OpenAI has unveiled a preview of Reinforcement Fine-Tuning (RFT) for its O1 series of models, enabling users to tailor models for specific tasks using reinforcement learning. Unlike traditional fine-tuning, which often focuses on imitation, RFT emphasizes teaching models to reason, leading to impressive performance gains with minimal data. This innovative technology, already in use at OpenAI, is now available through a research program for universities, researchers, and businesses, with a public release expected early next year. A recent demonstration highlighted RFT's effectiveness, showing how it enhanced a smaller model's performance in diagnosing complex genetic diseases, surpassing that of a larger model.
