ReTerm trains terminal agents to recover from their own errors, +14.6 pts on Terminal-Bench 2.0
jiqizhixin · x · 2026-11-22
Researchers from Fudan University, CUHK MMLab and Shanghai Jiao Tong University introduce ReTerm, targeting an overlooked bottleneck: terminal agents recovering from their own mistaken commands or faulty implementations.
- Data: failure states are collected and useful recovery continuations filtered into ReTerm-10K, a 10k high-quality recovery trajectory dataset for supervised fine-tuning
- RL: ReTerm-GRPO forks multiple recovery attempts from the agent's own failed states and ranks them by eventual task completion, placing the model inside situations created by its previous errors
- Results: built on Qwen3.5-9B, ReTerm-9B hits 34.8% Pass-at-one on Terminal-Bench 2.0, 14.6 percentage points above its base model
More from coding & agent
- 9.5-hour local test: Inkling 2 Mini on a 3060 scores 94% across 162 Claude Code tasks — SummarizedAnu · 2026-11-22
- Running a Super Agent via WeChat to Manage All Your Issues Daily — jiayuan_jy · 2026-11-22
- FitMetrics runs Grok Build + 27 Grok Bots, merging 1,800 PRs in one month — XFreeze · 2026-11-22
- OpenAI ships 'Sign in with ChatGPT': local apps run GPT-6.x on your plan quota — alexcovo_eth · 2026-11-22
- The top programming languages of 2027: YAML, JSON, Markdown, and Bash — kmeanskaran · 2026-11-22
- Adversarial review makes AI agents flip-flop on every pass, user finds — EconomicsEvery4493 · 2026-11-22