ReTerm trains terminal agents to recover from failures, +14.6 on Terminal-Bench
Wei Xie · hf · 2026-10-14
Researchers from Fudan, CUHK and SJTU present ReTerm, a self-improving terminal agent that learns from its own failed tasks:
- Data: a custom pipeline yields ReTerm-10K, high-quality recovery demonstrations for SFT cold start
- Algorithm: ReTerm-GRPO forks from the agent's own failure states to generate independent self-recovery attempts, using task success as the RL signal
- Results: built on Qwen3.5-9B, ReTerm-9B reaches 34.8% Pass@1 on Terminal-Bench 2.0, +14.6 points over the base model, with consistently higher recovery rates than standard GRPO
- Code, data and model are open-sourced
More from coding & agent
- Codex autocompaction gets stuck: one steer and 10 million tokens later it's still parroting stale state — Aryvyo · 2026-10-14
- Agentic engineering vs AI engineering: Hugo Bowne-Anderson's clean split — Al_Grigor · 2026-10-14
- PostHog to devtools: stop asking 'how to add AI,' ask 'what's the agent-first version' — edwin · 2026-10-14
- Devin SWE-2 jumps to #3 on attention-to-detail coding test, tying Claude Sonnet 5.5 — silasalberti · 2026-10-14
- MCP hits ~500M monthly downloads but stays free — the money goes to key-management security firms — FinanceYF5 · 2026-10-14
- One-shot flying dragon: Gemini 4 Argon with Blender MCP and Tripo 3D builds a pure .blend scene — sidahuj · 2026-10-14