Rethinking "RL Generalizes, SFT Memorizes": The Role of SFT Data
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
This paper revisits the claim that "RL generalizes, SFT memorizes" by examining how SFT training data affects generalization on Qwen2.5 and Qwen3 models. The authors find that data source matters: a mixed SFT dataset outperforms data from a larger model alone.