Cross-Modal Attribution

Computes the partial derivative of a target modality's output with respect to a source modality's input features. This reveals which pixels in an image were most influential for a specific text token's representation.

• Mechanism: Backpropagation through the fusion layers to the input.
• Key Metric: Saliency maps that highlight cross-modal influence regions.
• Use Case: Debugging why a model associated the word "running" with a specific region of an image containing water, rather than a person.

Linearly combines attention weight matrices across all transformer layers to trace how information propagates from one modality to another. This accounts for the mixing of attention heads in deeper layers.

• Mechanism: Matrix multiplication of attention weights, adding identity matrices to simulate residual connections.
• Output: A single heatmap showing the aggregate flow from image patches to text tokens.
• Advantage: More faithful than raw attention weights alone, as it captures the cumulative effect of attention across the network's depth.

Applies game-theoretic Shapley values to fairly distribute credit for a prediction among coalitions of features from different modalities. This quantifies the marginal contribution of a text phrase given the presence of specific image regions.

• Process: Systematically removes and reintroduces features across modalities to measure their interaction effect.
• Output: A scalar importance score for each feature that satisfies efficiency, symmetry, and additivity axioms.
• Challenge: Computationally expensive, requiring approximations like Kernel SHAP for high-dimensional multimodal inputs.

A causal intervention technique that zeroes out or replaces one entire modality with a neutral baseline to measure the resulting drop in prediction confidence. This quantifies the overall reliance on cross-modal information.

• Procedure: Compare the model's output distribution with and without the visual modality.
• Interpretation: A large drop in accuracy indicates high cross-modal dependence; a small drop suggests the model defaults to a unimodal shortcut.
• Variation: Partial ablation masks specific tokens or image regions to localize the critical cross-modal link.

Computes the path integral of gradients from a neutral baseline input to the actual multimodal input. This satisfies the completeness axiom, ensuring the sum of all feature attributions equals the difference in the model's output.

• Baseline: A blank image and zero-embedding text.
• Path: Linear interpolation in the joint pixel-embedding space.
• Output: Attribution maps for both image pixels and text tokens that are theoretically grounded and implementation-agnostic.

Forces the model to predict a set of human-interpretable concepts that span both modalities before making the final prediction. This architectural constraint enables direct inspection of the cross-modal reasoning logic.

• Example Concept: "Is the object metallic?" or "Is the person smiling?"
• Attribution: The weight of each concept on the final decision is directly readable.
• Benefit: Transforms cross-modal attribution from a post-hoc analysis into a first-class architectural feature, enabling intervention on misattributed concepts.

Visualizations of the attention weights between tokens from different modalities, such as image patches and text words. These maps reveal how a vision-language model grounds linguistic concepts in specific visual regions.

• Directly visualizes the raw attention scores from cross-modal transformer layers.
• Essential for debugging failures in vision-language grounding.
• Often the first diagnostic tool used before applying more complex attribution methods.

An attribution method that computes the path integral of the gradients for all input modalities from a neutral baseline to the actual input. It satisfies the completeness axiom, ensuring the sum of attributions equals the prediction difference.

• Requires defining a suitable multimodal baseline (e.g., black image, zero text embedding).
• Provides a theoretically rigorous decomposition of the output across pixels and tokens.
• A foundational method for feature attribution in the multimodal domain.

A metric that evaluates whether the features identified as important by a cross-modal explanation truly influence the model's prediction. It is measured by perturbing or removing the attributed features and observing the output change.

• High faithfulness means the explanation accurately reflects the model's true reasoning process.
• Contrasts with plausibility, which measures human interpretability.
• Critical for validating the reliability of any cross-modal attribution method.

An explainability method that systematically removes or zeroes out one input modality to measure its causal contribution to the model's final output. This assesses cross-modal reliance.

• A simple but powerful causal intervention technique.
• Can reveal if a model is ignoring one modality entirely (e.g., text bias).
• Often used in conjunction with modality importance weighting to produce scalar contribution scores.

A technique from causal inference applied to multimodal models to identify specific cross-modal neurons or representations that function as causal mediators between an input concept and the model's output.

• Goes beyond correlation to establish causal mechanisms within the network.
• Involves intervening on internal activations and measuring the downstream effect.
• A key tool in the emerging field of mechanistic interpretability for multimodal systems.

An architectural intervention that forces a model to predict a set of human-interpretable concepts spanning multiple modalities before making a final prediction. This enables direct inspection of the cross-modal reasoning process.

• Concepts act as an explicit, auditable intermediate representation.
• Allows for direct intervention and correction of the model's logic.
• Builds upon the Concept Activation Vectors framework for multimodal settings.