Ask-to-Distill: synthesizing questions for answer-free on-policy distillation
UIUC-CS · hf · 2026-10-22
A UIUC team released the paper "Ask to Distill: Learning to Synthesize Data for On-Policy Distillation" (Oct 20, 2026).
Key ideas:
- On-policy distillation (OPD) can succeed with surprisingly few training questions, suggesting questions are carriers of distributional knowledge transfer
- Core challenge: assessing whether a newly generated question is valuable for distillation — without access to correct answers
- Ask-to-Distill (A2D) treats synthetic questions as distribution elicitors and optimizes the teacher supervision they expose along student-generated trajectories
- A student-initialized synthesizer is optimized with a distributional surrogate for question value, based on teacher-student disagreement and reachability, requiring no answer labels
The method improves reasoning, coding, and agentic task performance; code, models and data are open-sourced on GitHub and HF.
More from Research
- S-Agent 2.0 uses a coding agent for spatial reasoning, +20.5 on ReVSI — liuziwei7 · 2026-10-22
- Open textbook explains when controller stability is topologically impossible — burkov · 2026-10-22
- OPAL: a reactive optimizer that watches your training and tunes it for you — vincesitzmann · 2026-10-22
- Toward inverse symbolic learning: back out physics equations from learned deep-net world models — bradneuberg · 2026-10-22
- Endless Powder streams an infinite ski mountain with a bounded 64 MiB working set — sterlingcrispin · 2026-10-22
- Retracting old papers: 'Is it still load-bearing?' beats 'Is it too old?' — iskander · 2026-10-22