, Yuxiao Chen, Ran Tian, Yurong You, Yan Wang, Wenjie Luo, Yulong Cao, Philipp Krahenbuhl, Marco Pavone, Boris Ivanovic
The paper introduces AgentDiff, a metric that quantifies how much LLM agents’ answers differ when inputs are altered by meaning‑bearing rewrites (paraphrases, synonym substitutions) versus presentation changes (reordering, formatting, distractors). Across 68 model–benchmark–scaffold combinations involving ten LLMs and over 1,500 questions, meaning‑bearing rewrites consistently produce a roughly 20‑percentage‑point higher inconsistency rate than presentation changes, a gap that persists across severity proxies and remains significant even outside the Qwen family. Trace analysis reveals that meaning‑bearing rewrites preserve the first action but reduce thought similarity from the second step onward, extending the divergence cascade—a phenomenon termed “stealth divergence.”
By Liyun Zhang, Jiayi Guo
The paper introduces RoMem, a temporal knowledge graph module that treats time as continuous phase rotation rather than discrete labels. RoMem uses a Semantic Speed Gate to assign volatility scores to relations, allowing evolving facts to rotate quickly while persistent facts remain stable, thereby preventing the need for deletion or costly LLM calls. The method achieves state‑of‑the‑art performance on ICEWS05‑15 and improves temporal reasoning in agentic memory benchmarks such as MultiTQ, LoCoMo, and FinTMMBench.
By Weixian Waylon Li, Jiaxin Zhang, Xianan Jim Yang, Tiejun Ma, Yiwen Guo
The paper introduces Gavel, a framework for evaluating large language models (LLMs) on long-context legal summarization tasks. Gavel includes a reference-based component (Gavel-Ref) with checklist, residual-fact, and writing-style checks, and a reference-free component (Gavel-Agent) that assesses factual coverage directly from source documents. Experiments on 12 frontier LLMs reveal that models tend to omit key information more than hallucinate, perform well on simple checklist items but struggle with rare, complex items, and their performance degrades with longer cases. Gavel-Agent cuts token usage by at least 36% compared to traditional methods while maintaining competitive accuracy, and it also generalizes effectively to the medical domain.
By Yao Dou, Benjamin Mamut, Wei Xu
The paper investigates how agents that inherit consolidated memory can mistakenly rely on constraints that were once true but have since been superseded by newer records. By modeling supersession explicitly and limiting verification to two records, the study shows that native allocation often leads to stale-consistent decisions, while reallocating one verification slot to the critical provenance path markedly improves consistency. The results suggest that memory systems may need separate freshness or supersession signals beyond relevance to avoid such errors.
By Kazuki Nakayashiki
The paper introduces P4-DT, a personalized patient preference predictor that uses dilemma training to elicit context‑dependent decision reasoning. In a study of 12 patient‑surrogate pairs, P4‑DT achieved 81.7% accuracy in predicting patient treatment choices, outperforming unassisted surrogates (55.0%) and surrogates aided by a simpler P4 model (61.7%). The authors show that incorporating contextual scenarios and open‑ended text into prompts improves accuracy by 15 percentage points over static value ratings.
By Natasha Ureyang, Sebastian Porsdam Mann, Yuxin Liu, Zuriel Hassirim, Melanie Almonte, Wenhao Chen, Joyce Ng, Thant Nay Lin, Aung Thiha, Gerald CH Koh, Brian David Earp, Pin Sym Foong
The paper introduces a novel framework that combines vision‑language model (VLM) generated preferences with the Plackett‑Luce (PL) ranking model for reward learning in reinforcement learning. Unlike traditional pairwise Bradley‑Terry approaches, the PL formulation allows listwise rankings of multiple candidates, enabling the use of different ranking sizes (K = 3, 4, 5). Experiments on Meta‑World manipulation tasks show that PL‑based reward models train robotic policies as effectively as, or better than, pairwise, K‑wise, and RL‑VLM‑F baselines, achieving up to an 86% mean final success rate and matching the Oracle baseline on the Drawer Open task.
By Srivalli Katkuri, Maxwell Kawada, Juan Wachs
The paper introduces AWM, a framework that treats the terminal working memory of long‑document VQA agents as an answerable evidence artifact. It proposes a memory‑only answerability diagnostic and incorporates this signal into the GRPO reward, giving higher advantage to trajectories whose final memory can answer the question alone. Experiments on MMLongBench‑Doc and LongDocURL show that AWM‑GRPO boosts final‑answer accuracy by up to 11.9 points and reduces the rate of correct answers that cannot be supported by memory alone.
By Dongzhuoran Zhou, Yuqicheng Zhu, Yule Liu, Zhen Yang, Rui Lu, Yuxiao Dong, Jie Tang, Evgeny Kharlamov