Announcement
MiRL targets recurring mistakes in terminal-based AI agents
MiRL’s researchers say training starts where recurring errors occur, rewarding runs only if they pass tests and avoid the error.
TLDR
MiRL’s researchers propose identifying recurring mistakes made by AI agents that work through a command shell, then starting reinforcement-learning runs where those mistakes occur. They report that eight misbehaviors accounted for 52.1% of wrong steps on Qwen3.5-4B. MiRL scored 51.8% versus TMax-4B’s 47.4% on Terminal-Bench Lite while using 35% of its rollouts, they say.
Combined views
3.3K
3 Sources, first seen ago
