ReTerm-9B

ReTerm-9B starts from Qwen3.5-9B and combines supervised training on recovery demonstrations with ReTerm-GRPO reinforcement learning. The method branches into multiple attempts from states left by the agent’s own failed terminal actions, then uses task-verifier outcomes as learning signals.

The Apache-2.0 model weights were uploaded on October 6. The English preprint was submitted October 13, and the GitHub training repository’s initial-release commit is dated October 14. ReTerm-10K provides 9,712 recovery trajectories under ODC-BY. Reported task-completion scores are the authors’ evaluations and depend on the agent harness and inference budget.