Unifying ICL, SFT, KL-Regularized RL Through a Bayesian Lens

On September 7, 2026, arXiv published a paper proposing a unified framework based on Bayesian perspectives, aimed at integrating supervised fine-tuning (SFT), few-shot context learning (ICL), and KL-regularized reinforcement learning (RL). The study formalized these three paradigms as摊销 and projection operations for Bayesian posterior predictions. The method consists of two steps: first, constructing a generalized Bayesian posterior using prior models and utility signals; second, approximating it as a parameterized distribution through forward KL projection, corresponding to weights within the context (SFT/RL) or within the context itself (ICL). The paper further demonstrated that KL-regularized RLHF/RLVR, reward-weighted SFT/ICL, and advantage-weighted SFT all involve forward KL projection for posterior predictions induced by rewards or advantages. Additionally, the article explored the implications of this framework for modern inference pipelines, including treating RLHF/RLVR as a “posterior design plus projection” process, and emphasizing the necessity of cold start or supervised预热 for weighted KL projection, while also mentioning…