Researchers have investigated the information encoded within the hidden states of language models during chain-of-thought (CoT) reasoning. By using activation patching on the GSM8K dataset, they found that individual CoT tokens contain task-relevant information that can significantly improve answer accuracy when transferred to a direct-answer generation process. This task-solving information is more concentrated in correct CoT runs and is unevenly distributed across tokens, appearing earlier in the reasoning trace and in mid-to-late model layers. The study also revealed that language tokens are more crucial for steering correct reasoning, while mathematical tokens primarily encode answer-proximal content. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT Provides new insights into how language models represent and fail during reasoning, potentially guiding future model development.
RANK_REASON Academic paper analyzing the internal workings of language model reasoning.