On-policy self-distillation (OPSD) is a training scheme in which a student model learns from trajectories it generates itself, supervised by a reference-conditioned version of the same base model. A new paper on arXiv asks a basic question about this setup: does every trajectory actually provide distinct supervision that deserves to be distilled?
The paper, titled "Not Every Rollout Needs Distillation: Correctness-Filtered On-Policy Self-Distillation," argues that the answer may be no. The proposed method is a correctness-filtered variant of OPSD, which selects which rollouts to distill based on correctness rather than treating all generated trajectories as equally valuable. The abstract itself stops at posing the question, so the specific results and filtering criteria are not available from the abstract alone.
Because the source is a single preprint, there are no independent results to compare against. The significance lies in the challenge to a common assumption: if some rollouts are redundant or unhelpful, filtering them could make self-distillation more efficient or more effective.