On-Policy Self-Distillation with Sampled Demonstrations Reduces Output Diversity
Identifies a hidden cost of self-distillation for practitioners using LLMs, showing that improved average performance can come at the expense of diversity and robustness.
On-policy self-distillation with sampled demonstrations improves pass@1 accuracy but reduces rollout diversity, flattening pass@k curves. The method amplifies existing probability gaps, concentrating mass on dominant modes, leading to lower functional and semantic diversity and failure on out-of-distribution tasks.
On-policy self-distillation achieves strong pass@1 accuracy by using a single model as both teacher and student, with the teacher conditioned on a correct demonstration to provide dense token-level feedback. We show that this could come at a hidden cost: rollout diversity decreases and pass@k curves flatten (i.e., generating more rollouts fails to improve accuracy). We trace this to compounding biases in the design of self-distillation with sampled demonstrations. The teacher scores each student rollout while conditioned on a sampled correct rollout, channeling its feedback through the model's own biases. We theoretically analyze the optimal self-distillation policy and show that it tilts the base distribution by a pointwise conditional mutual information score between the student's rollout and the correct rollout used as context. Unlike the ideal optimal on-policy reinforcement learning (RL), which preserves probability ratios among equally correct rollouts, self-distillation can amplify existing probability gaps, concentrating mass on already-dominant modes. On a controlled graph path-finding task and science question-answering benchmarks, self-distilled models match or exceed RL on average performance but exhibit substantially lower functional and semantic diversity, failing on out-of-distribution settings that require diverse strategies.