BioPro: Training-Free Framework for Difference-Aware Gender Fairness in VLMs

A recent study presents BioPro (Bias Orthogonal Projection), a framework that operates without the need for training, aimed at achieving gender fairness that is aware of differences in Vision-Language Models (VLMs). This research expands the principle of difference-aware fairness, previously limited to text-only models, into the multimodal sphere, focusing on tasks like image captioning and text-to-image generation. Existing fairness measures typically apply a one-size-fits-all approach across demographic groups, neglecting the need for context-specific neutrality versus valid group characteristics. BioPro selectively reduces bias in neutral scenarios while maintaining necessary distinctions in explicit contexts. The framework identifies a low-dimensional subspace for bias reduction without model retraining. This paper, published on arXiv under ID 2512.00807v2, supersedes an earlier version and builds on recent advancements in difference-aware fairness for text-only models, formalizing the issue for multimodal AI systems.