Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Free

Intermediate reasoning steps boost large language model performance on complex tasks.

FreeFree tier

Inputs: textOutputs: text

Type

Open Source

Visit Website GitHub (40 stars)

About Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Chain-of-Thought Prompting is a method introduced in a 2022 research paper that significantly improves the ability of large language models (LLMs) to perform complex reasoning. By generating a series of intermediate reasoning steps (a 'chain of thought') before arriving at an answer, the technique elicits reasoning capabilities that emerge in sufficiently large models. The paper demonstrates that providing just a few chain-of-thought exemplars in a prompt yields striking empirical gains on arithmetic, commonsense, and symbolic reasoning tasks. For example, prompting a 540B-parameter model with eight chain-of-thought exemplars achieves state-of-the-art accuracy on the GSM8K math word problem benchmark, surpassing even fine-tuned GPT-3 with a verifier. The method is simple, requires no model fine-tuning or architectural changes, and has become a foundational technique in prompt engineering.

Key Features

Generates a series of intermediate reasoning steps (chain of thought) before final answer

Improves performance on arithmetic, commonsense, and symbolic reasoning tasks

Achieves state-of-the-art accuracy on GSM8K math word problem benchmark with 8 exemplars

Simple few-shot prompting method requiring no model fine-tuning or architectural changes

Reasoning abilities emerge naturally in sufficiently large language models (e.g., 540B parameters)

Applicable to a range of reasoning tasks with minimal exemplar engineering

Pros & Cons

Pros

Simple and easy to implement—just add few-shot examples with intermediate steps
Requires no model fine-tuning, only prompt modification
Significant performance gains across diverse reasoning tasks
Generalizable; works with multiple large language models
Transparent reasoning process—outputs intermediate steps for interpretability

Cons

Effectiveness is highly dependent on model scale; smaller models may not show gains
May require carefully crafted exemplars for optimal performance
Limited to prompting paradigm—some tasks may still benefit from fine-tuning
Not a standalone tool; requires a compatible large language model (e.g., 540B parameters)
Paper is a research contribution, not a deployed product or API

Best For

Solving math word problems (e.g., GSM8K benchmark)Commonsense reasoning tasksSymbolic reasoning tasksEducational AI for step-by-step problem solvingPrompt engineering to extract better reasoning from LLMsResearch on emergent reasoning capabilities in large language models

FAQ

What is chain-of-thought prompting?

Chain-of-thought prompting is a method where a few exemplars that include intermediate reasoning steps are provided in the prompt. This encourages the language model to generate a chain of thought—a series of reasoning steps—before arriving at the final answer, significantly improving performance on complex reasoning tasks.

Which models were tested in the paper?

The paper experiments with three large language models, including a 540B-parameter model. The specific model names are not listed in the abstract, but the results show that reasoning abilities emerge in sufficiently large models.

What tasks benefit from chain-of-thought prompting?

The paper shows improvements on a range of arithmetic, commonsense, and symbolic reasoning tasks. The most notable gain is on GSM8K, a benchmark of math word problems.

Does chain-of-thought prompting require fine-tuning?

No. The method uses few-shot prompting with exemplars; no fine-tuning or architectural changes to the model are needed.

Is chain-of-thought prompting effective for all model sizes?

The paper observes that reasoning abilities emerge naturally in sufficiently large language models. Smaller models may not show the same level of improvement.