Reinforcement Learning via Self-Distillation

TL;DR

This paper introduces Self-Distillation Policy Optimization (SDPO), a reinforcement learning method that uses rich textual feedback to improve learning efficiency and accuracy in verifiable domains, surpassing traditional scalar reward methods.

Contribution

The paper presents SDPO, a novel reinforcement learning approach that leverages self-distillation from rich feedback without external teachers, enhancing sample efficiency and performance.

Findings

SDPO outperforms strong RLVR baselines in scientific reasoning, tool use, and programming tasks.

SDPO effectively uses implicit feedback from successful rollouts to improve learning.

SDPO accelerates discovery in binary-reward tasks with fewer attempts.

Abstract

Large language models are increasingly post-trained with reinforcement learning in verifiable domains such as code and math. Yet, current methods for reinforcement learning with verifiable rewards (RLVR) learn only from a scalar outcome reward per attempt, creating a severe credit-assignment bottleneck. Many verifiable environments actually provide rich textual feedback, such as runtime errors or judge evaluations, that explain why an attempt failed. We formalize this setting as reinforcement learning with rich feedback and introduce Self-Distillation Policy Optimization (SDPO), which converts tokenized feedback into a dense learning signal without any external teacher or explicit reward model. SDPO treats the current model conditioned on feedback as a self-teacher and distills its feedback-informed next-token predictions back into the policy. In this way, SDPO leverages the model's ability to retrospectively identify its own mistakes in-context. Across scientific reasoning, tool use, and competitive programming on LiveCodeBench v6, SDPO improves sample efficiency and final accuracy over strong RLVR baselines. Notably, SDPO also outperforms baselines in standard RLVR environments that only return scalar feedback by using successful rollouts as implicit feedback for failed attempts. Finally, applying SDPO to individual questions at test time accelerates discovery on difficult binary-reward tasks, achieving the same discovery probability as best-of-k sampling or multi-turn conversations with 3x fewer attempts.