Which Modality Decides? Counterfactual Modality Attribution for Multimodal LLMs

2026-07-29

AI Summary

Existing interpretability methods struggle to determine whether predictions from multimodal large language models are driven primarily by visual or textual inputs, and they fail to uncover reasoning flaws stemming from reliance on erroneous evidence. To address this, this work proposes Counterfactual Modality Attribution (CMA), a novel framework that enables attribution analysis at the modality level for the first time. CMA leverages a coupled diffusion prior to generate both unimodal and multimodal counterfactual samples and employs Shapley values to quantify each modality’s contribution to model decisions. By extending interpretability from the feature level to the modality level, CMA demonstrates strong efficacy on both controlled synthetic benchmarks and real-world clinical data, accurately identifying the dominant modality in 98% of cases—significantly outperforming baseline approaches—and offers a powerful new tool for auditing multimodal models in safety-critical applications.

Abstract

Multimodal large language models (MLLMs) increasingly support high-stakes decision making by combining complementary information from images and text. While existing explainability methods identify influential image regions or text tokens, they cannot answer a fundamental question: which modality drives a prediction? Consequently, a model may produce the correct output while relying on the wrong source of evidence, masking shortcut learning and unsafe reasoning. We formulate modality attribution as a complementary explainability objective for multimodal foundation models and propose Counterfactual Modality Attribution (CMA), the first framework for quantifying modality-level contributions in MLLMs. CMA generates image-only, text-only, and joint multimodal counterfactuals using coupled diffusion priors and converts them into principled modality attribution scores through a cooperative game-theoretic formulation based on Shapley values. We evaluate CMA on controlled synthetic benchmarks with known ground-truth modality reliance and on a real-world multimodal clinical dataset. CMA correctly identifies the decision-driving modality in 98% of controlled cases and consistently outperforms baselines, revealing failures of cross-modal reasoning that remain invisible to predictive accuracy alone. Our results establish modality attribution as a complementary dimension of explainability beyond feature attribution, providing a principled framework for auditing multimodal foundation models in safety-critical applications.

Problem

• modality attribution
• multimodal LLMs
• explainability
• counterfactual reasoning
• shortcut learning

Innovation

• Counterfactual Modality Attribution
• Multimodal LLMs
• Shapley values
• Explainability
• Modality Attribution