MiRL: Error-Targeted RL Matches Full Agent Training With 35% of Rollouts
mohitban47 · x · 2026-10-28
Researchers introduce MiRL (Misbehavior-driven RL) for terminal agents: instead of scaling up RL, it profiles recurring "generalizable misbehaviors," then starts rollouts at the steps where they occur and rewards only rollouts that pass tests while avoiding them. On Qwen3.5-4B, 8 dominant misbehaviors accounted for 52.1% of wrong steps; using <200 envs from TMax-15K and 35% of TMax's rollouts, MiRL matches the full TMax-4B recipe at a fraction of the compute.
More from coding & agent
- Complexity cheat sheet: pick your algorithm by input constraints — cneuralnetwork · 2026-10-28
- SpaceXSI sales team builds shared Grok Bots for deals, saving hundreds of hours a month — elonmusk · 2026-10-28
- FastAPI Cloud unveiled: deploy a FastAPI app with a single command — jxnlco · 2026-10-28
- Scaling coding agents: one developer's path from 5-agent teams to 10,000 parallels — Al_Grigor · 2026-10-28
- Claude's answer to the most dangerous daily AI habit: handing over permissions — alex_verem · 2026-10-28
- Parallel launches Decisions API: web-grounded yes/no answers with confidence scores at $0.001/request — heyshrutimishra · 2026-10-28