Beneath the Surface of Chains-of-Thought: A Mechanistic Interpretation of Reasoning Operations in LLMs
NAVER AI Lab
Finds that functional reasoning operations in CoT — problem formulation, goal decomposition, deduction — are linearly separable in held-out hidden states, peaking in middle layers and not explained by lexical or positional confounds. Attention-masking interventions show operation-aligned representations at chunk onset depend on preceding reasoning context. EMNLP 2026 Main, code on github.com/naver-ai/beneath-cot.
Why it matters
Gives mechanistic evidence that CoT text corresponds to internal geometric structure — a step toward verifying whether written reasoning actually reflects the model's computation rather than post-hoc rationalization.
Importance: 2/5
Notable mech-interp paper accepted to EMNLP 2026 Main