Measuring Faithfulness In Chain Of Thought Reasoning

Shows that stated reasoning may not reflect what drives a model's answer, and that this worsens with scale on most tasks. Matters when you monitor or audit agents through their reasoning traces.

• chain-of-thought — Having a model write out intermediate reasoning steps before its answer, which markedly improves performance on hard problems.

• articleExploit Evals
• articleAttack Navigator
• articleN Days
• articleScaling Laws And Interpretability Of Learning From Repeated Data

• articleQuestion Decomposition Improves The Faithfulness Of Model Generated ReasoningAnthropic research
• articleLanguage Models Don T Always Say What They Think Unfaithful Explanations In Chain Of Thought Prompting 2023 05 07Cohere
• articleChain-of-Thought Prompting Elicits Reasoning in Large Language ModelsJason Wei et al.

Checking sign-in…

Loading comments…