Distinct reasoning operations are geometrically separable in middle layers.

arXiv 2609.04753 asks whether chain-of-thought operations like problem formulation, goal decomposition and deduction have structure in hidden representations, and finds they're separable in held-out representations with separability peaking mid-network, verified against lexical and positional confounds. Identical surface tokens are represented differently depending on the operation of the surrounding chunk, and attention-masking shows operation-aligned representations at chunk onset depend on preceding reasoning context. A handle for steering reasoning phase rather than reasoning content, which is a knob nobody currently has.