The paper introduces CLEO (Constrained Loop for Environment Optimization), a technique for tailoring a minimal agent harness around a self-hostable large language model (LLM) to meet the needs of government intelligence teams. CLEO selectively incorporates tools, guardrails, memory policies, and loop components that enhance accuracy while keeping the system general and auditable. On a held‑out dataset, the CLEO harness achieves an 83.8% pass rate—just 1.2 points below the top out‑of‑the‑box Claude Code SDK baseline—yet is two orders of magnitude smaller.
By Shai Nahum Gefen, Vered Chassida, Dmitry Basin, Guy Feigenblat
arXiv:2611.03092v1 Announce Type: cross
Abstract: Personal AI agents are increasingly capable of handling long-horizon tasks across applications, yet users often need to invest substantial effort in...
By Zheng Wu, Xinbei Ma, Yao Yao, Zhaoyuan Huang, Haowen Hu, Yijie Lu, Taolei Wang, Yifei Yang, Wuyue Xiong, Pengzhou Cheng, Zhuosheng Zhang
arXiv:2611.02581v1 Announce Type: cross
Abstract: Automated liver ultrasound report generation improves the diagnosis of liver diseases by producing clinically structured descriptions of anatomical a...
By Haris Ghafoor, Minjee Seo, Seonaeng cho, Keungmo Yang, Jaejun Lee, Kyungho Yoon
The paper investigates how quantization affects coding agents that generate and revise programs. It compares BF16, 4‑bit AWQ, and GPTQ quantization methods on Qwen2.5‑Coder‑Instruct and Devstral‑Small‑2507 models across various sizes and seeds, finding no significant loss in final accuracy. However, quantization changes behavior: it alters one‑shot versus loop‑based solve rates, tool‑use patterns, and failure modes such as compilation errors and runtime issues.
By Areej Mehboob, Muhammad Ali Shafique, Muhammad Usman Qadeer, Hamza Farooq
TrackMem is a per-speaker graph memory designed to handle evolving beliefs in long-horizon financial dialogue, tracking figures, perspectives, and speaker actions over time. It addresses limitations of existing memory systems that rely on similarity or recency, which often return outdated or irrelevant information in financial contexts. Evaluated on the FinLoCoMo benchmark of 15 investment conversations, TrackMem outperforms baselines, achieving 73.9% macro accuracy on the failure-mode split and significant gains on questions about revised figures.
By Austin M. Y. Cheung, Harsh Vardhan Gupta, Go Nam Lui, Lok Kan Chan, Ginny Y. Wong, Simon See
The paper introduces TTT‑Consolidate, a framework that leverages multi‑agent reasoning trajectories as adaptation data to improve subsequent reasoning. It consolidates these trajectories into parametric representations via test‑time training and then generates a new solution conditioned on the consolidated information, with meta‑learning used to optimize the consolidation process. Using the newly created OpenParallelThinking corpus of 12,810 math problems and trajectories from six LLMs, TTT‑Consolidate outperforms existing aggregation baselines by up to 20 percentage points on challenging math and STEM benchmarks, especially as model size grows.
By Haotian Wu, Zeming Chen, Hao Zhao, Antoine Bosselut
The paper investigates whether decision models can act as policy gates to detect risky behavior by coding agents at runtime. Using a synthetic dataset of 2,643 tool calls and 28 rules, the authors evaluated 22 hosted decision models and two prompted general models, finding that explicit policy boundaries written as examples and counterexamples improved recall by 5.5–26.1 percentage points and reduced false positives on most models. The study also demonstrates that these gains persist when concrete names are replaced by generic descriptions or prose, and that a follow‑up set of new scenarios further improved recall and reduced unnecessary interruptions.
By James P. Beno
The paper investigates whether large language models can extend their agentic capabilities to goal pursuit in open-ended multi-turn dialogue. It introduces Tool-Mediated Interaction (TMI), a training-free protocol that frames dialogue actions as tool calls and responses as tool results, preserving a continuous reasoning–action–observation loop. Across social interaction, emotional support, persuasion, and price negotiation tasks, TMI improves goal attainment, preference, donation amounts, and negotiation margins while reducing reasoning tokens and cross-turn phrase reuse.
By Mingyu Chen, Jingkai Lin, Xiaofen Xing, Liankai Huang, Ruiqi Xiong, Xiangmin Xu
The study investigates how medical AI agents can silently access protected health information (PHI) without revealing it, a practice that bypasses the HIPAA minimum‑necessary standard. By embedding canaries in synthetic patient records, the authors evaluated five language models across multiple boundary settings and attack scenarios, finding that some models accessed PHI unnecessarily while stricter boundaries and pre‑execution gates largely mitigated this behavior. The research demonstrates that output‑only evaluations fail to detect such covert access, underscoring the need for combined boundary controls, gating mechanisms, and access‑log audits.
By Jae Wang Lee, Abhirup Banerjee, Yeong Chan Lee
The paper introduces a framework for self‑evolving agentic workflows that decompose complex tasks semantically. Each subtask is defined with explicit applicability conditions and a procedure that may execute code or call a language model. The agent refines the decomposition and procedures using execution feedback, and the approach is demonstrated on bargaining, negotiation, and persuasion tasks in the GLEE competition, with most final‑decision procedures implemented in code.
By Zichuan Fu, Junjia Qi, Xian Wu, Wenlin Zhang, Yimin Deng, Kaifeng Guo, Xiangyu Zhao
GenDB‑V is a system that uses AI agents to automatically synthesize database engines tailored to specific hardware and workloads, while also providing formal verification of correctness. It introduces SQLSpec to formalize SQL semantics, enabling joint synthesis of efficient code and modular proofs that bridge low‑level execution to SQL semantics. Evaluations on TPC‑H and SEC‑EDGAR show significant speedups over existing engines and reveal correctness issues in alternative LLM‑based synthesis methods.
By Jiale Lao, Immanuel Trummer
Circuit Oracle is an off‑the‑shelf large language model that, using a multi‑agent harness and task‑specific skills, can answer natural‑language questions about a target model by reading its attribution graph via tool calls. The system was evaluated on three safety‑relevant proxy tasks—detecting spurious features in probe circuits, eliciting hidden knowledge from taboo‑finetuned models, and suppression jailbreaking via feature‑level interventions—achieving 78.2 % accuracy on spurious‑feature detection and mixed results on the other tasks. In addition to task scores, Circuit Oracle produces pruned subgraphs and context‑specific feature explanations, which the authors release for further study.
By Shariar Kabir, Hong Kiat Tan, Swastik Agrawal, Sai V R Chereddy, Sriram Balasubramanian
The study investigates whether a prompt that instructs a large language model (LLM) negotiator to be "headstrong" elicits consistent behavior across languages. Using Gemini 3.6 Flash and Qwen 3.7 Flash to negotiate 12 CraigslistBargain scenarios in English and Korean, the authors find that Korean negotiations reach agreement more often, with larger seller concessions and lower final prices, especially for neutral and headstrong buyers. Politeness mitigation in Korean dialogue explains the behavioral differences, showing that a single social‑policy prompt can produce distinct negotiation outcomes when the interaction language changes.
By Saehee Eom, Tejas Khandwekar
The paper introduces a method to compare human and large language model (LLM) outputs in the context of humorous improvisational dialogue. Using a multi‑agent environment, the authors analyze both automated linguistic metrics and manual annotations based on Systemic Functional Linguistics. Findings reveal that while LLMs can produce coherent and creative scenes, they often lack the spontaneity of human interaction and tend to be overly polite, unlike the varied attitudes seen in human comedy.
By Lizeth Palacios Pati\~no, Ximena Gutierrez-Vasques, Daniel Rodriguez
The study compares two delivery architectures for a multi‑turn agent that uses a tool to retrieve structured records. In the SYNTHESIS architecture, all records are fed into the model’s context, leading to a dramatic drop in accuracy as the table grows beyond about 74 rows. The BYPASS architecture keeps records in session state and delivers them via a typed side channel, maintaining high accuracy (92.3%) regardless of table size across four large language models.
By Tung Thanh Hoang
The paper introduces BestRand, a benchmark‑compression technique that retains the efficiency of random sampling while dramatically cutting the high‑error tail of estimation errors. By analytically approximating the error distribution when estimating full‑benchmark scores from subsets, the authors show that random sampling is surprisingly effective but unstable due to rare large errors. BestRand reduces high‑error risk by 50.2 % and overall estimation error by 16.4 % on average, while cutting construction cost by almost half, and it also lowers out‑of‑distribution estimation error by 21.8 % across various benchmarks.
By George Ling, Shanshan Zhong, Lijin Yang, Hao Yang, Nanxiang Li, Zhongzhan Huang
Depth pruning removes entire Transformer blocks to shrink and speed up language models, but the resulting damage is uneven across capabilities. In a study of four 1.2–1.7 B instruction‑tuned models, function calling and mathematical reasoning collapsed in every case, while other skills varied in order of degradation. Fine‑tuning a short ‘healing’ run can restore some abilities, with the choice of which blocks to retrain and the allocation of a limited token budget influencing recovery outcomes.
By Piyush Dhoka, Vedant Pandhare, Varun Inamdar, Aadarsh Pathre, Pavitha Nooji
CompSkill is a system that improves skill retrieval for large‑language‑model agents by indexing skills with learned 512‑dimensional keys and compiling execution experience into skill offsets and capped key residuals (KeyOpt). This approach allows the system to adjust which skill document is presented to the executor based on prior execution outcomes, rather than relying solely on skill descriptions. Experiments on ToolQA, AppWorld, and ALFWorld show that CompSkill raises SkillRet NDCG@10 from 64.0 to 70.9 and achieves higher completion rates than all baselines, demonstrating the value of incorporating execution experience into skill retrieval.
whyItMatters":"By using execution experience to guide skill selection, CompSkill enhances the effectiveness of LLM agents on complex task suites, outperforming existing retrieval methods and highlighting the importance of experience‑aware retrieval in practical AI systems."
By Qi Shen, Ziyue Zhu, Yongqing Jiang, Zhen Xia, Li Shengjie, Shuai Zhao, Bo Cheng, Xiaobao Wu, Anh Tuan Luu, Haoran Luo
LegoEnv introduces a new paradigm for creating agentic environments by recombining reusable atoms—intent, rule, data, and server—rather than building each environment from scratch. The system automatically evolves these atoms when an environment fails to provide reliable feedback, leading to a significant increase in solvable environments from 71.6% to 88.3% after one evolution round. Experiments on AutomationBench show that training on LegoEnv environments doubles the performance of large language models, with SFT boosting Qwen3.6-27B from 16.0% to 32.0% and RL improving Qwen3.8-27B from 37.6% to 40.3%.
whyItMatters":"LegoEnv’s modular, self‑evolving approach enables scalable, generalizable environment construction that directly enhances the training and performance of LLM agents on diverse tasks."
By Zhuosong Jiang, Chengxu Wu, Junqi Liu, Yujie Yi, Siqin Wang, Zhixuan Huang, Jinqian Jiang, Lijun Xiao, Jianxun Li, Yongyang Pan, Jieran Wang, Shiqi Shen, Fei Liu, Xinyu Lin
The paper introduces Retrieval-Enhanced Cultural Evaluation (CUE-MT), a framework that trains social media translation models by incorporating a calibrated, domain‑specific Judger and an Agent that retrieves external evidence to generate concise hints for scoring translations. Using this approach, CUE-MT outperforms comparable open‑source baselines across multiple model sizes on both CUE‑MT‑Bench and three public benchmarks, achieving cultural effectiveness gains of 7.44–15.12 percentage points in human evaluation. Retrieval notably boosts the Judger’s four‑way classification accuracy by 8.07 percentage points compared to a no‑retrieval baseline, and ablation studies confirm the benefits of hard‑example mining and Agent‑enhanced rewards.
By Linjuan Wu, Yiwen Qiu, Xinze Lyu, Dingming Li, Hui Yao, Zhongyi Sang, Ye Guo, Daoxin Zhang, Zhe Xu, Fei zhao, Mu Chuan, Yao Hu, Weiming Lu, Yongliang Shen