MiRL: RL from recurring mistakes matches full terminal-agent training with 35% of rollouts
mohitban47 · x · 2026-10-28
A new paper introduces MiRL (Misbehavior-driven RL) for efficiently training terminal agents to avoid misbehavior.
Method: profile a policy's errors to find mistakes recurring across many environments ("generalizable misbehaviors"), then train with RL starting rollouts at those failure steps, rewarding a rollout only if it passes tests and avoids the misbehavior.
Results:
- On Qwen3.5-4B, 8 dominant misbehaviors account for 52.1% of wrong steps and appear in 75.4% of environments;
- With <200 envs and 35% of TMax's rollout budget, MiRL matches the TMax-4B full recipe;
- It beats TMax-4B on Terminal-Bench Lite (51.8% vs 47.4%) and matches it on Terminal-Bench 2.1 (21.3%).
Related event: MiRL: Misbehavior-driven RL Reduces Terminal Agent Errors(2 posts)→
More from coding & agent
- Multimodal GraphRAG models missing and contradictory evidence for storm damage assessment — JeremyCMorgan · 2026-10-28
- Vibe coding in real life: 20 hours of personal-project work done in 2 — gleech · 2026-10-28
- LangChain open-sources Content Studio, a Slack design agent, and shares its harness design — BraceSproul · 2026-10-28
- "If AI can do it, humans shouldn't": team lets AI agent review PRs — Yuchenj_UW · 2026-10-28
- Sequoia, Citi and OpenAI launch No Limit: build a poker bot in 60 minutes, $100K grand prize — polynoamial · 2026-10-28
- Nyxis: a Lean framework with formally proven authorization safety for AI-written backends — udmrzn · 2026-10-28