On-policy distillation study: 4 prompts match 3,840, but prompt utility depends on the teacher-student pair
Tencent-Hunyuan · hf · 2026-10-14
A Nanjing University–Tencent study systematically examines how prompt quantity, source, and selection shape transfer in on-policy distillation (OPD). Key findings: OPD is highly prompt-efficient—4 DAPO prompts match the math score of 3,840 DeepMath prompts—but prompt utility is relational, not intrinsic: swapping only the teacher can reverse the relative effectiveness of math vs. code prompts. Targeted selection doesn't consistently beat uniform random sampling, and filtering weak sources yields no consistent gain. Effective data choice depends on the teacher-student pair and target capability, with random sampling a competitive baseline. Code: github.com/wyy-1112/dissecting-opd
More from Research
- RUC and Kling AI propose EventLens: event-structure RL to fix video LLMs' temporal understanding — Guowen Zhang · 2026-10-14
- JSIR dataset drops: 267 jazz recordings with beat, chord and structure annotations — umpedronosapato · 2026-10-14
- There is no such thing as an AI drug, argues researcher Patrick Kidger — PatrickKidger · 2026-10-14
- More thinking, worse scores: refusals tank closed-model evals at max reasoning effort — _lewtun · 2026-10-14
- Integer multiplication repo becomes an agent harness benchmark — and collaboration loses — iskander · 2026-10-14
- Recurrent self-conditioning: framing recent diffusion methods as depth decoupled from time — alec_helbling · 2026-10-14