Ask to Distill: Learning to Synthesize Data for On-Policy Distillation

The paper studies on-policy distillation (OPD), which can succeed with surprisingly few training questions. It treats questions as carriers of distributional knowledge transfer and asks how to assess whether a newly generated question is valuable for distillation, and how to learn to generate better questions without access to a correct answer. The supplied text is truncated and does not state the method's results.

Key takeaways

• On-policy distillation can succeed with surprisingly few training questions.
• The paper frames questions as carriers of distributional knowledge transfer.
• The authors pose the problem of scoring whether a generated question is valuable for distillation.
• The approach aims to learn question generation without access to correct answers.

Why it matters

Reducing the number of training questions needed for distillation could lower data-creation costs. Generating useful questions without verified answers may make distillation practical in settings with little labeled data.

Read with care

The supplied text is truncated, so the method's details and empirical results are not available; source claims not independently checked.