Zero-shot CoT

Append 'Let's think step by step.' Kojima et al 2022. Works without examples on large models. Cheap win.

Advertisement

Few-shot CoT

Provide examples where the reasoning is shown. Model imitates. Stronger than zero-shot CoT on hard tasks.

Advertisement

Why it works

Generates 'scratchpad' tokens. Model uses them as working memory. Emerges at sufficient scale (~50B parameters historically).

Failure modes

Confidently wrong reasoning. Post-hoc rationalization. Small models fake steps without actual reasoning.