Weak-to-Strong Generalization via Direct On-Policy Distillation

2026-07-06

AI Summary

This work addresses the high computational cost of Reinforcement Learning with Verifiable Rewards (RLVR) on large language models, which has become a bottleneck in post-training scaling. The authors propose Direct On-Policy Distillation (Direct-OPD), a novel approach that leverages policy shifts induced by weak-model reinforcement learning as implicit, dense, and cross-scale transferable reward signals to directly guide the policy optimization of stronger student models—without requiring an explicit reward model or sparse-reward RL on the target model. By integrating policy distillation with implicit reward modeling, Direct-OPD improves Qwen3-1.7B’s performance on AIME 2024 from 48.3% to 62.4% in just four hours using eight A100 GPUs, significantly outperforming direct RL under equivalent training steps and enabling multi-stage composition of policy shifts.

Abstract

Reinforcement learning with verifiable rewards (RLVR) is a powerful recipe for improving language-model reasoning, but it is expensive to repeat on every new strong model because the target model must generate many rollouts during training. As models scale, post-training itself becomes a bottleneck. We study a weak-to-strong alternative: run RL on a smaller model where rollouts are cheaper, then reuse what that RL run learned to improve a stronger target model. Directly distilling the post-RL weak teacher is not enough, because the teacher's final policy mixes useful RL gains with the limitations of the smaller model. We propose Direct On-Policy Distillation (Direct-OPD), which transfers the teacher's RL-induced policy shift instead. Direct-OPD compares the post-RL teacher with its own pre-RL reference and treats their log-ratio as a dense implicit reward for the student. In plain terms, the checkpoint pair tells us which actions RL made the weak model more or less likely to take, and Direct-OPD applies that signal on the stronger student's own on-policy states. This directly reuses the weak model's RL supervision signal without training an explicit reward model or running sparse-reward RL on the target model. Empirically, Direct-OPD consistently leverages weaker teachers to improve stronger target models; notably, it boosts Qwen3-1.7B from 48.3% to 62.4% on AIME 2024 in just 4 hours on 8 A100 GPUs. It outperforms step-matched direct RL and enables the sequential composition of multiple policy shifts. Our results show that RL outcomes can be reused across model scales as implicit reward signals, not merely as final models to imitate.

Problem

• weak-to-strong generalization
• reinforcement learning
• policy distillation
• language model reasoning
• implicit reward

Innovation

• Direct On-Policy Distillation
• weak-to-strong generalization
• implicit reward
• policy shift
• reinforcement learning