MultiSHAP explains cross-modal interactions in multimodal AI

It does not require access to model internals, only a query interface allowing separate input perturbation and task-relevant score return. The method provides instance-level explanations of how specific visual and text components jointly influence outputs, along with dataset-level summaries of interaction patterns. Experiments show it captures behaviorally relevant interactions and reveals information beyond unimodal saliency in patch-token correspondences.

Summarised from the original article by an open-source model, then checked automatically: every number and English name must appear in the source. The original article is authoritative.