Faithfulness Metric

The gold standard for measuring faithfulness involves counterfactual interventions on the model's internal state. By surgically altering a specific intermediate reasoning step and observing whether the final answer changes accordingly, the metric verifies a direct causal link. If the reasoning trace states 'The patient has a fever, therefore they have an infection,' a faithfulness test would ablate the 'fever' representation and check if the 'infection' conclusion disappears. This distinguishes causal dependence from mere correlation in the generated text.

A faithful reasoning chain must be sensitive to counterfactual edits. The metric works by introducing a minimal, semantically targeted change to the input context—such as flipping a single premise—and measuring whether the reasoning trace adapts coherently. A high faithfulness score requires that the new trace reflects the altered premise while preserving logical structure. If the model changes its answer but keeps an identical reasoning path, the original trace is flagged as post-hoc rationalization rather than a genuine causal driver.

This metric decomposes the relationship between intermediate reasoning steps and the final answer. It measures the statistical correlation between the correctness of each step and the correctness of the outcome. A faithful model exhibits a strong positive correlation: correct intermediate steps reliably lead to correct answers, and errors propagate. A low correlation suggests the model is using spurious heuristics or the Clever Hans effect, where the reasoning trace is a decorative narrative unrelated to the actual decision process.

A direct quantitative metric computed by systematically removing or zeroing out the attention heads, MLP layers, or specific neurons identified as responsible for a reasoning step. The faithfulness score is the magnitude of change in the output distribution when the component is ablated. A high score indicates that the identified component was necessary and sufficient for the reasoning step. This method, rooted in mechanistic interpretability, provides a hardware-level verification that the generated explanation maps to real computational circuits.

This metric tests faithfulness by injecting irrelevant or misleading distractor sentences into the context and measuring whether the reasoning trace remains logically anchored to the relevant premises. A faithful model's trace will ignore the distractors, demonstrating that its reasoning is causally grounded in the true evidence. If the trace incorporates or is derailed by the distractor, the metric flags the reasoning as brittle and non-causal, revealing that the model is pattern-matching rather than performing robust logical deduction.

This metric compares the token-level attention weights or gradient-based saliency maps against the explicit claims in the reasoning trace. For each statement in the chain-of-thought, the metric verifies that the tokens cited as evidence actually received high attention scores during computation. Misalignment—where the trace claims to use premise A but attention was focused on premise B—indicates confabulation. This provides a fine-grained, token-by-token faithfulness audit of the entire reasoning trajectory.