Sign InOpen Brain
arXivPaperNeeds Review

Rethinking On-Policy Distillation of Large Language Models II: One Training Example

On-policy distillation recovered most full-data gains from one query; diverse rollouts mattered more than dataset size, while slow student alignment remained the bottleneck.

arXiv · Sep 3, 2026
Open Source Open MarkdownOpen JSON
Source Summary

A single-query OPD run reached **71.5% state coverage** and recovered most of full-data OPD’s gain across tested tasks and model families. With **16 queries**, coverage rose to **98.9%** and matched full-data training.

Practical Implication

If you distill smaller models for agent workloads, optimize prompts for diverse visited states before collecting a large task dataset. The experiments suggest rollout coverage, not query count or topical content alone, is the useful data signal.

Agent-Ready Context
A single-query OPD run reached **71.5% state coverage** and recovered most of full-data OPD’s gain across tested tasks and model families. With **16 queries**, coverage rose to **98.9%** and matched full-data training.

If you distill smaller models for agent workloads, optimize prompts for diverse visited states before collecting a large task dataset. The experiments suggest rollout coverage, not query count or topical content alone, is the useful data signal.

This does not make training cheap: student-teacher alignment still took **hundreds of steps**, even on fixed states. The claims concern OPD experiments, so deployment quality and transfer to a specific coding-agent workload still need validation.
Connected Context · Feed7 Judgment

This sharply refines data-efficiency guidance for OPD: a prompt set can be tiny if its rollouts traverse diverse states, so query count and topical breadth are poor proxies for useful supervision. It does not remove the optimization cost or establish coding-agent transfer, making state coverage a prompt-selection criterion rather than evidence that adaptation is broadly cheap.

Data Quality Is the Compute Multiplier — Ari Morcos, DatologyAITurns the general emphasis on information-rich data into a measurable OPD criterion: visited-state coverage rather than dataset size alone.OctoLong: Mid-Training On Cross-Repository Code Contexts Enhances Long-Context ModelingBoth favor structurally diverse training experience, but OctoLong constructs linked code contexts whereas OPD obtains diversity through teacher rollouts from carefully chosen prompts.Post-Training Language Models for Gold-Medal Performance in Coding CompetitionsThe coding result makes specialized post-training consequential, while this work suggests such adaptation datasets should be designed around behavioral state coverage; coding-workload transfer still requires direct validation.Consolidating RLVR Capabilities Across Domains: A Deep Dive into Fusion ParadigmsThe unchanged solution coverage in RLVR fusion contrasts with OPD’s strong dependence on state coverage, reinforcing that average task gains do not by themselves show broader reasoning exploration.
Context Map
model#reasoning#model-selection
Uncertainty
This does not make training cheap: student-teacher alignment still took **hundreds of steps**, even on fixed states. The claims concern OPD experiments, so deployment quality and transfer to a specific coding-agent workload still need validation.