SRFT: A Single-Stage Method with Supervised and Reinforcement Fine-Tuning for Reasoning

TL;DR

SRFT is a novel single-stage training method that combines supervised fine-tuning and reinforcement learning for improved reasoning in large language models, leveraging entropy-based mechanisms for better integration.

Contribution

The paper introduces SRFT, a unified single-stage approach that effectively integrates supervised and reinforcement fine-tuning for reasoning tasks in LLMs.

Findings

SRFT achieves 59.1% average accuracy on reasoning benchmarks.

Outperforms zero-RL methods by 9.0% on five benchmarks.

Improves out-of-distribution performance by 10.9%.

Abstract

Large language models (LLMs) have achieved remarkable progress in reasoning tasks, yet the optimal integration of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) remains a fundamental challenge. Through comprehensive analysis of token distributions, learning dynamics, and integration mechanisms from entropy-based perspectives, we reveal key differences between these paradigms: SFT induces coarse-grained global changes to LLM policy distributions, while RL performs fine-grained selective optimizations, with entropy serving as a critical indicator of training effectiveness. Building on these observations, we propose Supervised Reinforcement Fine-Tuning (SRFT), a single-stage method that unifies both fine-tuning paradigms through entropy-aware weighting mechanisms. Our approach simultaneously applies SFT and RL to directly optimize the LLM using demonstrations and self-exploration rollouts rather than through two-stage sequential methods. Extensive experiments show that SRFT achieves 59.1% average accuracy, outperforming zero-RL methods by 9.0% on five mathematical reasoning benchmarks and 10.9% on three out-of-distribution benchmarks.