Quantifying and Mitigating Unimodal Biases in Multimodal Large Language Models: A Causal Perspective

Recent advancements in LLMs have facilitated the development of Multimodal LLMs (MLLMs). Despite their impressive capabilities, MLLMs often suffer from an over-reliance on unimodal biases (e.g., language bias and vision bias), leading to incorrect answers in complex multimodal tasks. To investigate this issue, we propose a causal framework to interpret the biases in Visual Question Answering (VQA) problems. Within our framework, we devise a causal graph to elucidate the predictions of MLLMs on VQA problems, and assess the causal effect of biases through an in-depth causal analysis. Motivated by the causal graph, we introduce a novel MORE dataset, consisting of 12,000 VQA instances. This dataset is designed to challenge MLLMs' abilities, necessitating multi-hop reasoning and the surmounting of unimodal biases. Furthermore, we propose two strategies to mitigate unimodal biases and enhance MLLMs' reasoning capabilities, including a Decompose-Verify-Answer (DeVA) framework for limited-access MLLMs and the refinement of open-source MLLMs through fine-tuning. Extensive quantitative and qualitative experiments offer valuable insights for future research. Our project page is at https://opencausalab.github.io/MORE.