Paper finds CoT reasoning operations are geometrically organized in hidden states

dair_ai · x · 2026-09-08

The paper "Beneath the Surface of Chains-of-Thought" asks whether distinct reasoning operations — problem formulation, goal decomposition, deduction — are geometrically organized in hidden representations. They are: operations are separable in held-out representations, with separability peaking in middle layers. Lexical and positional confounds are ruled out; identical surface tokens get different representations depending on the surrounding operation, so geometry tracks function rather than wording. Attention masking shows operation-aligned representations at chunk start depend on preceding reasoning context — the operation label is constructed from the trace, not read off the current token.

Related event: NAVER Study Reveals Geometric Structure of CoT Reasoning Steps in Hidden Layers(2 posts)→

Original post →

More from Research

Research channel →