RetireOPD: Self-Retiring On-Policy Distillation for Agentic Reinforcement Learning
RetireOPD trains multi-turn agents via self on-policy distillation with adaptive teacher retirement, lifting Qwen2.5 ALFWorld success by up to 18.8%.
RetireOPD first trains a decoupled, skill-conditioned teacher with environment rewards, then trains a skill-free student jointly with reinforcement learning and on-policy distillation. Its Adaptive Retirement mechanism drops the teacher once their discrepancy stops shrinking and the student reaches a target fraction of the teacher's success rate. Across Qwen2.5 models from 1.5B to 7B, RetireOPD improves ALFWorld success by 14.1-18.8% and WebShop accuracy by 11.8-19.0% over RL baselines, surpassing its own teacher in every setting.
- Dense token-level supervision from a self-teacher combined with RL
- Adaptive Retirement drops the teacher when discrepancy stops shrinking
- Tested on Qwen2.5 1.5B-7B across ALFWorld and WebShop tasks
- Student surpasses its own skill-conditioned teacher in all settings
Full article172 words · extracted from arxiv.org · click to collapse
Multi-turn agents trained with reinforcement learning (RL) receive a single scalar reward per trajectory, which motivates self on-policy distillation (OPD) to supply dense token-level supervision from a self-teacher with privileged task skills, letting a skill-free student internalize them. This recipe, however, is undermined by two findings in agentic tasks: privileged information alone does not always make a teacher reliable, and the benefit of teacher supervision is stage-dependent. We therefore propose RetireOPD (Self-Retiring On-Policy Distillation), which first optimizes a decoupled, skill-conditioned teacher with environment rewards and then trains a skill-free student jointly with RL and OPD. Rather than following a predefined distillation schedule, RetireOPD adopts Adaptive Retirement: the student drops the teacher on its own once their discrepancy stops shrinking and it reaches a target fraction of the teacher's success rate, after which training proceeds with RL alone. Across Qwen2.5 models from 1.5B to 7B, RetireOPD improves ALFWorld success rate over RL baseline by 14.1% to 18.8% and WebShop accuracy by 11.8% to 19.0%, and surpasses its own skill-conditioned teacher in every setting.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.20784