MiRL: error-targeted RL cuts terminal agent misbehaviors 43-57%
_akhaliq · x · 2026-10-28
Researchers (Runchu Tian, Mohit Bansal et al.) propose MiRL (Misbehavior-driven RL), bringing error-targeted learning to terminal agents.
- Method: profile a policy's errors to find generalizable "misbehaviors" — recurring wrong step patterns across environments — then run RL rollouts starting at those steps, rewarding only rollouts that pass tests and avoid the misbehavior
- Key finding: on Qwen3.5-4B, eight misbehaviors account for 52.1% of all wrong steps
- Results: trained policy repeats misbehaviors 43-57% less on unseen prefixes; beats TMax-4B on Terminal-Bench Lite (51.8% vs 47.4%), matches it on Terminal-Bench 2.1 (21.3% vs 20.6%), using only 35% of its rollouts
The paper notes existing error analyses are used for eval/debug rather than training, and prior error-targeted methods target single-turn tasks; MiRL shows the approach works on realistic long-horizon terminal tasks.
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