Faithfulness Metrics

Measures whether the features highlighted in an explanation are independently sufficient to reproduce the model's original prediction.

• Core Logic: If you mask all input features except those identified as important by the explanation, the model should still output the same class with high confidence.
• Failure Mode: An explanation with low sufficiency is omitting critical variables that the model secretly relies on.
• Practical Use: Essential for validating that a rationale captures the complete causal picture, not just a partial justification.

Quantifies the impact of removing the features deemed important by the explanation. A faithful explanation should cause a significant drop in model confidence when its key features are ablated.

• Calculation: Compare the original prediction probability against the probability after zeroing out the top-k% of attributed features.
• Interpretation: A larger drop indicates the explanation successfully identified the true drivers of the decision.
• Contrast with Sufficiency: While sufficiency tests inclusion, comprehensiveness tests the effect of exclusion.

Evaluates the alignment between a generated textual rationale and the model's internal attention weights or saliency maps.

• Methodology: Compare the tokens highlighted in the natural language explanation against the input tokens receiving the highest attention scores.
• Critical Caveat: High correlation does not guarantee true faithfulness, as attention weights can sometimes be misleading; however, a consistent mismatch is a strong indicator of unfaithful rationalization.
• Use Case: A rapid diagnostic for detecting post-hoc rationalization that invents logic not present in the forward pass.

Measures the rate at which modifying only the explanation-identified features causes the model to change its prediction.

• Procedure: Perturb the input by altering or removing the features cited in the rationale while keeping non-cited features constant.
• Target Metric: A high flip rate suggests the explanation is faithful to the decision boundary. A low flip rate implies the explanation is focusing on irrelevant or non-causal correlations.
• Relevance: Directly validates counterfactual faithfulness—if the reason changes, the outcome must change.

Assesses whether a human observer can use the generated explanation to predict the model's output on new, unseen inputs.

• Experimental Setup: Present users with the explanation and a new input (without the model's prediction), then measure how accurately they can guess the model's decision.
• Implication: A high simulatability score indicates the explanation captures the model's true functional logic in a communicable way.
• Limitation: This metric conflates human interpretability with faithfulness; a simple but wrong explanation can sometimes be highly simulatable.

Directly compares the tokens cited in a rationale against a ground-truth explanation derived from mechanistic analysis or synthetic data.

• Gold Standard: Requires access to a dataset where the true causal features are known a priori (e.g., synthetic tasks or heavily annotated benchmarks).
• Precision & Recall: Calculates how many cited tokens are actually causal (precision) and how many causal tokens were successfully cited (recall).
• Advantage: The only metric that provides an absolute measure of faithfulness, rather than a proxy signal, but is rarely available in real-world enterprise settings.

Measures how much a model's prediction changes when the features identified as important by the explanation are removed. A high comprehensiveness score indicates the explanation captured the truly influential inputs.

• Perturbation-based metric: Systematically removes top-k attributed features
• Delta measurement: Compares original prediction probability to the degraded output
• Key insight: Faithful explanations cause a steep drop in confidence when removed

Evaluates whether the features highlighted by an explanation are alone sufficient to maintain the original prediction. Only the top-ranked features are retained while all others are masked.

• Inverse of comprehensiveness: Keeps only the important features
• Minimal set validation: Tests if the explanation identifies a complete causal set
• Practical use: Detects explanations that omit critical secondary factors

The correlation between the importance scores assigned by an explanation method and the true sensitivity of the model to those features. High fidelity means the explanation ranks features in the correct order of influence.

• Rank correlation metrics: Spearman's ρ or Kendall's τ
• Ground truth comparison: Requires a known causal structure or oracle model
• Synthetic benchmarks: Often tested on models with deliberately planted dependencies

A metric that measures the necessary condition for explanation faithfulness by testing whether the explanation changes when the model's decision boundary is altered. A faithful explanation must be sensitive to model internals.

• Model randomization test: Progressively randomizes model layers from top to bottom
• Divergence tracking: Measures how explanation diverges as model parameters change
• Theoretical grounding: Based on the axiom that explanations must be a function of model parameters

Quantifies the stability of an explanation under small, semantically meaningless perturbations to the input. A faithful explanation should not fluctuate wildly when noise is added to irrelevant features.

• Local Lipschitz estimate: Maximum change in explanation per unit of input perturbation
• Adversarial testing: Applies controlled noise to non-salient regions
• Practical significance: Unstable explanations erode user trust even if locally accurate

Measures whether a human can use the generated explanation to correctly predict the model's output on new, unseen inputs. This operationalizes faithfulness as a human-in-the-loop behavioral test.

• Forward simulation: Human subjects predict model outputs given only the explanation
• Agreement rate: Percentage of predictions matching the model's actual decision
• Gold standard: Directly tests if the explanation transfers understanding of the model's logic