#machine learning
Mar 2026
Stabilizing Policy Optimization via Logits Convexity
This work investigates the stability gap between SFT and RL from a gradient-based perspective, and shows that the convexity of the SFT loss with respect to model logits plays a key role in enabling stable training.
Hongzhan Chen, Tao Yang, Yu-Hua Zhu et al.
· arXiv.org · 0 citations