New paper: SFT can generalize as well as RL given enough diverse data

josh_wills · x · 2026-11-18

Since DeepSeek-R1, "RL generalizes, SFT memorizes" has been treated as gospel, with plateaued SFT runs blamed on the algorithm. A new paper argues the comparison was unfair: RL gets a huge rollout budget per prompt while SFT trains on a tiny data slice. Testing on Qwen2.5 and Qwen3, the authors show that mixing data sources instead of distilling from one giant teacher and scaling up data lets SFT generalize just as well as RL. Practitioners conclude the mid-training/SFT boundary is increasingly meaningless — high-quality, diverse data is always the best bet.

Related event: New Paper Challenges "RL Generalizes, SFT Memorizes" Consensus(2 posts)→

Original post →

More from Research

Research channel →