The paper introduces TerminalHorizon, an automated engine that synthesizes long‑horizon trajectories for terminal agents by reconstructing environments from real‑world professional work and progressively expanding tasks to require further exploration and improvement. Using this engine, the authors create 1.5K task environments and fine‑tune Qwen3.5‑27B, producing TerminalHorizon‑27B, which shows significant performance gains across 12 benchmarks in coding, agentic production tasks, and reasoning. Case studies demonstrate the model’s ability to experiment beyond initial solutions, refine methods based on feedback, and preserve effective improvements in final deliverables.
By Shun Zou, Ziyu Ma, Yi Zou, Yong Wang, Lin Chen, Zehui Chen, Guanhua Chen, Xiangxiang Chu, Feng Zhao
The paper introduces Strategy Accumulation and Guided Execution (SAGE), a two-stage framework that makes automated fine-tuning of large language models cumulative. In the first stage, a multi-agent pipeline uses Monte Carlo Tree Search to explore training strategies while a Distillation Agent records task-specific insights and cross-task confidence scores into a structured repository. In the second stage, SAGE retrieves relevant experience from this repository to guide training on new tasks, achieving a 12.4‑percentage‑point improvement over a baseline pipeline without accumulated experience on nine unseen tasks.
By Haoran Zhao, Wei Du, Dingwen Yang, Jixuan Huang, Junlin Shang, Lingyong Fang, Ya Guo, Tao Gui, Qi Zhang, Xuanjing Huang
arXiv:2601. 21754v3 Announce Type: replace Abstract: While Large Language Models (LLMs) excel in language-based agentic tasks, their applicability to unseen, nonlinguistic environments (e.
By Haoyu Wang, Guozheng Ma, Shugang Cui, Yilun Kong, Haotian Luo, Li Shen, Mengya Gao, Yichao Wu, Xiaogang Wang, Dacheng Tao
DAEDALUS is a method that builds reusable memory for large‑language‑model agents by having an explorer agent generate self‑created tasks and a solver agent attempt them. When the solver fails, a heuristic is extracted and only accepted after repeated successful use, then added to a memory bank for future test‑time use. Experiments on AppWorld, τ²‑bench, and AutomationBench show that DAEDALUS raises mean success rates by up to 15.9 points and pass⁵ by up to 2.2× compared to a no‑memory baseline, while also providing a cost‑effective alternative to training‑task or oracle‑verifier approaches.
By Antoine Edy, Max Conti, Victor Xing, Marc-Antoine Allard, Nawfal Benhamdane, Gautier Viaud
arXiv:2511. 20297v2 Announce Type: replace Abstract: Large Language Model (LLM)-based agents are increasingly capable of complex, multi-step tasks such as GUI automation, tool use, and data manipulation, yet they cannot learn from experience: each new session rediscovers solutions from scratch.
By Shashank Kirtania, Param Biyani, Priyanshu Gupta, Yasharth Bajpai, Roshni Iyer, Sumit Gulwani, Gustavo Soares
arXiv:2610.14224v1 Announce Type: new
Abstract: LLM agents' capabilities depend not only on their underlying models but also on their execution harnesses. Recursive self-improvement (RSI) offers a pr...
By Ruiyao Xu, Yanfei Chen, Zhongying CuiZhu, Bhavana Dalvi Mishra, Yifei Ming, Han Yu, Rujun Han, Chen-Yu Lee, Tomas Pfister
arXiv:2608. 15071v1 Announce Type: new Abstract: Learning from experience is critical for developing capable, self-improving large language model (LLM) agents.
By Tianxin Wei, Zhan Shi, Minhua Lin, Bing He, Zewen Liu, Yisi Sang, Yuanchen Bei, Xuying Ning, Jiaru Zou, Ting-Wei Li, Xiao Lin, Yanjun Zhao, Chi Wang, Benoit Dumoulin, Dakuo Wang, Jingrui He, Hanqing Lu
ReGE is a framework that transforms recorded interactive task experience into guidance for new attempts. Its implementation, ReGE‑Syn, synthesizes execution records and artifacts into concise, self‑contained instructions using a single coach call, without updating model parameters. Experiments on four benchmarks show that ReGE improves accuracy by roughly 9–10 percentage points for GPT and Claude, and significantly reduces guidance‑construction tokens while maintaining success rates.
By Jiarong Liang, Yubo Wang, Zijie Liu, Yuxuan Zhang, Ping Nie, Wenhu Chen
Agentick is a unified benchmark for sequential decision‑making agents that evaluates RL, LLM, VLM, hybrid, and human agents on 37 procedurally generated tasks across six capability categories, four difficulty levels, and five observation modalities via a single Gymnasium‑compatible interface. It includes a Coding API, oracle reference policies, pre‑built SFT datasets, a composable agent harness, and a live leaderboard. An evaluation of 27 configurations and over 90,000 episodes shows no single approach dominates, with GPT‑5 mini leading overall, PPO excelling in planning and multi‑agent tasks, and the reasoning harness boosting LLM performance by 3‑10×, while ASCII observations outperform natural language.
By Roger Creus Castanyer, Pablo Samuel Castro, Glen Berseth
MetaExp is a framework that trains language‑model agents to learn reusable experience across tasks by using verifiable cross‑task environment scaling. It introduces two designs—Shared‑Tool families that keep black‑box tools constant and Shared‑Workflow families that preserve latent procedures—allowing agents to acquire operational and procedural experience respectively. After fine‑tuning Qwen3.5‑122B‑A10B on thousands of trajectories from 1,000 task families, MetaExp achieves significant performance gains on benchmarks such as SpreadsheetBench, APEX‑Agents, and AutomationBench, with improvements persisting even when domain‑excluded training is applied.
By Renhao Li, Xinjie Shen, Yang Su, Jianhong Tu, Derek F. Wong, Dayiheng Liu, Min Yang
arXiv:2606. 01667v1 Announce Type: new Abstract: Test-time scaling has become a major way to improve large language model reasoning, but its orchestration has remained designer-engineered: a fixed sample budget, a fixed refinement loop, a fixed scoring rule, or a fixed search policy decides how compute is spent, leaving the model in charge of solving but not of orchestration.
By Peijia Qin, Qi Cao, Pengtao Xie
The paper presents the first dataset that captures how humans learn and use mechanisms in unfamiliar environments, using ARC‑AGI‑3 as a testbed. It includes 144 trajectories from 64 participants across 25 public games, each paired with time‑synchronized actions and think‑aloud audio, and 250 trajectories from large language model agents with and without harnesses. The dataset is annotated for mechanism triggers, action purposes, errors, and recovery, enabling a comparison of human and LLM cognition, highlighting differences in hypothesis testing, exploration, and use of qualitative representations.
By Shuzhi Liu, Archana Warrier, Dat Nguyen, Chengqi Zheng, Kevin Ellis, Yewen Pu