Paper: LLM reasoning operations are internally separable — wrong math still carries the right signature

rohanpaul_ai · x · 2026-09-24

The arXiv paper "Beneath the Surface of Chains-of-Thought" offers a mechanistic interpretability analysis of chain-of-thought reasoning:

Reasoning has internal structure: what the model is trying to do and whether it does it correctly appear to be separable internally, opening a route to monitoring reasoning from inside the model.
8 reasoning moves leave distinct fingerprints: fact extraction, decomposition, recall, deduction, algebra and calculation each produce a distinct internal pattern, clearest around middle layers; even the same token looks different depending on the reasoning job it performs.
Context shapes these states: blocking access to the previous 30 tokens weakens the signal for the next reasoning operation.
Most striking: a wrong calculation or deduction can still carry the correct operation signature — the model can represent what kind of reasoning it is attempting without getting it right.

More from Research