The paper introduces MERIT, a benchmark that evaluates the marginal benefit of long‑term memory for tool‑using large language model agents while explicitly accounting for cost. MERIT provides episodic tool‑use tasks across three domains, verifies dependence on earlier‑episode facts, and measures memory operations in tokens and dollars. Experiments on GPT‑4.1‑mini, Claude Haiku 4.5, and Claude Sonnet 5 show that memory can significantly improve task success, but its utility varies widely across models and memory implementations, and full replay is rarely cost‑effective.
By Shweta Mishra, Shashank Mishra
The paper introduces TEAM-Design, a rule that assigns two replay probabilities to each task—one for a human-only replay and one for an agent-only replay—based on how difficult it is to predict the missing baseline outcome and the cost of replay. It addresses the challenge of deciding whether to keep a human-AI workflow or replace it with a single actor when only one outcome can be observed after deployment. The authors prove that TEAM-Design solves the budgeted design problem and controls error rates, and demonstrate its effectiveness on clinical and coding benchmarks, noting it excels when one comparison is clearly harder than the other.
By Hamed Khosravi, Xiaoming Huo
RubricRefine is a training‑free pre‑execution refinement method that generates task‑specific rubrics from tool documentation, scores candidate code against explicit contract checks, and iteratively repairs failures before execution. It achieves an average score of 0.86 across seven models on M3ToolEval without any execution attempts, outperforming prior inference‑time baselines while incurring lower latency. The approach shows consistent performance on single‑step API‑Bank tasks and maintains an advantage in multi‑turn settings on AppWorld, with its effectiveness tied to the quality of the supplied documentation.
By Will LeVine, Brendan Evers, Sam Saltwick, Abhay Venkatesh
The paper introduces an attack that reconstructs text generated by locally hosted large language models by monitoring CPU cache activity during detokenization. It uses Flush+Reload on shared tokenizer code to time decoding, then Prime+Probe to capture token‑dependent cache traces, followed by a clustering‑and‑language‑model pipeline to recover the output text. The method is evaluated across various datasets, hardware, inference frameworks, and model families, successfully retrieving semantically accurate outputs from real‑world local LLM deployments, including agentic systems.
By Roy Weiss, Benyamin Konstantinov, Eitam Sheetrit, Tomer Simon, Yisroel Mirsky
The paper introduces AMA, a framework that uses multiple agents—Constructor, Retriever, Judge, and Refresher—to manage memory for large language model agents. AMA’s hierarchical memory design dynamically adjusts retrieval granularity to match task complexity, while the Judge and Refresher ensure relevance, consistency, and timely updates. Experiments on long-context benchmarks show AMA outperforms existing baselines and cuts token usage by about 80% compared to full-context approaches.
By Weiquan Huang, Zixuan Wang, Hehai Lin, Sudong Wang, Bo Xu, Qian Li, Beier Zhu, Linyi Yang, Chengwei Qin
The paper proposes a method for identifying the current traffic situation of a vehicle using Bird's‑Eye‑View images. It combines object detection with semantic segmentation and a situation identification neural network built on the Cam2BEV projective transformation. The approach was validated in the CARLA simulator and on Cityscapes and nuScenes datasets, demonstrating that it can prioritize relevant sensor data for efficient distribution in connected vehicles.
By Falk Dettinger, Akshay Narla, Michael Weyrich
The paper presents a framework for causal attribution in agentic AI systems, outlining estimators and conditions where they fail. It distinguishes between marginal total effects and common‑random‑number total effects, introduces a natural direct effect under pinned downstreams, and derives a coupling method to keep direct effects estimable. The authors also propose a traceability specification to meet upcoming regulatory requirements for high‑risk AI systems.
By Ajay Pravin Mahale (Hochschule Trier)
APPSim-Bench is a new benchmark for mobile GUI agents that uses controllable simulated apps to balance realism and reproducibility. It includes 557 tasks across 17 popular Chinese and English apps, with a coding-agent-assisted and human-verified workflow that ensures deterministic evaluation. Evaluation of 19 agents shows that autonomous mobile execution is still far from perfect, with the best model completing only 50.27% of tasks and many failures in longer workflows and numerical reasoning.
By Jintian Feng, Long Chen, Xiao Yu, Jiayi Dai, Chenglong Liu, Haoru Wang, Zizhen Xue, Yuxuan Shi, Ziyang Wang, Yichen Gong
The paper introduces SAEScientist-Bench, a benchmark that tests whether AI agents can autonomously conduct mechanistic interpretability research using Sparse Autoencoders (SAEs). Agents are tasked with designing contrastive probes and navigating a large feature dictionary in Gemma-2-9B-IT to identify optimal features for a target concept, with performance measured against expert-curated references on activation rank, concept selectivity, and causal steering. Results show that while frontier agents can discover features and outperform controls, they still lag behind expert baselines, especially in causal steering, highlighting both the potential and current limitations of closed-loop autonomous AI research.
By Yuqiao Tan, Shizhu He, Jun Zhao, Kang Liu
TamilEOT is a new dataset and two audio‑only models for detecting semantic end‑of‑turn in Tamil telephone speech, comprising 18,485 labeled turn boundaries from 116 real conversations. The models, fine‑tuned from Smart Turn v3, achieve 83.71% and 86.13% accuracy on a held‑out set, with ROC‑AUC rising from 0.751 to 0.921, and run in under 150 ms on a laptop CPU. The paper also reports the cost of building the dataset, the accuracy of rule‑derived labels versus human agreement, and the impact of encoder capacity on performance.
whyItMatters":"The study provides the first publicly available Tamil end‑of‑turn detection resource, enabling more accurate voice agents for South Indian languages and demonstrating efficient, low‑latency models that outperform zero‑shot baselines."
By Santhoshkumar V
The paper critiques current evaluations of efficiency methods for large language model–based multi‑agent systems, arguing that reported gains are often inflated by method‑specific prompts and starting topologies. It introduces a controlled, MAS‑demanding diagnostic benchmark that standardizes the backbone model, agent registry, and runtime, and systematically varies topology, scale, depth, and tool use. The authors find that many claimed efficiency improvements are setup‑dependent, sometimes stemming from structural collapse or random pruning rather than genuine, robust gains.
By Jiamu Zhang, Lingxi Zhang, Pengjun Lu, Qiyue Zhang, Yu-Neng Chuang, Zhengchen Li, Shuai Xu, Vipin Chaudhary, Hanjie Chen
arXiv:2609. 07344v1 Announce Type: cross Abstract: Large language model (LLM) based agents are increasingly applied to cybersecurity tasks such as vulnerability discovery and automated penetration testing.
By Weizhe Wang, Yitong Zhang, Yao Zhang, Xiaoqiang Di, Zhigang Li, Bin Wu, Guangquan Xu
The paper introduces a three-tier persona vector for user simulation in evaluating LLM agents, comprising 23 dimensions across demographics, behavioral traits, and emotional states, plus a query-complexity overlay. It demonstrates that these nuanced personas generate diverse, scenario-reactive conversations, leading to significant variations in agent goal achievement and compliance across different contexts. The model’s design allows for reproducible, auditable user behavior patterns without relying on learned covariance matrices.
By Rahul Khedar, Eshita, Sneha Teja Sree Reddy Thondapu, Mayank Malhotra, Arup Kumar Das, Jitesh Chandra Mishra, Arun Menon, Avinash Karn, Mouli V
SocialRL is a multi-turn reinforcement learning framework that refines large language models’ social intelligence. It uses PPO to propagate delayed outcome rewards across turns, enabling long‑horizon planning, and introduces six process reward dimensions—such as goal advancement and relational attunement—to capture the goal‑relationship trade‑off. A reward model provides fine‑grained scoring and a stage‑aware weight schedule prioritizes relationship building early, goal pursuit mid‑way, and balanced closure later, yielding an average 9.2 percentage‑point improvement in goal achievement across multiple social‑dialogue benchmarks.
By Jianing Wang, Xintao Wang, Aili Chen, Jie Shi, Hongcheng Guo, Jun Gao, Wenxuan Zhao, Chengkun Lang, Yuanli Guo, Yanghua Xiao
arXiv:2504.00285v2 Announce Type: replace
Abstract: Large Language Models (LLMs) are effective at deceiving when prompted to do so. Models that demonstrate better performance on reasoning tasks are a...
By Samuel M. Taylor, Benjamin K. Bergen
arXiv:2609.08399v1 Announce Type: new
Abstract: Universal machine-learning interatomic potentials (u-MLIPs) aim to generalize across diverse configurations. Benchmarks enable reproducible evaluation...
By Ryuhei Okuno, Nontawat Charoenphakdee, Kaoru Hisama, Yuta Tsuboi
arXiv:2606.03371v4 Announce Type: replace
Abstract: Reliable proactive agents must choose an action and judge whether current evidence is sufficient to act. We study retail service from sparse third-...
By Honghui Zhang, Anna Min, Chenmeinian Guo, Yujia Zhang, Yichen Yu, Zezhou Zhang, Guanyu Liu, Yongming Qin, Chongguo Song, Mengyue Yang, Lei Yu, Tianyu Shi
arXiv:2609.06059v1 Announce Type: new
Abstract: As large language models evolve from question-answering systems into general-purpose agents, evaluation must move beyond static answer correctness to a...
By Yu Liu, Zhilin Liu, Zhiwei Yang, Shaojie Zhang, Zheyuan Deng, Tingwei Huang, Zhenbo Luo, Lei Jiang, Yanbing Liu, Pei Fu
arXiv:2609.09881v1 Announce Type: new
Abstract: Semantic segmentation for autonomous driving requires reliable detection of vulnerable road users (VRUs) despite heavy class imbalance. We introduce CL...
By Toomas Tahves, Mauro Bellone, Raivo Sell
arXiv:2609.10540v1 Announce Type: new
Abstract: Recent video world models generate increasingly realistic and interactive visual experiences, yet lack reliable mechanisms for maintaining persistent w...
By Zheng-Hui Huang, Guixu Lin, Jiacheng Lin, Yi-Chuan Huang, Ruihan Yu, Muyao Niu, Siqi Yang, Yu-Lun Liu, Yung-Yu Chuang, Kaipeng Zhang, Zhixiang Wang