Counterfactual Modality Attribution: New Framework for Multimodal LLMs

A new framework called Counterfactual Modality Attribution (CMA) has been developed by researchers to measure how much each modality—image and text—contributes to the predictions of multimodal large language models (MLLMs). This innovation fills a significant void in explainability, as current techniques can highlight key image areas or text tokens but fail to identify which modality primarily influences a model's output. This distinction is crucial, given that a model may arrive at accurate conclusions based on misleading evidence, which could obscure shortcut learning and unsafe reasoning. CMA creates counterfactuals for image-only, text-only, and joint multimodal scenarios using coupled diffusion priors, transforming them into modality attribution scores through a cooperative game-theoretic approach rooted in Shapley values. This framework is the first of its kind for MLLMs. The research paper can be found on arXiv with the identifier 2608.00076, and its abstract notes a cross-listing. This study underscores the increasing significance of explainability in AI systems that aid in critical decision-making by integrating diverse information from various modalities.