ReTerm trains terminal agents to recover from their own failed tasks
An arXiv paper submitted on 13 October presents ReTerm, a self-improving terminal agent trained by learning how to recover from its own failed terminal tasks. The authors built a data pipeline producing ReTerm-10K, a dataset of recovery demonstrations for supervised fine-tuning cold start, and propose ReTerm-GRPO, an online RL algorithm that forks from the agent's own failure states to generate multiple independent self-recovery attempts and learns from whether they complete the task. Built on Qwen3.5-9B, ReTerm-9B reaches 34.8% Pass@1 on Terminal-Bench 2.0, which the authors say is 14.6 points above the base model, and they report consistently higher recovery rates across error categories than standard GRPO training; code, data and model are stated to be released.
The authors argue terminal agents often struggle to recover from command errors or faulty implementations during long-running execution, eventually failing tasks — the failure mode ReTerm is trained to address.
Computer Science > Artificial Intelligence
arXiv:2610.18484 (cs)
[Submitted on 13 Oct 2026]
Title:ReTerm: Recovery Training for Self-Improving Terminal Agents
Authors:Wei Xie, Kaituo Feng, Shiming Fu, Yong Liu, Xiangyu Yue, Shuangjia Zheng
View a PDF of the paper titled ReTerm: Recovery Training for Self-Improving Terminal Agents, by Wei Xie and 5 other authors
#ReTerm#ReTerm-9B#ReTerm-GRPO#ReTerm-10K#Qwen3.5-9B#Terminal-Bench 2.0