Large Language Models are Zero-Shot Reasoners

Problem

Framing

Zero-shot prompting in large LLMs still fails on multi-step reasoning unless the prompt includes task-specific exemplars. The paper shows that one trigger phrase, \"\text{Let's think step by step}\", elicits useful reasoning traces without exemplars, raising MultiArith accuracy from 17.7% to 78.7% and GSM8K from 10.4% to 40.7%.

Currently Used Methods

Foundational

• @brownGPT3_2020 — few-shot in-context prompting with task examples.
  • Limitation in context: zero-shot reasoning remains weak on multi-step tasks.
• @weiCoT2022 — few-shot chain-of-thought exemplars elicit intermediate reasoning.
  • Limitation in context: requires hand-written, task-specific reasoning demonstrations.
• @ouyangInstructGPT2022 — instruction-tuned GPT-3 strengthens zero-shot instruction following.
  • Limitation in context: instruction tuning alone does not unlock stable reasoning traces.
• @kaplanScalingLaws2020 — larger models improve predictably on many language tasks.
  • Limitation in context: reasoning gains stay flat without explicit stepwise prompting.

Proposed Method

Architecture

The method keeps the LLM frozen and changes only inference prompts. It uses two stages: first generate a rationale with the trigger phrase, then extract the final answer with a short answer cue.

Loss / Objective

The paper does not optimize model parameters. Inference uses the pretrained conditional distribution directly.

Algorithm

Zero-shot-CoT forms a reasoning prompt, then an answer-extraction prompt.

Training Procedure

• No finetuning.
• No gradient updates.
• Main model: text-davinci-002.
• Additional models: InstructGPT, GPT-3, PaLM.
• Two inference passes per example.
• Answer-extraction cue varies with answer format.

Evaluation

Datasets

• Arithmetic: SingleEq, AddSub, MultiArith, GSM8K, AQUA-RAT, SVAMP.
• Commonsense: CommonSenseQA, StrategyQA.
• Symbolic: Last Letter, Coin Flip.
• Logical: Date Understanding, Tracking Shuffled Objects.

Metrics

• Accuracy on every benchmark.
• Model-scale accuracy curves for MultiArith and GSM8K.

Headline results

• MultiArith, text-davinci-002: 78.7%.
• GSM8K, text-davinci-002: 40.7%.
• MultiArith, zero-shot baseline: 17.7%.
• GSM8K, zero-shot baseline: 10.4%.
• PaLM 540B: GSM8K 12.5% $\rightarrow$ 43.0%, MultiArith 25.5% $\rightarrow$ 69.5%.

Ablations

• Prompt template: instructive templates stay strong; misleading templates can collapse accuracy.
• Example mismatch: Few-shot-CoT degrades when exemplar task and answer format mismatch.
• Model scale: gains appear mainly in larger models.
• Commonsense tasks: little or no improvement versus zero-shot.

Results table

Table 5: Few-shot-CoT robustness to cross-task exemplars

$\dagger$ CommonsenseQA samples are used as exemplars.

Method Strengths and Weaknesses

Strengths

• No training cost; gains come entirely from prompt changes.
• One trigger transfers across arithmetic, symbolic, and logical tasks.
• Large arithmetic gains: MultiArith improves by 61.0 points.
• Similar gains appear on a second model family, PaLM 540B.

Weaknesses

• Still trails few-shot CoT with curated exemplars on key arithmetic tasks.
• Needs two decoding passes per question.
• Gains are weak on CommonSenseQA and StrategyQA.
• Rationales can be fluent yet logically wrong or overgenerate options.

Suggestions from the authors

• Analyze why a single trigger phrase elicits multi-step reasoning.
• Study broader zero-shot reasoning abilities hidden in pretrained LLMs.
• Improve robustness across tasks and answer formats.
• Explore higher-level, multi-task cognitive capabilities beyond benchmark prompting.

Links

Prior Papers

• @brownGPT3_2020 — establishes the few-shot and zero-shot prompting baseline that this paper strengthens for reasoning.
• @weiCoT2022 — introduces chain-of-thought exemplars, the direct method that Zero-shot-CoT removes.
• @ouyangInstructGPT2022 — provides the instruction-tuned text-davinci models used for the main experiments.

Further Papers

No vault papers identified as further work yet.