The paper introduces Driving with DINO (DwD), a framework that uses Vision Foundation Module (VFM) features to bridge simulation and real-world domains for autonomous driving video generation. It addresses the consistency‑realism dilemma by projecting VFM features onto a principal subspace, dropping high‑frequency texture elements, and applying a Random Channel Tail Drop to preserve structural detail. Additional components— a learnable Spatial Alignment Module and a Causal Temporal Aggregator— enhance control precision, spatial alignment, and temporal stability, reducing motion blur and ensuring realistic, consistent outputs.
By Xuyang Chen, Conglang Zhang, Chuanheng Fu, Zihao Yang, Kaixuan Zhou, Yizhi Zhang, Yanfeng Zhang, Mingwei Sun, Zhen Dong, Xiaoxiao Long, Zengmao Wang, Liqiu Meng
BC-Bench is a new benchmark for assessing agentic engineering systems on the domain‑specific language AL used in Microsoft Dynamics 365 Business Central. It contains 101 manually curated tasks drawn from real Microsoft production repositories and evaluates not only code generation but also test creation and multimodal problem statements. The study shows that performance gaps between models are larger than those between different agent harnesses, and that gains on general‑purpose benchmarks do not reliably transfer to the AL domain.
By Haoran Sun, Klaus Marius Hansen
The paper introduces BUMEX, a reinforcement learning exploration strategy that leverages a set of prior models containing the true transition kernel and reward function. By optimizing over this model set, the method derives upper and lower bounds on the Q‑function to guide exploration, providing theoretical guarantees of convergence to the optimal policy. When the model set follows a bounded‑parameter MDP structure, the optimization becomes convex, enabling finite‑time convergence under mild assumptions and demonstrating accelerated learning in simulations.
By J. S. van Hulst, W. P. M. H. Heemels, D. J. Antunes
The paper introduces an LLM-as-a-Judge framework for evaluating the outputs of an agentic drug discovery assistant, ChatInvent, deployed at AstraZeneca. It defines four quality dimensions—Completeness, Relevancy, Structural Clarity, and Scope Adherence—alongside deterministic Tool Call Correctness checks, and validates the judge against five expert annotators. After optimizing the best-performing judge with few-shot demonstrations, alignment with human majority votes improves from 0.80 to 0.86, and the framework reveals that informal question phrasing does not degrade output quality.
By Emma Granqvist, Roc\'io Mercado, Samuel Genheden
The paper introduces a multi‑modal traffic sign detection framework that fuses camera and LiDAR data using an Intensity‑Aware Deformable Fusion module to align retro‑reflective LiDAR cues with visual features. It also presents a dual motion‑model tracker to handle non‑linear perspective changes and a semantic attribute classification pipeline that estimates occlusion, readability, sign embeddedness, and road relevance. Evaluated on a dataset covering more than 60 countries and 2,500 hours of driving, the system achieves an Object Miss Ratio of 0.49% across 221,068 sequences, indicating strong global generalization for autonomous driving.
By Meda Lazar, Sourab Sridhar, Shashwata Gupta, Alexandra Tripcea, Varun Ravi, Senthil Yogamani
MAESTRO is a large language model agent that automates the full screening pipeline for metal‑organic frameworks (MOFs). It parses extensive MOF literature, links publications to crystal structures, curates a computation‑ready database, and then applies a progressively more expensive computational strategy to identify promising candidates. The identified materials for wet flue gas separation come from unrelated studies, demonstrating the agent’s ability to uncover high‑performance materials across domains.
By Chen Yuntong, Huang Ju, Liu Yu, Zhao Dan, Sun Mingqi, Ju Chentian, Liu Yanbing, Huang Lijiang, Zhao Guobin
TH-GNN is a heterogeneous temporal graph neural network designed to detect shilling attacks generated by large language model (LLM) agents. It combines a two‑layer Heterogeneous Graph Transformer with per‑type and per‑relation attention, learnable sinusoidal temporal encodings, cross‑modal attention that fuses user embeddings with frozen RoBERTa representations of reviews and item descriptions, and a GRU that models log inter‑arrival times. Across five attack families and four benchmark datasets, TH‑GNN achieves a grand‑mean F1 score of 0.870, surpassing the best text‑only baseline on Agent4SR attacks by 10.9 percentage points and 11.5 percentage points at the lowest injection rate.
By Shivam Swarup, Divya Prakash Shrivastava, Rakesh Thakur
The paper introduces SUSVIBES, a benchmark of 186 real‑world software engineering tasks where human programmers have committed vulnerable code. It evaluates 12 popular coding‑agent settings on these tasks and finds that all agents perform poorly in terms of security, with only 11.8% of solutions from SWE‑Agent with Claude 4 Sonnet being secure despite 57% being functionally correct. Attempts to mitigate security issues by adding vulnerability hints to the prompts do not improve results.
By Songwen Zhao, Danqing Wang, Kexun Zhang, Jiaxuan Luo, Zhuo Li, Lei Li
The study examines how adding a security-requirements section to prompts affects web applications generated by a large language model. Six distinct applications were produced twice—once with a baseline prompt and once with a security-aware prompt—yielding 12 programs. Analysis of these programs revealed 75 confirmed security findings, with the security-aware variants showing fewer issues (24 vs. 51) and no Critical or High severity problems.
By Darko Andro\v{c}ec
arXiv:2608.21247v1 Announce Type: new
Abstract: Token compression has become a key technique for reducing the inference cost of large foundation models, with approaches such as token pruning and KV-c...
By Zhuoyuan Li, Rui Zhao, Jin Wang, Hanwei Zhu, Cong Zhang, Giuseppe Valenzise, Weisi Lin, Kin-Man Lam
Tree-of-Experience (ToE) is a hierarchical experience-management framework designed for large language model agents that aligns stored experiences with the agents’ reasoning hierarchy. By organizing experiences into a shared tree of analytical perspectives and reasoning paths, ToE calibrates reliability through environmental outcomes, enabling systematic updating, cross-task transfer, and efficient retrieval. Experiments on Game of 24 and FinEvolveBench demonstrate that ToE yields significant performance gains—31.4% accuracy improvement on Game of 24 and a 41.24% average improvement in tsIC on FinEvolveBench—outperforming both experience-free baselines and conventional experience-management methods.
By Zihao Deng, Yining Zhu, Leiming Wang, Junbo Wang, Jingfei Lu
AudioWorldSim is an open‑source platform that generates realistic binaural audio datasets for training and evaluating audio‑based machine learning models, especially world models. It extends Meta’s SoundSpaces 2.0 by automating random agent navigation and correcting continuous sound composition. The project is publicly available on GitHub to support reproducibility in research.
By Luis Vitor Zerkowski, Luiz Velho
LingShu is a large-scale, symptom‑centric knowledge graph that bridges Traditional Chinese Medicine (TCM) and modern biomedicine. It contains 17.33 million entity records and 39.47 million relation records, combining 17.19 million semantic triples with 22.29 million contextualized quadruples to encode conditional medical associations. The graph integrates data from electronic medical records, TCM texts, biomedical ontologies, and curated knowledge bases, and is supported by a web platform offering visualization, reasoning, and evidence‑grounded question answering.
By Rui Hua, Zixin Shu, Kai Chang, Dengying Yan, Jianan Xia, Hui Zhu, Shujie Song, Shurui Yang, Tongxin Wang, Yue Yin, Yu Wei, Lijuan Pei, Yunhui Hu, Hao Xu, Mingzhong Xiao, Xiaodong Li, Haibin Yu, Runshun Zhang, Wenjia Wang, Baoyan Liu, Xuezhong Zhou
EditPPT is a multi‑agent framework that turns slide editing into a constrained tool‑selection task, using the native PowerPoint COM interface to perform localized shape‑level operations. By separating validation across modalities, its dual‑modal validators assess both instruction fidelity and visual quality, achieving high execution and accuracy rates even on long decks. The authors also introduce DeckEdit‑Bench, a benchmark of 28 human‑authored decks with 582 slides and 183 editing prompts across varying deck lengths.
By Jiheon Kim, Kyudan Jung, Jaegul Choo
BridgeGuard is an air‑gapped, agentic Retrieval‑Augmented Generation system designed to automate compliance verification for the Federal Highway Administration’s bridge inspection requirements. It combines vector search over the FHWA Recording and Coding Guide with structured SQL queries against National Bridge Inventory data, guided by a stateful multi‑step ReAct planning loop that runs locally on commodity edge hardware. The system achieves near‑perfect classification and citation accuracy on large bridge inventories while operating without external network access, and ablation studies show that both vector search and the agentic loop are essential for correct compliance reasoning.
By Viraj Nishesh Darji, Hemaliben Rakeshkumar Darji
AsmEvo is an agentic assembly-level optimizer that targets compiled AMDGPU code objects, reconstructing a reassemblable representation and applying low-level edits guided by a long-horizon agent. It rebuilds ABI-preserving optimized objects and verifies functional equivalence through differential testing against the original binary. Experiments show significant speedups—up to 1.35× geometric mean on MI308X and 1.18× on MI300X—while maintaining correctness.
By Ji Liu, Puyuan Yang, Rongzhang Zheng, Fan Wang, Jinglin Wang, Muhammad A. Awad, Mortis Huang, Andy Chang, Zekai Li, Zeping Li, Zihao An, Yue Liu, Yuchen Yang, Jianghui Wang, Chushi Chen, Ziqiong Liu, Fuwei Yang, Dong Li, Wen Heng Chung, Shengcai Liu, Emad Barsoum
AEGIS is a policy enforcement component designed to prevent resource abuse in the Model Context Protocol (MCP), an open‑source JSON‑RPC protocol that allows large language models to interact with external systems via tools. By leveraging large language models to analyze, categorize, and normalize diverse tool invocations across text, images, video, and location modalities, AEGIS provides administrators with fine‑grained safeguards that can be enforced through the Open Policy Agent and ContextForge AI Gateway. This approach enables detection and mitigation of abusive behaviors—such as excessively large search requests or long video calls—while maintaining the flexibility of MCP‑based agent ecosystems.
By Shriti Priya, Teryl Taylor, Frederico Araujo
The paper introduces OneModel, a paradigm that consolidates complex business workflows and SOPs into a single model’s parameters, replacing traditional modular pipelines. By using Continual Pre‑Training and logic‑compilation SFT, the authors transform fragmented rules into unified attention‑based reasoning. In a global financial service deployment, OneModel cuts end‑to‑end latency by over 50 % and raises the Intelligent Resolution Rate from 64.3 % to 83.3 %.
By Chang Liu, Chaoyang Ning, Dayi Jiang, Enrui Gu, Fang Ran, Hongyan Xue, Huaqing Li, Hui Cai, Jia Liu, Jiang-Ming Yang, Jianshe Li, Jiawei Luo, Jin Zhou, Leshen Zhu, Lihui Chen, Liying Ma, Lyuxin Xue, Mengjian Ji, Ruijia Xu, Wei Ren, Wei Wu, Xiaoling Qu, Xiaoyun Feng, Xin Zhang, Xixie Zhou, Xuanwei Hu, Yan Chen, Yichao Wang, Yongqi Tong, Yu Liu, Yuhong Zhou, Zemin Sun, Zhenwen Xu, Zhiling Liu, Zifan Wang
The paper investigates how small lexical changes in prompts can cause large performance swings in large language models. Using a dataset of 132,000 prompt variants, the authors uncover a scaling law linking higher average task performance to lower variance and greater robustness. They identify domain-specific terminology and explicit action directives as key linguistic factors that stabilize prompts, and propose an automated Prompt-Refining Agent that reduces performance variance by 40.7% in code generation while maintaining or improving mean performance.
By Qipeng Xie, Zi Liang, Jiafei Wu, Yufei Chen, Weizheng Wang, Wenao Ma, Zhong Ming, Haiqin Yang, Kaishun Wu
Tree-of-Concerns is a multi‑agent framework that uses specialized skeptic personas to conduct parallel debate trees, each focusing on a specific category of potential limitations in scientific papers. The system employs structured, evidence‑grounded argumentation and a panel review mechanism to correct drift and miscalibration, ultimately extracting unstated limitations. Experiments on the ToC‑Bench benchmark show that the approach improves precision by 79% and coverage by 11% over leading baselines, providing reviewers with specific, evidence‑based concerns for systematic evaluation.
By Sahil Mishra, Niranjan Rajeev, Tanmoy Chakraborty