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.

Source

AlphaSignal

Published

Oct 1, 2026, 5:45 PM

Author

AlphaSignal Newsroom

Read

1 min read

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

Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.

Next reads

$Shanghai AI Lab's RePro Cuts Reasoning Tokens 47% While Boosting Math Scores$

$MIT Finds a Simple Word Prefix Boosts AI Math Scores by 36 Points$