Training Differences Between RFT and RL
heghbalz · x · 2026-07-10
This repost discusses the differences between Rejection Fine-Tuning (RFT) and RL.
The core conclusions are:
- RFT often shows rapid early gains but subsequently hits a plateau;
- The key difference isn't "exploration volume," but selectivity: RFT is more prone to generating shortcut-driven, not necessarily effective rewrites;
- RL focuses exploration more heavily on effective and reusable structures.
The latter half of the post emphasizes that with only final answer rewards, RL can still solve held-out problems the base model fails at. It first sharpens basic skills, then combines them into a more stable, programmatic toolbox.
More from Research
- Structural ensembles beat single predictions in TCR:pMHC generalization study — quaidmorris · 2026-07-22
- RSS launches under OMSF to push structural biology data modeling at scale — MoAlQuraishi · 2026-07-22
- enFoldX tops 8 neoantigen scans and an unseen-peptide benchmark — quaidmorris · 2026-07-22
- enFoldX reaches AUC 0.82 on human VDJdb and transfers to mouse at 0.76 — quaidmorris · 2026-07-22
- enFoldX gains accuracy as AF3 ensemble disagreement rises for non-binders — quaidmorris · 2026-07-22
- A 3D ray plot shows how hard this Jacobian counterexample is to read — moultano · 2026-07-22