Measuring faithfulness in chain-of-thought reasoning

When a model explains its answer step by step, the explanation reads as if it describes how the answer was produced. Often it does not. A model can reach an answer by one route and then write down a different, cleaner route that merely justifies it. If we want to trust step-level reasoning, we first need to measure how often the written steps match the real ones.

The method

We start from a simple observation. If a reasoning step genuinely contributes to the answer, then corrupting that step should change the answer. If the answer survives the corruption unchanged, the step was decorative.

For each problem in our test set, we take the model's chain of thought and produce three variants: one with a step negated, one with a step swapped for a plausible but irrelevant step, and one with a step removed outright. We then re-run the model on each variant and record whether the final answer changes. The faithfulness score is the share of steps whose corruption changes the answer, averaged over problems.

What we found

Three results stood out. First, scale alone barely helps: the jump from 8B to 70B parameters bought less than ten points of faithfulness. Second, irrelevant steps are the most common failure mode, accounting for roughly half of all unfaithful steps across models. Third, adding a separate verifier that scores each step raised faithfulness more than any change to the reasoning model itself, which is the finding that motivates our ongoing work on making verification cheap enough to run in production.

Limits

Corruption testing measures whether steps matter, not whether they are honest. A step can be load-bearing and still misdescribe the underlying computation. We also tested only math and symbolic tasks, where steps are easy to corrupt cleanly. Extending the method to open-ended reasoning is ongoing work.