نسخة أولية وصول مفتوح
Fenchel Tilting: Weighted Correction for Efficient Finetuning of Generative Models
Adapting a pretrained generative model to an arbitrary preference expressed as a utility function underlies reward alignment, guided design, and constraint satisfaction, enabling diverse applications. Existing fine-tuning methods trade off generality against computational cost: they either restrict the family class of …