Self-Distillation Zero: Self-Revision Turns Binary Rewards into Dense Supervision

Hugging Face Daily Papers Papers

Summary

Self-Distillation Zero (SD-Zero) is a novel training method that converts sparse binary rewards into dense token-level supervision through dual-role training where a model acts as both generator and reviser, achieving 10%+ improvements on math and code reasoning benchmarks with higher sample efficiency than RL approaches.

Current post-training methods in verifiable settings fall into two categories. Reinforcement learning (RLVR) relies on binary rewards, which are broadly applicable and powerful, but provide only sparse supervision during training. Distillation provides dense token-level supervision, typically obtained from an external teacher or using high-quality demonstrations. Collecting such supervision can be costly or unavailable. We propose Self-Distillation Zero (SD-Zero), a method that is substantially more training sample-efficient than RL and does not require an external teacher or high-quality demonstrations. SD-Zero trains a single model to play two roles: a Generator, which produces an initial response, and a Reviser, which conditions on that response and its binary reward to produce an improved response. We then perform on-policy self-distillation to distill the reviser into the generator, using the reviser's token distributions conditioned on the generator's response and its reward as supervision. In effect, SD-Zero trains the model to transform binary rewards into dense token-level self-supervision. On math and code reasoning benchmarks with Qwen3-4B-Instruct and Olmo-3-7B-Instruct, SD-Zero improves performance by at least 10% over the base models and outperforms strong baselines, including Rejection Fine-Tuning (RFT), GRPO, and Self-Distillation Fine-Tuning (SDFT), under the same question set and training sample budget. Extensive ablation studies show two novel characteristics of our proposed algorithm: (a) token-level self-localization, where the reviser can identify the key tokens that need to be revised in the generator's response based on reward, and (b) iterative self-evolution, where the improving ability to revise answers can be distilled back into generation performance with regular teacher synchronization.
Original Article
View Cached Full Text

Cached at: 04/20/26, 08:29 AM

Paper page - Self-Distillation Zero: Self-Revision Turns Binary Rewards into Dense Supervision

Source: https://huggingface.co/papers/2604.12002

Abstract

Self-Distillation Zero trains a model to transform binary rewards into dense token-level self-supervision through dual-role training and on-policy self-distillation, achieving superior performance in reasoning tasks with reduced sample efficiency requirements.

Current post-training methods in verifiable settings fall into two categories. Reinforcement learning (https://huggingface.co/papers?q=Reinforcement%20learning)(RLVR) relies on binary rewards (https://huggingface.co/papers?q=binary%20rewards), which are broadly applicable and powerful, but provide only sparse supervision during training. Distillation (https://huggingface.co/papers?q=Distillation) provides dense token-level supervision (https://huggingface.co/papers?q=token-level%20supervision), typically obtained from an external teacher or using high-quality demonstrations. Collecting such supervision can be costly or unavailable. We propose Self-Distillation (https://huggingface.co/papers?q=Self-Distillation)Zero (SD-Zero), a method that is substantially more training sample-efficient than RL and does not require an external teacher or high-quality demonstrations. SD-Zero trains a single model to play two roles: a Generator, which produces an initial response, and a Reviser, which conditions on that response and its binary reward to produce an improved response. We then perform on-policy self-distillation (https://huggingface.co/papers?q=on-policy%20self-distillation) to distill the reviser into the generator, using the reviser’s token distributions conditioned on the generator’s response and its reward as supervision. In effect, SD-Zero trains the model to transform binary rewards (https://huggingface.co/papers?q=binary%20rewards) into dense token-level self-supervision. On math and code reasoning benchmarks with Qwen3-4B-Instruct and Olmo-3-7B-Instruct, SD-Zero improves performance by at least 10% over the base models and outperforms strong baselines, including Rejection Fine-Tuning (RFT), GRPO, and Self-Distillation (https://huggingface.co/papers?q=Self-Distillation)Fine-Tuning (SDFT), under the same question set and training sample budget. Extensive ablation studies show two novel characteristics of our proposed algorithm: (a) token-level self-localization (https://huggingface.co/papers?q=token-level%20self-localization), where the reviser can identify the key tokens that need to be revised in the generator’s response based on reward, and (b) iterative self-evolution (https://huggingface.co/papers?q=iterative%20self-evolution), where the improving ability to revise answers can be distilled back into generation performance with regular teacher synchronization (https://huggingface.co/papers?q=teacher%20synchronization).

View arXiv page (https://arxiv.org/abs/2604.12002)View PDF (https://arxiv.org/pdf/2604.12002)Add to collection (https://huggingface.co/login?next=%2Fpapers%2F2604.12002)

Get this paper in your agent:

hf papers read 2604.12002

Don’t have the latest CLI?curl -LsSf https://hf.co/cli/install.sh | bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2604.12002 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2604.12002 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2604.12002 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to a collection (https://huggingface.co/new-collection) to link it from this page.

Similar Articles

Self-Distilled Agentic Reinforcement Learning

Hugging Face Daily Papers

SDAR enhances multi-turn agent training by integrating self-distillation with a sigmoid gate to selectively strengthen positive token-level guidance while mitigating negative teacher rejections, achieving significant improvements over GRPO across multiple benchmarks.

Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information

Hugging Face Daily Papers

Proposes Anti-Self-Distillation (AntiSD) which reverses the knowledge transfer direction in self-distillation to improve math reasoning efficiency and accuracy, achieving GRPO baseline accuracy in 2-10x fewer steps and up to 11.5 points higher final accuracy across models from 4B to 30B parameters.

G-Zero: Self-Play for Open-Ended Generation from Zero Data

Hugging Face Daily Papers

This paper introduces G-Zero, a verifier-free framework that enables autonomous large language model self-improvement through co-evolutionary training using intrinsic rewards and hint-based guidance. It aims to overcome the limitations of proxy LLM judges in open-ended tasks by deriving supervision from internal distributional dynamics.