RetireOPD: self-retiring on-policy distillation lifts agent RL success rates by up to 18.8%

Yan Yu · hf · 2026-09-18

RetireOPD addresses sparse rewards in multi-turn agent RL by first training a decoupled, skill-conditioned teacher with environment rewards, then training a skill-free student with joint RL and on-policy distillation. Instead of a fixed schedule, Adaptive Retirement lets the student drop the teacher once their discrepancy plateaus and it reaches a target fraction of the teacher's success rate. On Qwen2.5 models (1.5B–7B), it improves ALFWorld success by 14.1–18.8 points and WebShop accuracy by 11.8–19.0 points over RL baselines, surpassing its own teacher in every setting.

Original post →

More from Research

Research channel →