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 narrows data-efficiency claims for distillation: topical dataset breadth may be unnecessary when a few prompts generate broad state coverage, but optimization remains costly because alignment still requires hundreds of steps. For agent adaptation, prompt selection should be judged by visited-state diversity, not query count alone.

Data Quality Is the Compute Multiplier — Ari Morcos, DatologyAIMakes the general data-curation principle operational for OPD by identifying rollout-state coverage as the relevant information-value signal.OctoLong: Mid-Training On Cross-Repository Code Contexts Enhances Long-Context ModelingBoth favor structurally informative training exposure over raw volume: dependency-linked contexts in mid-training and diverse visited states in distillation.Consolidating RLVR Capabilities Across Domains: A Deep Dive into Fusion ParadigmsAdds a useful metric contrast: OPD gains track visited-state coverage, while RLVR fusion did not broaden solution coverage despite similar average performance.Adaption Labs: Gradient-Free Continual Learning — Sara Hooker, AdaptionSupplies a concrete constraint for automated adaptation systems: discovering a tiny effective prompt set would not remove the substantial student-alignment compute.
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.