#artificial intelligence
May 2026
Complementing reinforcement learning with SFT through logit averaging in the post training of LLMs
A novel method that averages the logits of a frozen reference policy and a trainable policy and the reference anchor is coupled through the logit averaging structure to leverage the reasoning expertise of the trainable policy while maintaining the formatting advantage of SFT.
X. Gan, Ying-Tao Zhu
· arXiv.org · 0 citations