The story’s score and claims

Comments on No. 73 of 144 · arXiv · edition of Thursday 19 November 2026

Rethinking "RL Generalizes, SFT Memorizes": The Role of SFT Data

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.

Comments

Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules