Problem
Research questions and friction points this paper is trying to address.
On-Policy Distillation
Multi-turn Agents
Error Compounding
State Distribution Shift
Model Compression
On-Policy Distillation with Curriculum Turn-level Guidance for Multi-turn Agents
🤖 AI Summary
This work addresses the high inference cost of large language model–based agents in multi-turn tasks and the failure of conventional online policy distillation (OPD) due to error accumulation from early student mistakes, which causes trajectory deviation and renders supervision ineffective. To mitigate this, the authors propose Guided-OPD, a novel algorithm featuring an episode-level curriculum guidance mechanism that dynamically blends teacher and student actions within each trajectory. The probability of teacher intervention is gradually reduced over training: strong initial guidance keeps student trajectories close to the teacher’s distribution, while later stages smoothly transition to a purely student-driven policy. Evaluated on ALFWorld, ScienceWorld, and WebShop using Qwen3-30B-A3B as the teacher, Guided-OPD yields a 21.1% average score improvement and a 25.5% higher success rate when distilling smaller models, with even greater gains observed for more compact student architectures.
📝 Abstract
Multi-turn agents that plan, invoke tools, and interact with environments offer a promising paradigm for solving complex tasks, yet their capabilities typically rely on very large models whose inference cost is prohibitive in practice.On-Policy Distillation (OPD) is a natural recipe for transferring such capabilities to smaller students, but we find that it suffers a characteristic failure mode in this setting: small student errors compound across turns and push the trajectory out of the teacher's familiar state distribution, so the teacher's supervision becomes least reliable precisely where the student needs it most.We propose Guided On-Policy Distillation (Guided-OPD), a simple yet effective algorithm that mixes teacher- and student-generated turns within each rollout and schedules the teacher's intervention probability along a curriculum that decays to zero.Strong guidance keeps early trajectories close to the teacher distribution and is then gradually withdrawn to recover the purely on-policy regime used at inference.On ALFWorld, ScienceWorld, and WebShop, distilling Qwen3 students from a Qwen3-30B-A3B teacher, Guided-OPD improves Score by 21.1\% and Success Rate by 25.5\% over vanilla OPD on average, with larger gains on smaller students.
Innovation
Methods, ideas, or system contributions that make the work stand out.
On-Policy Distillation
Curriculum Learning
Multi-turn Agents
Teacher-Student Guidance
Trajectory Distribution Alignment
🔎 Similar Papers
No similar papers found.
💼 Related Jobs
