We found Chains-of-Thought largely aren’t “faithful”: the rate of mentioning the hint (when they used it) was on average 25% for Claude 3.7 Sonnet and 39% for DeepSeek R1.

→ View original post on X — @anthropicai