A Closed-Form Solution for Debiasing Vision-Language Models with Utility Guarantees Across Modalities and Tasks

TL;DR

This paper presents a training-free, closed-form debiasing method for vision-language models that guarantees fairness improvements with limited utility loss across multiple modalities and tasks.

Contribution

It introduces a novel closed-form, utility-guaranteed debiasing approach that is training-free and applicable across visual and textual modalities without annotated data.

Findings

Outperforms existing debiasing methods on fairness metrics

Maintains task performance while reducing biases

Effective across diverse downstream tasks and datasets

Abstract

While Vision-Language Models (VLMs) have achieved remarkable performance across diverse downstream tasks, recent studies have shown that they can inherit social biases from the training data and further propagate them into downstream applications. To address this issue, various debiasing approaches have been proposed, yet most of them aim to improve fairness without having a theoretical guarantee that the utility of the model is preserved. In this paper, we introduce a debiasing method that yields a \textbf{closed-form} solution in the cross-modal space, achieving Pareto-optimal fairness with \textbf{bounded utility losses}. Our method is \textbf{training-free}, requires \textbf{no annotated data}, and can jointly debias both visual and textual modalities across downstream tasks. Extensive experiments show that our method outperforms existing methods in debiasing VLMs across diverse fairness metrics and datasets for both group and \textbf{intersectional} fairness in downstream tasks such as zero-shot image classification, text-to-image retrieval, and text-to-image generation while preserving task performance.