New Paper Challenges "RL Generalizes, SFT Memorizes" Consensus
A new NeurIPS paper from NUS/Yale researchers challenges the widely assumed dichotomy that RL generalizes while SFT only memorizes, showing that with properly composed SFT data, SFT can outperform RL in out-of-domain reasoning on models like Qwen2.5.
2026-11-17 ~ 2026-11-18 · 2 related posts
- NeurIPS paper: with mixed data and RL-matched scale, SFT can generalize as well as or better than RL — kastnerkyle · 2026-11-17
- New paper: SFT can generalize as well as RL given enough diverse data — josh_wills · 2026-11-18