The authors studied error-targeted learning for terminal agents, which solve long-horizon tasks by interacting with a command shell.
Error-targeted RL corrects terminal agent misbehaviors by training on failure steps
MiRL profiles an agent's errors to find recurring misbehaviors, starts RL rollouts at their occurrence, and achieves state-of-the-art performance on command-shell tasks with far fewer training environments.
Academic
Runchu Tian · Zaid Khan · Zun Wang · Justin Chih-Yao Chen · Mohit Bansal
University of North Carolina at Chapel Hill
Research Digest··3 min read
The authors introduce Misbehavior-driven RL (MiRL), a method that first profiles a terminal agent's errors to identify generalizable misbehaviors (recurring types of wrong steps) and then trains the agent with reinforcement learning starting from the steps where those misbehaviors occur.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.