arXiv 2026 Preprint
Rethinking On-Policy Distillation of Large Language Models II: One Training Example
Investigating why on-policy distillation can recover much of full-data training’s gain from a single query.
01 / The question
How much training data does on-policy distillation actually need?
02 / The approach
One-Shot On-Policy Distillation
The work compares one-shot and full-data distillation across models and tasks, measuring the states reached by student rollouts and the pace of teacher–student alignment. It connects gains from query diversity to state coverage and examines why absorbing that supervision still takes many updates.
03 / Citation
Cite this work
Select the citation to copy it manually.
04 / Versions & records
Publication record
READ HERE · 29 PAGES
Full paper
Loading document…
Scroll through every page. Select text to copy, or use search to find a passage.