ReTerm: Recovery Training for Self-Improving Terminal Agents
To this end, we present ReTerm, a self-improving terminal agent trained by learning how to recover from its own failed terminal tasks.
ProofPaper ↗
Key points
- Terminal agents have shown strong abilities in complex coding 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.
- 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.
Sources (2)
- [1]ReTerm: Recovery Training for Self-Improving Terminal AgentsHugging Face Daily Papers · Oct 13, 12:00 AM
To this end, we present ReTerm, a self-improving terminal agent trained by learning how to recover from its own failed terminal tasks.
Terminal agents have shown strong abilities in complex coding tasks.
- [2]ReTerm: Recovery Training for Self-Improving Terminal AgentsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 02:44 PM · same content
Extractive summary: sentences quoted from the sources.
Before this
- Oct 12, 2026[AINews] Meta Muse Wonder & Wonder Full: MSL’s Claude 5.5-tier Frontier Model
- Oct 12, 2026AUGUstus: Autonomous Driving VLA Model with Unified Generation and Understanding
- Oct 12, 2026TPPO: Tracking-Value PPO For Language Model Reasoning
- Oct 11, 2026From Zero to Leaderboard: Agent Evaluation — Wolfram Ravenwolf, Weights & Biases
- Oct 11, 2026Scalable Weak-to-Strong Verifier Hardening for Agent Environments
- Oct 10, 2026Predicting how model propensities generalize out of distribution
