Step-Level On-Policy Distillation: Interpolating Between On-Policy Distillation and Supervised Fine-Tuning

2026-08-17

AI Summary

This study addresses the limitation of traditional on-policy distillation, which typically yields fragmented corrections without complete repair pathways. To overcome this, we propose a step-level on-policy distillation framework that unifies the strengths of supervised fine-tuning and on-policy learning. By employing trajectory-conditioned generation based on student states, our method enables precise, long-horizon supervision and alignment throughout the student's generation process. Experiments demonstrate that this framework significantly outperforms existing baselines in reasoning and agent tasks, achieving a 13.4 percentage point improvement in success rate on ALFWorld. These results effectively validate the superiority of step-level correction mechanisms in complex sequential decision-making scenarios.

Abstract

On-policy distillation (OPD) aligns a student model with a teacher's logit distribution on student-generated trajectories. This approach has achieved strong empirical gains and can often surpass conventional off-policy distillation with substantially less data. However, standard token-level OPD can provide only fragmented corrections along an erroneous student trajectory and cannot unfold a complete and correct repair path. Motivated by this limitation, we propose \emph{Step-Level On-Policy Distillation} (SOPD), which combines the long-horizon correction of supervised fine-tuning (SFT) with the on-policy advantage of OPD to provide step-level supervision over complete student-generated trajectories. We show that, at different limits of step length, SOPD reduces to SFT or approximates OPD. Compared with SFT, the teacher responses in SOPD are conditioned on student trajectories and therefore align more closely with student-visited states; compared with OPD, SOPD provides longer-horizon corrections rather than fragmented token-level guidance. Across both reasoning and agent tasks, SOPD substantially outperforms conventional SFT and OPD. For example, on ALFWorld, SOPD improves the average success rate by 13.4 points over Vanilla OPD. We hope this work offers a new perspective for future research on distillation methods.

Problem

• On-Policy Distillation
• Token-level correction
• Long-horizon correction
• Knowledge Distillation

Innovation

• Step-Level On-Policy Distillation
• On-Policy Distillation
• Supervised Fine-Tuning
• Trajectory Alignment
• Long-Horizon Correction