Technical Summary: Counterfactual Modality Attribution for Multimodal LLMs

Problem Statement

Multimodal Large Language Models (MLLMs) are increasingly deployed in high-stakes domains, such as clinical decision support, where they integrate complementary information from images and text. While existing explainability methods (e.g., saliency maps, attention weights) successfully identify influential features within individual modalities (e.g., specific image regions or text tokens), they fail to answer a fundamental question: which modality actually drives a prediction?

Current methods operate locally within single channels and cannot quantify the relative contribution of entire modalities or their interactions. Consequently, a model may produce a correct output by relying on a "shortcut" from a single modality while ignoring critical evidence in another (e.g., correctly diagnosing pneumonia from an X-ray but ignoring a text note about a penicillin allergy). This hidden modality reliance masks unsafe reasoning patterns that remain invisible to standard predictive accuracy metrics. The paper argues that modality attribution is a distinct and complementary explainability objective to feature attribution, necessary for auditing multimodal foundation models.

Methodology: Counterfactual Modality Attribution (CMA)

The authors propose Counterfactual Modality Attribution (CMA), a model-agnostic framework that quantifies modality-level contributions using a combination of coupled diffusion priors and cooperative game theory. The framework operates in three stages:

1. Coupled Diffusion & Guided Counterfactual Generation

Unlike prior counterfactual methods that intervene on a single modality, CMA generates synchronized image-only, text-only, and joint multimodal counterfactuals within a unified optimization framework.

• Priors: The method utilizes independently pretrained diffusion models: a Denoising Diffusion Probabilistic Model (DDPM) for images and a Riemannian Diffusion Language Model (RDLM) for text.
• Synchronization: During the reverse diffusion process, the image and text models are synchronized at the same timestep. They are jointly guided by gradients from a pretrained MLLM classifier to minimize a Carlini–Wagner margin loss ($L_{CW}$). This ensures that counterfactuals are realistic (lying on the data manifold) and represent minimal interventions that change the prediction.
• Modes:
  • Image-only CF: Text is fixed; the image is optimized to flip the prediction.
  • Text-only CF: Image is fixed; the text is optimized to flip the prediction.
  • Joint CF: Both modalities are optimized simultaneously to capture cross-modal dependencies.

2. Coalition Value Definition

The generated counterfactuals define a two-player cooperative game where the Image ($I$) and Text ($T$) modalities act as players. The value function $v(S)$ for a coalition $S \subseteq \{I, T\}$ is defined as the logit score assigned to the original predicted class $\hat{y}$ when modalities in $S$ remain unchanged and modalities not in $S$ are replaced by their counterfactual counterparts.
The four required coalition values are:

• $v(\{I, T\})$: Original sample $(x_I, x_T)$.
• $v(\{I\})$: Image original, Text counterfactual $(x_I, x'_T)$.
• $v(\{T\})$: Image counterfactual, Text original $(x'_I, x_T)$.
• $v(\emptyset)$: Both counterfactuals $(x'_I, x'_T)$.

3. Shapley Value Attribution

The framework computes Shapley values ($\theta_I, \theta_T$) to quantify the marginal contribution of each modality. For a two-player game, these admit a closed-form solution:
$$\theta_I = \frac{1}{2} [v(\{I\}) - v(\emptyset) + v(\{I, T\}) - v(\{T\})]$$
$$\theta_T = \frac{1}{2} [v(\{T\}) - v(\emptyset) + v(\{I, T\}) - v(\{I\})]$$
The final attribution scores are normalized to percentages (Image% and Text%) to indicate which modality primarily drives the decision.

Key Contributions

1. Complementary Explainability Objective: The paper formulates modality attribution as a necessary complement to feature attribution, addressing the gap in understanding which modality drives a decision.
2. Novel Counterfactual Framework: CMA is the first framework to generate synchronized, realistic image-only, text-only, and joint counterfactuals for pretrained MLLMs using coupled diffusion priors.
3. Principled Attribution Formulation: It introduces a game-theoretic approach using Shapley values to convert multimodal counterfactuals into interpretable scores that quantify individual contributions and non-linear cross-modal interactions.
4. Extensive Validation: The method is validated on controlled synthetic benchmarks and a real-world clinical dataset, demonstrating superior performance over existing baselines in identifying decision-driving modalities.

Experimental Results

The authors evaluated CMA on two benchmarks:

1. Synthetic Multimodal MNIST: A controlled benchmark with known ground-truth modality reliance (Image-Biased, Text-Biased, and Balanced variants).
2. Multiclass OpenI: A real-world chest X-ray dataset paired with radiology reports.

Performance:

• Accuracy: On the Synthetic MNIST benchmarks, CMA correctly identified the ground-truth decision-driving modality in 98% of controlled cases.
• Baseline Comparison: CMA significantly outperformed baselines including Input $\times$ Gradient, Perturbation SHAP, EAGLE, MM-SHAP, and MultiViz. Notably, MM-SHAP and MultiViz achieved 0% and 3% accuracy, respectively, on image-biased samples, failing to recover known model biases.
• Qualitative Findings: In the OpenI dataset, CMA revealed that a model predicting "Nodule" relied on spurious correlations with ECG leads (visible in the image) rather than clinical evidence. Text-only counterfactuals failed to flip the prediction because the image shortcut remained, demonstrating the necessity of joint counterfactual analysis.
• Ablation Study: Progressive corruption experiments showed that CMA attribution scores adapt continuously to the available evidence. As one modality is corrupted, its attribution drops to 0%, shifting entirely to the intact modality, revealing unimodal collapse that accuracy metrics miss.

Significance and Claims

The paper claims that modality attribution is a critical, previously unaddressed dimension of explainability for multimodal AI. By moving beyond feature-level explanations, CMA provides a principled framework for auditing whether a model relies on complementary evidence or unsafe shortcuts.

The authors assert that CMA enables:

• Detection of Hidden Failures: Identifying cases where correct predictions rely on the wrong modality (e.g., ignoring allergy warnings in text).
• Auditing Cross-Modal Reasoning: Quantifying non-additive interactions between modalities that independent unimodal interventions cannot capture.
• Trustworthy Deployment: Providing a mechanism to verify that high-stakes decisions (e.g., in healthcare) are driven by the appropriate combination of evidence sources.

The work concludes that while CMA currently operates at the modality level and incurs computational costs due to diffusion generation, it establishes a foundational approach for systematically auditing multimodal foundation models in safety-critical applications.