Flow Matching Method for Few-Shot Vision-Language Adaptation

A new paper on arXiv (2605.05054) critiques existing flow matching (FM) methods for few-shot adaptation of vision-language models, identifying three key limitations from a polar decomposition perspective: angular dynamics distortion due to radial-angular coupling, neglect of radial dynamics from feature normalization discarding modality confidence, and context-agnostic unconditional generation. The authors propose Direct Product Flow Matching to decouple radial and angular dynamics, aiming to improve adaptation performance.