ReTerm: Recovery Training for Self-Improving Terminal Agents
The paper trains terminal agents specifically on recovering from their own failed execution states.
ReTerm builds a recovery dataset, ReTerm-10K, for supervised cold-start training. It also introduces ReTerm-GRPO, which forks from failure points and tests multiple independent recovery attempts. The reported ReTerm-9B model, built on Qwen3.5-9B, reaches 34.8% Pass@1 on Terminal-Bench 2.0, 14.6 points above the base model. The authors say code, data, and model are released. HF Daily Papers' note
ReTerm builds a recovery dataset, ReTerm-10K, for supervised cold-start training. It also introduces ReTerm-GRPO, which forks from failure points and tests multiple independent recovery attempts. The reported ReTerm-9B model, built on Qwen3.5-9B, reaches 34.8% Pass@1 on Terminal-Bench 2.0, 14.6 points above the base model. The authors say code, data, and model are released. HF Daily Papers' note
score 5