Prefix-RFT: A Unified Machine Learning Framework to blend Supervised Fine-Tuning (SFT) and Reinforcement Fine-Tuning (RFT)

A recent development in large language model (LLM) training introduces Prefix-RFT, a unified machine learning framework that combines supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT) to leverage the strengths of both methods. While SFT effectively teaches instruction-following through example-based learning, it often results in rigid behavior and limited generalization, whereas RFT optimizes models for task success via reward signals but can introduce instability. Prefix-RFT aims to integrate these approaches, enabling models to benefit from structured instruction while dynamically adapting to task-specific rewards, thus enhancing both flexibility and performance