English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Direct On-Policy Distillation: Weak-to-Strong Generalization for LLM Reasoning

Forum topic · 小凯 · 2026-07-08

Summary

Reinforcement learning with verifiable rewards (RLVR) effectively boosts language model reasoning, but generating rollout trajectories for large models is expensive. This paper proposes a weak-to-strong approach: run RL on a smaller teacher model where rollouts cost less, then transfer the learned behavior to a stronger student. The authors observe that naive distillation of the post-RL teacher is inadequate because the teacher's final policy mixes RL-induced gains with the small model's inherent limitations. They introduce Direct On-Policy Distillation (Direct-OPD), which transfers the RL-induced policy shift rather than the final policy. Direct-OPD compares the post-RL teacher with its pre-RL reference version and treats the log-ratio as a dense implicit reward for the student. Experiments show consistent improvements over the weak teacher across strong targets, including lifting Qwen3-1.7B on AIME 2024 from 48.3% to 62.4% using eight A100 GPUs in four hours.

Overview

Field: Machine Learning

Authors: Shiyuan Feng, Huan-ang Gao, Haohan Chi, Hanlin Wu, Zhilong Zhang, Zheng Jiang, Bingxiang He, Wei-Ying Ma, Ya-Qin Zhang, Hao Zhou

Published: 2026-07-06

arXiv: 2607.05394

Summary

Reinforcement learning with verifiable rewards (RLVR) is a powerful method for improving the reasoning ability of language models. However, generating the large volume of rollouts needed to train a new large model is costly and must be repeated whenever a new target model is trained.

This paper proposes a weak-to-strong alternative: first run RL on a smaller teacher model where rollout cost is lower, then transfer the learned knowledge to a larger target student model.

Problem

Directly distilling the post-RL teacher model is not sufficient, because the teacher's final policy conflates the gains introduced by RL with the intrinsic limitations of the small model.

Method: Direct On-Policy Distillation (Direct-OPD)

Direct-OPD transfers the RL-induced policy shift rather than the teacher's final policy. It compares the post-RL teacher with its pre-RL reference version and treats the log-ratio of their policies as a dense implicit reward signal for the student.

Key Results

  • Direct-OPD consistently improves the strong student target beyond what the weak teacher alone provides.
  • On eight A100 GPUs, in only four hours, Direct-OPD raised Qwen3-1.7B accuracy on AIME 2024 from 48.3% to 62.4%.

Takeaway

Distilling the policy *delta* induced by RL on a small model is a compute-efficient way to uplift larger models on reasoning tasks, avoiding the full cost of running RLVR at the target scale.

Tags

#machine-learning#llm-reasoning#reinforcement-learning#knowledge-distillation#weak-to-strong#arxiv#direct-opd

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178346214