Get Started
Home
Topics
Search
Library
Research questionWhen are SFT, ICL, and KL-regularized RL mathematically equivalent, and how does test-time reasoning differ?These paradigms are often treated as fundamentally different, even though several can implement related posterior projections. Their relationship changes depending on whether the signal comes from likelihoods, rewards, or advantages and whether learning occurs in weights or at inference time.
AI
LLM Pretraining & Post-training
Machine Learning
Reasoning
Reinforcement Learning
Statistical Machine Learning
Latest papersRecent research connected to this question, newest first.Unifying ICL, SFT, KL-Regularized RL Through a Bayesian LensThe source is a conceptual note covering SFT, ICL, KL-regularized RLHF/RLVR, on-policy distillation, reward-weighted ICL and SFT, advantage-weighted SFT, and test-time search. It formalizes where objective-level and first-order equivalences hold, while noting practical constraints such as the role of supervised warm-up for importance-weighted projections; it does not provide a broad empirical comparison.research paper · Sep 4, 2026
Related questions
Should reasoning LLMs combine on-policy distillation and verifiable-reward RL jointly or sequentially during post-training?How should vision-language model post-training balance perception and reasoning under SFT versus reinforcement learning?How can reasoning-model fine-tuning improve pass@1 without reducing pass@k coverage?How can on-policy LLM reasoning turn unreliable skill-derived signals into reliable token-level supervision?