Read the original at HF Daily Papers
Researchers propose a metric to measure how well large language model chain-of-thought traces align with internal computations, finding limited agreement across three models and tasks.
Carried by: HF Daily Papers. First seen: .