Definition
Plain language
A mixed-policy reasoning post-training method.
As stated in the literature
Self-Refinement Fine-Tuning, a mixed-policy post-training method blending supervised and on-policy data, used as a baseline in math-reasoning post-training comparisons where headline gains turned out to depend on infrastructure bugs.
Why it matters: Mixed-policy methods like SRFT looked like a clear win until follow-up work showed their gains depended on infrastructure quirks — a reminder that headline improvements in this field often hide subtle bugs.
For example, during post-training the model is updated on a blend of expert-curated solutions and its own attempts, mixing imitation with reinforcement.
Heard on the show
“LUFFY, ReLIFT, SRFT, Prefix-RFT, HPT — methods that said: don't separate the stages, blend them.”Episode 009 — How Two Silent Library Bugs Quietly Invalidated a Wave of Reasoning Papers