MiRL: error-targeted RL matches full terminal-agent training with <200 envs and 35% rollouts

mohitban47 · x · 2026-10-28

Researchers introduce MiRL (Misbehavior-driven RL), an error-targeted post-training method for terminal agents: it first profiles recurring "generalizable misbehaviors" across environments, then starts rollouts at the steps where they occur and rewards rollouts only if tests pass and the misbehaviors are avoided.

Key results and claims:

On a Qwen3.5-4B-based terminal agent, 8 dominant misbehaviors account for 52.1% of wrong steps;
With fewer than 200 environments in TMax-15K and 35% of TMax's rollouts, MiRL matches the full TMax-4B recipe;
The authors argue the paradigm democratizes agentic post-training: instead of expensive, unpredictable thousands-of-rollouts runs, you pick which errors to fix and watch them disappear during training — a legible, accessible path to capability gains.

Original post →

MiRL: error-targeted RL matches full terminal-agent training with <200 envs and 35% rollouts

More from coding & agent