Explainability in Multi-Modal Models

Abstract

Multi-modal AI models -- systems that jointly process and reason over text, images, audio, video, and structured data - have achieved remarkable performance across a wide range of tasks, but their explanations have lagged behind their capabilities. Explainability in multi-modal models is substantially harder than in unimodal systems: explanations must attribute predictions not only across input features within a modality but across modalities, identifying whether a prediction was driven by visual context, textual content, their interaction, or some combination. The stakes are high in applications where multi-modal AI supports consequential decisions -- medical imaging with clinical notes, legal document analysis with evidence photos, financial report analysis with market data visualisations. This study evaluates eight explainability methods applied to six multi-modal architectures across four application domains: medical visual question answering, legal document understanding, financial report analysis, and autonomous driving scene comprehension. Explainability methods include gradient-based saliency (GradCAM-MM, Integrated Gradients), attention-based attribution (Cross-Attention Attribution, DIME), perturbation-based (LIME-MM, SHAP-MM), counterfactual explanation (CACE), and a proposed Cross-Modal Explanation Framework (CMEF). CMEF achieves the highest faithfulness score (0.882) and the best user comprehension rating (3.84/5) in a study with 84 domain expert evaluators. GradCAM-MM shows highest visual saliency alignment (IoU 0.724). All methods show significant degradation on cross-modal interaction explanations -- attributing predictions that depend on the interplay between modalities rather than individual modality contributions. A multi-modal explainability taxonomy and evaluation protocol are proposed.