Multimodal Faithfulness

The core mechanism for measuring faithfulness. Instead of passively observing attention weights, this method actively intervenes by removing, masking, or altering the specific cross-modal features identified as 'important' by an explanation.

• Modality Ablation: Zeroing out the visual input to test if a text-centric explanation is faithful.
• Feature Removal: Deleting specific image regions or text tokens flagged by the explanation.
• Fidelity Check: A faithful explanation will cause a significant, predictable drop in model confidence when its highlighted features are perturbed.

Two quantitative metrics that decompose faithfulness into distinct logical properties.

• Comprehensiveness: Measures how much the prediction changes when the 'important' features are removed. A high score indicates the explanation captured all necessary evidence.
• Sufficiency: Measures if the 'important' features alone are enough to maintain the original prediction. A high score indicates the explanation found the minimal set of decisive features.
• Trade-off: An ideal explanation maximizes both, proving it is both complete and concise.

Evaluates if the explanation captures genuine cross-modal interactions rather than isolated unimodal correlations.

• Synthetic Counterfactuals: Creating images where the visual object contradicts the text (e.g., an image of a dog with the text 'cat') to see if the explanation correctly attributes the conflict.
• Binding Verification: Checking if the explanation links the right words to the right objects, not just relying on statistical co-occurrence.
• Grounded Faithfulness: A faithful multimodal explanation must localize the specific vision-language grounding that drove the decision.

A faithful explanation should remain stable under minor, semantically meaningless perturbations to the input, while correctly identifying when the perturbation is significant.

• Visual Noise: Adding Gaussian noise to the image should not drastically change the explanation if the semantic content is preserved.
• Textual Paraphrasing: Replacing words with synonyms should yield a functionally identical explanation.
• Adversarial Sensitivity: The metric must distinguish between random noise and a true adversarial attack that changes the model's logic, exposing brittle correlations.

A direct operationalization of faithfulness that quantifies the divergence between the original prediction and the prediction on the perturbed input.

• Soft Score: Uses the difference in predicted probability or logit score, not just a hard label flip.
• Area Over the Perturbation Curve (AOPC): A robust metric that measures the average change in prediction probability as features are incrementally removed in order of their attributed importance.
• Baseline Comparison: The AOPC of a faithful explanation should be significantly higher than that of a random or intentionally unfaithful baseline.

Decomposes the overall faithfulness score to diagnose whether the model relies on one modality more faithfully than the other.

• Visual Bias Detection: Reveals if the model is a 'visual shortcut learner' that ignores text, even when the text explanation is faithful.
• Textual Primacy: Identifies if the model over-relies on language priors, making the visual explanation unfaithful.
• Balanced Integration: A truly faithful multimodal system will show high fidelity in both modalities, proving it performs genuine fusion rather than ignoring one data stream.

Measures the drop in prediction confidence when the features identified as important by an explanation are removed. A faithful explanation should cause a sharp decline in the target class probability upon removal.

• Calculation: Original_Score - Score(Input - Important_Features)
• High Comprehensiveness: Indicates the explanation captured truly critical features.
• Limitation: Can be confounded by feature interactions if removal breaks input structure.

Evaluates whether the important features alone are enough to sustain the model's original prediction. A faithful explanation should produce a similar output when only the important features are kept and the rest are masked.

• Calculation: Original_Score - Score(Important_Features_Only)
• Low Sufficiency Score: Suggests the explanation missed context the model relies on.
• Use Case: Detecting explanations that highlight a salient object but ignore necessary background context.

A meta-evaluation that tests the monotonic relationship between the importance rank of features and the effect of their removal. The most important features should cause the largest drop in accuracy when removed sequentially.

• Procedure: Remove features in order of attributed importance and measure the Area Under the Perturbation Curve (AUPC).
• Ideal Result: A steep, monotonic negative slope.
• Failure Mode: A flat or erratic curve indicates the explanation ranks features randomly relative to true causal impact.

A metric that measures how consistent an explanation is under slight perturbations that do not change the model's prediction. A faithful method should produce similar explanations for semantically identical inputs.

• Process: Generate N slightly varied inputs via Monte Carlo sampling or synonym replacement.
• Metric: Compute the maximum discrepancy between the explanation of the original input and any of the N perturbed inputs.
• Goal: Low Sensitivity-N indicates stability, a necessary but insufficient condition for faithfulness.

A rigorous framework borrowed from causal inference to test if a specific cross-modal neuron or representation is a causal mediator. It intervenes directly on internal activations rather than input features.

• Mechanism: Set the activation of a candidate mediator to its baseline value and measure the Total Effect on the output.
• Advantage: Isolates the causal role of internal components without destroying input structure.
• Application: Identifying specific attention heads in a vision-language model that mediate grounding between a noun phrase and an image region.

A benchmark framework that retrains a model from scratch on data where the most important features have been removed. If the explanation is faithful, the retrained model should perform significantly worse than one trained on data with random features removed.

• Comparison: Accuracy(ROAR_Important) vs Accuracy(ROAR_Random).
• Key Insight: Eliminates the confounding effect of out-of-distribution inputs that plague single-instance perturbation tests.
• Cost: Computationally expensive due to full retraining cycles.