GFT: From Imitation to Reward Fine-Tuning with Unbiased Group Advantages and Dynamic Coefficient Rectification
Unifying supervised fine-tuning and reinforcement learning through group advantage and dynamic stabilization.
Large language models are typically post-trained using supervised fine-tuning (SFT) and reinforcement learning (RL), yet effectively unifying efficient knowledge injection with robust generalization remains challenging. In this work, we provide a training-dynamics analysis showin...
ArXiv CS.AI
Source
Generated insight by OnAirToday AI
GFT addresses the fundamental instability in current post-training pipelines by treating SFT as a flawed optimization process with sparse rewards. By stabilizing gradient flow and diversifying response groups, this framework bridges the gap between imitation learning and more effective reinforcement learning.

Comments
Keep comments relevant and respectful. No links, advertising, self-promotion, or spam.