RetireOPD: Self-Retiring On-Policy Distillation for Agentic Reinforcement Learning

TL;DR

RetireOPD method improves RL model success rate by 18.8% in ALFWorld using adaptive retirement strategy.

cs.CL 🔴 Advanced 2026-09-18 16 views
Yan Yu Zhengxi Lu Yizhou Liu Yichen Pan Aozhe Wang Qipeng Chen Hua Yang Wenqi Zhang Weiming Lu Qianglong Chen Yongliang Shen
reinforcement learning policy distillation adaptive retirement multi-turn dialogue agent training

Key Findings

Methodology

RetireOPD first optimizes a skill-conditioned teacher model using environment rewards, then trains a skill-free student with RL and OPD. It employs adaptive retirement, ceasing teacher supervision when student-teacher discrepancy stops shrinking, continuing with RL alone.

Key Results

  • In ALFWorld, RetireOPD improves success rate by 14.1% to 18.8% over RL baseline, and WebShop accuracy by 11.8% to 19.0%.
  • RetireOPD surpasses its skill-conditioned teacher in all settings.
  • RetireOPD performs well across different scales of Qwen2.5 models, proving its effectiveness in multi-task environments.

Significance

RetireOPD offers a novel policy distillation approach in reinforcement learning, addressing stage-dependent teacher supervision issues with adaptive retirement strategy. It shows outstanding performance in multi-task environments, holding significant academic and practical value.

Technical Contribution

RetireOPD resolves the unreliability of teacher models in traditional policy distillation by optimizing skill-conditioned teachers and introducing adaptive retirement, providing new theoretical guarantees and engineering possibilities.

Novelty

RetireOPD is the first to propose adaptive retirement, differing from traditional fixed-schedule methods, allowing dynamic adjustment of teacher supervision timing, enhancing model flexibility and effectiveness.

Limitations

  • RetireOPD may require longer training times for optimal results in some tasks.
  • The parameter settings for adaptive retirement strategy can significantly impact performance across different tasks.

Future Work

Future research could explore RetireOPD's application in more complex tasks and optimize parameter settings for adaptive retirement strategy to further enhance model performance.

AI Executive Summary

In multi-turn dialogue tasks, traditional reinforcement learning methods often face sparse reward issues, leading to unsupervised intermediate decisions. The RetireOPD method introduces an adaptive retirement strategy, combined with optimized skill-conditioned teacher models, to address stage-dependent teacher supervision issues.

RetireOPD first optimizes a skill-conditioned teacher model using environment rewards, then trains a skill-free student with RL and OPD. The adaptive retirement strategy allows the student to cease teacher supervision when the discrepancy with the teacher stops shrinking, continuing training with RL alone. This method shows excellent performance in tasks like ALFWorld and WebShop, with significant improvements in success and accuracy rates.

RetireOPD holds significant academic value and offers new possibilities for practical applications. Future research could explore its application in more complex tasks and optimize parameter settings for adaptive retirement strategy to further enhance model performance.

Deep Analysis

Background

Reinforcement learning in multi-turn dialogue tasks faces sparse reward issues. Traditional methods like GRPO and OPD provide partial solutions but still face challenges in teacher model reliability and supervision timing.

Core Problem

In multi-turn dialogue tasks, effectively utilizing teacher models for policy distillation while avoiding long-term constraints on student models is a critical and challenging issue.

Innovation

RetireOPD addresses teacher model unreliability and supervision timing issues through adaptive retirement strategy and optimized skill-conditioned teacher models, enhancing model flexibility and effectiveness.

Methodology

  • �� Optimize skill-conditioned teacher model using environment rewards.
  • �� Train skill-free student with RL and OPD.
  • �� Employ adaptive retirement strategy, ceasing teacher supervision when student-teacher discrepancy stops shrinking, continuing with RL alone.

Experiments

Experiments conducted on ALFWorld and WebShop using different scales of Qwen2.5 models, comparing RetireOPD with traditional RL and hybrid baseline methods.

Results

RetireOPD improves success rate by 14.1% to 18.8% over RL baseline in ALFWorld, and WebShop accuracy by 11.8% to 19.0%. It surpasses its skill-conditioned teacher in all settings.

Applications

RetireOPD can be used in multi-turn dialogue task agent training, suitable for scenarios requiring dynamic adjustment of strategy supervision.

Limitations & Outlook

RetireOPD may require longer training times in some tasks, and parameter settings for adaptive retirement strategy can significantly impact performance across different tasks.

Plain Language Accessible to non-experts

Imagine a school where a teacher guides students in learning various skills. Initially, the teacher closely monitors the students' progress, but as students gradually master the knowledge, the teacher reduces supervision until they are fully independent. RetireOPD is like such a teacher, helping students gain more guidance in the early stages of learning but allowing them to explore independently as their abilities improve.

ELI14 Explained like you're 14

Imagine playing a complex game where an expert teaches you how to play. As you become more skilled, the expert gradually reduces guidance until you can play well on your own. RetireOPD is like this system, helping AI get guidance in the early learning stages and then letting them perform independently.

Glossary

RetireOPD (Adaptive Retirement Strategy)

A method for dynamically adjusting the timing of teacher supervision in policy distillation.

Used to enhance flexibility in multi-turn dialogue tasks.

GRPO (Group Relative Policy Optimization)

A reinforcement learning algorithm optimized through group-relative advantages.

Serves as one of the foundational RL methods for RetireOPD.

OPD (On-Policy Distillation)

A policy distillation method providing dense supervision from teacher models.

Used to enhance student model learning effectiveness.

ALFWorld

A text-based environment for testing agents in multi-turn dialogue tasks.

One of the experimental benchmarks for RetireOPD.

WebShop

An environment simulating online shopping scenarios for testing agent decision-making.

One of the experimental benchmarks for RetireOPD.

Open Questions Unanswered questions from this research

  • 1 How to apply RetireOPD in more complex tasks to further validate its effectiveness.
  • 2 How parameter settings for adaptive retirement strategy affect performance across different tasks.

Applications

Immediate Applications

Multi-turn Dialogue Systems

Apply RetireOPD in multi-turn dialogue tasks to enhance agent learning efficiency and flexibility.

Long-term Vision

Autonomous Agent Learning

Achieve autonomous learning and decision-making capabilities in agents through RetireOPD, reducing reliance on external supervision.

Abstract

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.

cs.CL cs.AI