Harvard's Projection Sampling Makes Standard Fine-Tuning Match Reinforcement Learning

Harvard researchers use MCMC sampling to reshape expert data into on-policy traces, letting supervised finetuning match or beat RL on generalization.