ActionOPD: Action-Level On-Policy Distillation for Agents
October 13, 2026
ActionOPD optimizes agent distillation by aligning student models with teacher semantic actions rather than individual tokens. This approach maximizes expected teacher-action likelihood at states encountered by the student, mitigating state-distribution mismatch common in offline imitation.
HOW THIS AFFECTS YOU
●
builderYou can build smaller, more capable agents by using action-level supervision.
●
researcherYou can improve distillation efficiency by targeting semantic units instead of raw tokens.