[2601.20802] Reinforcement Learning via Self-Distillation

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.

Authors:Jonas Hübotter, Frederike Lübeck, Lejs Behric, Anton Baumann, Marco Bagatella, Daniel Marta, Ido Hakimi, Idan Shenfeld, Thomas Kleine Buening, Carlos Guestrin, Andreas Krause View PDF HTML (experimental) 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,…

saved by

• Chris Shi

related reading

• Self-Distillation Enables Continual Learningarxiv.org
• Why Does Reinforcement Learning Generalize? A Feature-Level Mechanistic Study of Post-Training in Large Language Modelsarxiv.org
• On-Policy Distillation - Thinking Machines Labthinkingmachines.ai
• State of RL for reasoning LLMs | A. Weersaweers.de
• RLHF: Reinforcement Learning from Human Feedbackhuyenchip.com
• [2601.18734] Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Modelsarxiv.org
• Reinforcement Learning With Verifiable Rewards: How Data and Verifiers Shape RLVRsnorkel.ai
• Reward Hacking in Reinforcement Learning | Lil'Loglilianweng.github.io
• Pedagogical RL: Teaching Models to Teach Themselves from Privileged Information - Noah Ziemsnoahziems.com
• Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Modelsarxiv.org
• SFT, RL, and On-Policy Distillation Through a Distributional Lens | whnrehiew.github.io
• The State of Reinforcement Learning for LLM Reasoningmagazine.sebastianraschka.com