8 misbehaviors drive 52% of agent errors: MiRL trains terminal agents with 35% of rollouts
mohitban47 · x · 2026-10-28
Researchers propose MiRL (Misbehavior-driven RL), a targeted RL method for terminal agents.
- Error analysis shows just 8 recurring misbehaviors account for 52.1% of wrong steps across terminal tasks
- MiRL profiles a policy to find generalizable misbehaviors, then starts RL rollouts at the steps where they occur, rewarding only outcomes that pass tests and avoid the misbehavior
- With only 189 training environments and 35% of TMax's rollout compute, MiRL beats TMax-4B on Terminal-Bench Lite and matches it on Terminal-Bench 2.1 (on Qwen3.5-4B)
Takeaway: learning where and what to fix can matter more than training on more data.
Related event: MiRL: Misbehavior-Driven RL Cuts Terminal Agent Errors by Up to 57%(3 posts)→
More from coding & agent
- Dumb chunk-and-rerank baseline beats top agentic memory systems in 6 of 7 benchmarks — JeremyCMorgan · 2026-10-28
- Seven changes cut Hermes agent running costs by up to 95% — tomcrawshaw01 · 2026-10-28
- Microsoft's MANDATE: 23 failure categories from 163k Reddit posts on AI agents — dair_ai · 2026-10-28
- LangSmith now tracks cache hit rates for your agents — LangChain · 2026-10-28
- MOSS-Image ecosystem roundup: ComfyUI 0.43, QwenVL layout node, new style LoRAs — optimisticalish · 2026-10-28
- Ex-PC repairman builds JinnexGuard with Claude Code to scan MCP servers before install — Complex-Pin-4616 · 2026-10-28