Zero-Plus-Few-Shot CoT

Insert a zero-shot reasoning trigger into few-shot CoT exemplars so each demo still shows steps but also cues the model with the task-agnostic phrase.

Zero-Plus-Few-Shot-CoT mixes zero-shot chain of thought with Wei-style few-shot CoT: keep ⟨question, reasoning, answer⟩ exemplars, but insert the reasoning trigger (e.g. “Let’s think step by step”) into those demonstrations (and typically into the query). On MultiArith / GSM8K with InstructGPT / PaLM-class models, Kojima et al. report this hybrid above plain Few-shot-CoT and above Zero-shot-CoT alone.

Harness use: when you already maintain exemplar packs, try adding the same trigger used for zero-shot CoT rather than treating the two as mutually exclusive. Still gate with when chain of thought helps and offline prompt evaluation; stacking trigger + exemplars + self-consistency decoding multiplies cost. Prefer this before inventing new exemplar content when the gap is “forgot to reason” rather than “wrong domain demos” — cross-task few-shot CoT transfer.

---

Sources

• Large Language Models are Zero-Shot Reasoners (Kojima et al.), §3 Table 2