Less is More: Early Stopping Rollout for On-Policy Distillation

2026-05-26

AI Summary

This work addresses the issue of “off-policy teacher degradation” in policy distillation, where early deviations in student-generated trajectories cause the teacher policy to revert to pretraining-style completion behavior. To mitigate this, the authors propose Early Stopping Rollout (ESR), which trains exclusively on the initial segment of student-generated tokens. The study uncovers underlying mechanisms—“cascading alignment” and “submodality commitment”—demonstrating that truncated rollouts consistently enhance performance in ways not fully captured by KL divergence or entropy signals. ESR employs a position-based token truncation strategy and achieves significant improvements over full-rollout distillation across diverse model architectures, scales, and tasks. Notably, it enhances training stability and GPU efficiency while, in certain settings, even surpassing the teacher model’s performance.

Abstract

On-policy distillation has recently emerged as a promising alternative to standard sequence-level imitation, training a student by scoring its own rollouts with a teacher model. However, we observe ``Off-policy Teacher Decay'' problem in this paradigm: for the later tokens, with student's earlier trajectory as context that is off-policy to the teacher, the teacher's ability to produce a corrective score would decay, and may fall back to token-completion behavior learned in the pre-training stage. We empirically verify this problem, and we propose Early Stopping Rollout (ESR) to fix it: a simple yet effective distillation strategy that simply restricts the rollout generation to the first response tokens. We show that ESR both surpasses the full rollout OPD performance across model size, family, tasks and training regime, and exhibit much higher GPU efficiency and training stability, especially under cross model family scenarios. We further investigate the mechanism behind this surprising performance and discovered "Cascading Alignment" and "Sub-mode Commitment" effect of ESR that may explain why it works effectively and even sometimes exceeding the teacher model performance. Besides, we show that this position-based token selection strategy cannot be fully explainable by KL divergence and entropy signals.

Problem

• On-policy Distillation
• Off-policy Teacher Decay
• Sequence-level Imitation
• Teacher-Student Alignment
• Rollout Generation

Innovation

• Early Stopping Rollout
• On-Policy Distillation
• Off-policy Teacher Decay
• Cascading Alignment
• Sub-mode Commitment