On-Policy Self-Distillation with Sampled Demonstrations Reduces Output Diversity
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...
Summary
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 bi
Why it matters
This is part of the steady stream of AI work that reshapes how researchers and builders think about what’s possible. The full details are in the original source below — worth reading directly rather than relying on a brief summary.
Read the original
The primary source has the full paper, announcement, or reporting:
→ https://arxiv.org/abs/2606.26091v1
Curated by Nizam.Wiki — a daily signal in the AI noise.