Learning to Adapt SFT Data for Better Reasoning Generalization

Researchers introduced DART, a method that adapts supervised fine-tuning data to better match a model’s distribution.

DART uses reinforcement learning to transform existing SFT demonstrations into model-adapted supervision before fine-tuning.

The approach improves reasoning generalization across multiple language models and datasets.

It offers a practical way to make fixed training data more effective, without relying on reinforcement learning alone.