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

Inside the paper

The problem

Language models could answer questions from examples in a prompt, but multi-step arithmetic and other reasoning tasks remained difficult. A prompt containing only inputs and final answers gives no example of the intermediate work.

The idea

Include a worked reasoning trace alongside each answer in the prompt. The model is then asked to generate a similar trace and an answer for a new problem.

How it works

• The authors write a small set of examples pairing each question with intermediate reasoning and a final answer.
• These examples are placed in the input before a new question, without updating the model parameters for that task.
• They compare this format with prompts that provide the same questions and final answers without the intermediate steps.
• Experiments vary model scale, example wording and task type, including arithmetic, commonsense and symbolic reasoning.

What it showed

The approach improves performance on several reasoning benchmarks for the larger models studied, while smaller models benefit less consistently. Results vary across different written examples, though all tested sets of reasoning traces outperform the standard prompt on the tasks in the robustness analysis.

Read it for

Read Figure 1 and Section 2 for the prompt format, then Sections 3.2–3.3 for results and ablations. The long example collections in the appendix can be skimmed initially.

Written in our own words from the paper itself. A draft the editors have not yet reviewed.