Social Debiasing for Fair Multi-modal LLMs

TL;DR

This paper introduces a new dataset and debiasing strategy for multi-modal large language models to reduce social biases related to race and gender, while maintaining their reasoning capabilities.

Contribution

It presents a comprehensive counterfactual social concept dataset and a novel counter-stereotype debiasing method that effectively mitigates biases in MLLMs.

Findings

CMSC dataset enhances bias detection and mitigation.

CSD strategy reduces social biases effectively.

Models retain performance on reasoning benchmarks.

Abstract

Multi-modal Large Language Models (MLLMs) have dramatically advanced the research field and delivered powerful vision-language understanding capabilities. However, these models often inherit deep-rooted social biases from their training data, leading to uncomfortable responses with respect to attributes such as race and gender. This paper addresses the issue of social biases in MLLMs by i) introducing a comprehensive counterfactual dataset with multiple social concepts (CMSC), which complements existing datasets by providing 18 diverse and balanced social concepts; and ii) proposing a counter-stereotype debiasing (CSD) strategy that mitigates social biases in MLLMs by leveraging the opposites of prevalent stereotypes. CSD incorporates both a novel bias-aware data sampling method and a loss rescaling method, enabling the model to effectively reduce biases. We conduct extensive experiments with four prevalent MLLM architectures. The results demonstrate the advantage of the CMSC dataset and the edge of CSD strategy in reducing social biases compared to existing competing methods, without compromising the overall performance on general multi-modal reasoning benchmarks.