AI and the World's Hard Problems · Theoretical computer science · Anthropic · 2026-09-28
Measuring faithfulness in Chain-of-Thought reasoning
The study asks whether chain-of-thought reasoning reflects a model's actual decision process. The authors intervened on the reasoning, such as by adding mistakes or paraphrasing it, and checked how the predictions changed. Faithfulness varied by task, and on most tasks studied it declined as models became larger and more capable.