ReTerm-9B Hits 34.8% on Terminal-Bench 2.0 via Recovery Training

ReTerm-9B, a terminal agent from the Chinese University of Hong Kong's MMLab and Shanghai Jiao Tong University, scores 34.8% Pass@1 on Terminal-Bench 2.0 and 58.4% on SWE-bench Verified, improvements of 14.6 and 13.2 points over its Qwen3.5-9B base model. The paper on Hugging Face describes training the agent to recover from its own failures rather than only solving tasks from a clean environment.

The pipeline collects failed base-model attempts, restores the broken environment snapshot, and uses DeepSeek-V4-Pro as a teacher to generate recovery continuations, filtered down to roughly 10K trajectories for a supervised cold start. On top of that, ReTerm-GRPO 'forks from the agent's own failure states to generate multiple independent self-recovery attempts' and learns from which attempts pass the task verifier.

The abstract reports that the method 'achieves consistently higher recovery rates across diverse error categories compared with standard GRPO training.' Absolute levels stay low though: 'except for tool use error, the recovery rates of ReTerm-GRPO remain at or below 32% across all other categories.'

Training ran on 16 NVIDIA A100 (80GB) GPUs, with ReTerm-10K partitioned into 6K SFT trajectories and 4K RL instances. It joins Adobe-UCLA's Tev among the Terminal-Bench training-recipe papers we logged this week.