Zero-Shot Chain of Thought

Elicit multi-step reasoning with a fixed trigger such as “Let’s think step by step,” without per-task few-shot reasoning exemplars.

Zero-shot-CoT (Kojima et al., 2022) is chain-of-thought prompting that drops human- written step-by-step exemplars and uses one task-agnostic reasoning trigger (classically “Let’s think step by step”) across arithmetic, symbolic, and commonsense tasks. Compared with Few-shot-CoT (Wei et al.), it trades per-task exemplar engineering for a second LLM call under two-stage reasoning–answer extraction. It usually underperforms carefully crafted Few-shot-CoT but beats standard few-shot (answer-only exemplars, even at 8-shot) and massively beats plain zero-shot on MultiArith / GSM8K-class tasks when the model is large enough — and on some tasks (e.g. AQUA-RAT with PaLM 540B) Zero-shot-CoT already matches Few-shot-CoT before adding self-consistency decoding. Treat it as the minimalist strongest zero-shot baseline for difficult multi-step system-2 tasks before investing in exemplar packs or finetuning datasets — Kojima et al.’s closing claim for the method. See also zero-plus-few-shot CoT when you already have exemplars.

Few-shot-CoT degrades when exemplar question types mismatch the task; a single fixed zero-shot template can still produce a much better scaling curve on MultiArith / GSM8K-class benchmarks when the model is large enough (see scale-dependent chain of thought). Use it when you want multi-hop reasoning without maintaining exemplar packs, then measure against few-shot CoT with offline prompt evaluation — when chain of thought helps still applies (large multi-step / symbolic / logical gains; easy one-step arithmetic and some commonsense metrics can be flat or regress even when traces look sensible). Sweep the exact cue under reasoning trigger sensitivity; do not assume paraphrases are equal. When few-shot exemplars must come from another task, see cross-task few-shot CoT transfer. Prefer broad vs narrow prompting as the default design question: one task-agnostic trigger vs per-task demos.

---

Sources

• Large Language Models are Zero-Shot Reasoners (Kojima et al.), §1–§3; Table 1–2