The paper shows CoT boosts reasoning, but the mechanism is unclear. One hypothesis: CoT simply converts hard one-step problems into easier multi-step ones that happen to lie in the training distribution. Another: it induces genuine intermediate computation. Section 5 mentions emergent properties at 100B+ parameters — does this exclude the distribution-shift explanation, or are both in play?
The 'better distribution' hypothesis seems hard to rule out. But the fact that scratchpad steps need to be plausible (wrong intermediates hurt performance) suggests the model does use them computationally, not just as a format hint.
Right — if it were purely distributional you'd expect even syntactically valid but wrong steps to help, which they don't. That does push toward genuine computation.
The mechanistic interpretability literature hasn't fully settled this for transformers. Some circuits appear to implement multi-step look-up, others seem to shortcut. CoT probably helps both paths.