Hidden States Reveal Task-Relevant Information in Chain-of-Thought Reasoning

A recent research paper on arXiv (2604.23351) examines if chain-of-thought (CoT) tokens hold task-specific information beyond simple explanations. By employing activation patching on GSM8K, the researchers transferred hidden states at the token level from a CoT generation to a direct-answer response for the same query. The findings indicate that patching achieves greater accuracy than both direct-answer prompting and the initial CoT trace, suggesting that individual CoT tokens possess enough information to derive the correct answer, even if the original trace is flawed. Task-relevant information is found to be more abundant in correct CoT executions, unevenly spread across tokens, with a concentration in mid-to-late layers and appearing earlier in the reasoning process. The study also examines language tokens, including verbs.