Chain-of-thought (CoT) verification is a growing area of research, but current methods have a blind spot. They typically check faithfulness by looking at the model's final outputs or static activations, which can tell you that something is wrong but not how the unfaithfulness developed during generation. That limits their usefulness for debugging or understanding model behaviour.

The new paper introduces VERITAS, a method designed to close that gap. While the abstract does not detail the inner workings, the core idea is to verbalise the model's computation—essentially making the internal reasoning process explicit and inspectable. This would allow researchers to observe unfaithfulness as it unfolds, rather than only after the fact.

If VERITAS works as described, it could become a valuable tool for auditing AI systems. By showing exactly where a model's reasoning diverges from its stated chain of thought, it would give developers a clearer target for correction and give users more confidence in the reliability of CoT explanations. The paper is available on arXiv, though full details of the method are not yet public.