DIME: Fine-grained Interpretations of Multimodal Models via Disentangled Local Explanations

Multimodal AI systems (those combining images, text, audio, etc.) are increasingly used in high-stakes domains like healthcare and robotics, but their decision-making processes remain poorly understood. The paper addresses this gap by developing DIME, a method to explain how these complex systems work.

DIME breaks down model decisions into two understandable components: how individual data types (like images alone or text alone) contribute to decisions, and how interactions between different data types influence outcomes. This "disentangled" approach makes it easier to identify which parts of the system are driving specific predictions.

The method works across different types of multimodal systems and tasks without requiring redesign, suggesting it could be broadly applied to explain AI systems already deployed in real-world applications. This generality is important for practical implementation in regulated industries.

Better explanations of AI decision-making support critical responsible AI goals: helping stakeholders visualize and debug model behavior, building appropriate trust in AI systems, and enabling more effective human-AI collaboration rather than blind reliance on automated decisions.

Fine-grained interpretability of multimodal models is particularly important because these systems' complexity—combining multiple data sources—makes their failures harder to diagnose and their biases harder to detect without tools that can isolate which components caused problematic decisions.