ReTerm: Recovery Training for Self-Improving Terminal Agents

Wei Xie; Kaituo Feng; Shiming Fu; Yong Liu; Xiangyu Yue; Shuangjia Zheng

cs.AIOct 13, 2026

ReTerm: Recovery Training for Self-Improving Terminal Agents

Authors: Wei Xie, Kaituo Feng, Shiming Fu, Yong Liu, Xiangyu Yue, Shuangjia Zheng

Organizations: Fudan University · MMLab, The Chinese University of Hong Kong · Shanghai Jiao Tong University

Abstract

Terminal agents have shown strong abilities in complex coding tasks. However, they often struggle to recover from errors during long-running execution (e.g., command errors, faulty implementations), eventually leading to task failures. To this end, we present ReTerm, a self-improving terminal agent trained by learning how to recover from its own failed terminal tasks. To support its training, we construct a customized data pipeline and produce ReTerm-10K, a high-quality dataset that provides effective recovery demonstrations for SFT cold start. Furthermore, we propose ReTerm-GRPO, an online RL algorithm that forks from the agent's own failure states to generate multiple independent self-recovery attempts. By learning from whether these attempts successfully complete the task, the agent self-improves through its own recovery experience. Built on Qwen3.5-9B, our ReTerm-9B achieves 34.8% Pass@1 on Terminal-Bench 2.0, improving over the base model by 14.6 points. Further analysis shows that our method achieves consistently higher recovery rates across diverse error categories compared with standard GRPO training. All code, data and model are released at https://github.com/twothsnd/ReTerm.

Figures & tables

Figure 1: An inference example revealing the difference in recovery behavior between models trained by GRPO and our proposed ReTerm-GRPO. With ReTerm-GRPO, the model formulates a stronger recovery strategy.

Figure 2: Overview of our data curation pipeline. Starting from an executable task, we collect various failed attempts induced by the base model. Then, a stronger model is asked to recover from the resulting failure state based on the limited information. For the generated recovery continuations, we finally apply four-step quality filtering and construct a high-quality dataset.

Figure 3: ReTerm-10K dataset statistics. (a) Domain distribution of recovery trajectories. The nine domains have relatively even shares (8.4–14.2%), and file operations account for the largest share. (b) Distribution of output token count in recovery continuations. The 4–8k token range is the most common (37.1%), and 69.6% of recoveries contain 2–8k output tokens. (c) Distribution of assistant turns in the recovery continuation. 9–16 turns is the most common range (38.0%), and 87.8% of recoveries finish within 32 turns.

Figure 4: An illustration of our proposed ReTerm-GRPO.

Closed-Source Models
Model	Model Size	OpenThoughts- TBLite	Terminal-Bench 2.0	SWE-bench Verified
GPT-5 mini	–	–	38.2	–
Gemini 3 Flash	–	–	47.6	78.0
Claude Sonnet 4.5	–	–	51.0	77.2
Claude Opus 4.6	–	–	65.4	80.8
Gemini 3.1 Pro	–	–	68.5	80.6

Table 1: Performance comparison on OpenThoughts-TBLite, Terminal-Bench 2.0, and SWE-bench Verified. All results are Pass@1 (%).

Model	Ours	Terminus-2	mini-SWE-agent	OpenHands
Qwen3.5-9B	20.2	22.5	19.1	15.7
ReTerm-9B	34.8	32.6	25.8	28.1
$\Delta$	+14.6	+10.1	+6.7	+12.4

Table 2: Generalization performance across agent harnesses. All results are Pass@1 (%).

Model	AIME 2024	AIME 2025	Aider Polyglot	SpreadsheetBench-V
Qwen3.5-9B	76.7	73.3	17.3	37.8
ReTerm-9B	86.7	93.3	20.4	48.8
$\Delta$	+10.0	+20.0	+3.1	+11.0

Table 3: Generalization performance across task domains. All results are Pass@1 (%).

Method	Terminal-Bench 2.0
Base	20.2
SFT	25.8
GRPO	28.0
ReTerm-GRPO	34.8

Table 4: Ablation study on Terminal-Bench 2.0. All results are Pass@1 (%).

Figure 5: Recovery rates after encountering different error types.

Figure 6: Mean reward curves during ReTerm-GRPO training.

LLM Agent Training

Large language model agent training teaches models to plan, use tools, and learn from multi-step interaction with environments.

See all 582 papers

Tmax: A simple recipe for terminal agents
Jun 22, 2026Hamish Ivison, Junjie Oscar Yin, Rulin Shao +3Synthetic Data Generation Terminal Agents
α Allen Institute for AI · ω University of Washington
T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks
Sep 10, 2026Junyao Yang, Yucheng Shi, Zhongzhi Li +4RL for Tool Use Long-Horizon Agent Evaluation
Tencent Hy Foundation Model · National University of Singapore · University of Georgia +2
Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite
Oct 2, 2026Zongxia Li, Yucheng Shi, Zhongzhi Li +7LLM Agent Trajectory Synthesis Self-Training for Language Models
Tencent HY LLM Frontier · University of Maryland, College Park · University of Georgia +4

Jun 22, 2026J/K move · Enter open · S save

Tmax: A simple recipe for terminal agents

Hamish Ivison, Junjie Oscar Yin, Rulin Shao, Teng Xiao+2

α Allen Institute for AI · ω University of Washington

Synthetic Data Generation Terminal Agents Synthetic Data Generation for LLM Agents LLM Agent Training RL for LLM Agents

Open paper AskPDF arXiv

Terminal-using agents have quickly become the most popular downstream application of language models (LMs). Despite their prevalence, relatively little academic work has examined RL-based training of these models, likely due to difficult benchmarks, a lack of data, and a lack of simple baseline recipes. We present Tmax, the strongest open RL recipe for terminal agents to date, bringing open data recipes closer to the frontier. While simple, our recipe achieves 27% on Terminal-Bench 2.0 with only 9B parameters, outperforming much larger models from prior work. Concretely, we generate data using a novel taxonomy, combining difficulty control, personas, and verifier diversification, which allows us to cheaply generate large amounts of terminal environments for RL and SFT training. We open-source our terminal dataset, which is over 2.5x larger than previously released terminal-agent datasets. We then train open-weight models using RL with our data, using a simple, outcome-only recipe. We release our data, models, and code as a strong baseline for future open academic work on terminal agents at https://github.com/hamishivi/tmax.

Abstract

LLM Agent Training

Explore similar work

Tmax: A simple recipe for terminal agents

T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks

Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite