Google

Chain-of-Thought (CoT) prompting is a technique used to guide a large language model to detail its reasoning process before providing a final answer. This is often achieved by including a simple instruction like 'Let’s think step by step' in the prompt. By verbalizing its thought process, the model can break down complex problems into smaller, manageable parts, which often improves the accuracy and reliability of its conclusions.

Chain-of-Thought Prompting

When prompted to solve a problem, a large language model might begin its response with the phrase 'Let’s think step by step.' This serves as an explicit indicator that the model is about to articulate its reasoning process, a key characteristic of Chain-of-Thought prompting.

Example of a Model Initiating Chain-of-Thought Reasoning

A user wants a language model to solve the following logic puzzle but finds it often gives an incorrect final answer without any justification. Modify the user's prompt to encourage the model to break down its reasoning process before providing the solution.

**Original Prompt:**
'There are three boxes labeled 'Apples', 'Oranges', and 'Apples and Oranges'. You know that every label is incorrect. You are allowed to pick one fruit from one box to determine the correct labels for all three boxes. Which box should you pick from?'

Modifying a Prompt for Step-by-Step Reasoning

A user provides the prompt below to a large language model and receives an incorrect answer. Analyze the model's failure and describe how you would modify the original prompt to guide the model toward a more accurate result. Justify your modification by explaining the underlying principle that makes it effective for multi-step problems.

Improving AI Reasoning for a Multi-Step Problem

The accuracy of chain-of-thought reasoning in language models can be enhanced through various strategic modifications. These include sampling multiple distinct reasoning paths to find consensus (self-consistency), diversifying the provided few-shot demonstrations, and reducing complex problems into smaller, manageable sub-problems.

Improving Chain-of-Thought Reasoning

Zero-shot chain-of-thought reasoning enables large language models to perform step-by-step logical deductions without relying on few-shot examples. This is achieved by prepending simple trigger prompts, such as 'Let’s think step by step', just before the model generates its answer, which yields decent reasoning accuracy.

Zero-Shot Chain-of-Thought Reasoning

Multimodal chain-of-thought reasoning extends step-by-step logical prompting to inputs that combine both text and images. Language models utilizing these multimodal inputs achieve higher reasoning accuracy compared to when they rely on text input alone.

Multimodal Chain-of-Thought Reasoning

A user is trying to solve a word problem using a large language model. They use two different prompts and receive two different responses.

**Prompt 1:**
"A farmer has 5 pens, and each pen holds 8 chickens. The farmer buys 10 more chickens. How many chickens does the farmer have in total?"
**Response 1:**
"The farmer has 50 chickens in total."

**Prompt 2:**
"A farmer has 5 pens, and each pen holds 8 chickens. The farmer buys 10 more chickens. How many chickens does the farmer have in total? Let

Learn Before

Related