Chain of Thought

Definitions

1. Chain of Thought is a prompting technique and reasoning approach in large language models where the model is encouraged to show its step-by-step reasoning process rather than jumping directly to an answer. Research by Wei et al. at Google in 2022 demonstrated that asking models to think step by step significantly improves performance on complex reasoning tasks, particularly mathematics, logic puzzles, and multi-step problem solving. Instead of producing just a final answer, the model generates intermediate reasoning steps, which helps it track the state of a problem and avoid errors that occur when trying to compute everything in a single step. Chain-of-thought prompting can be triggered by including worked examples in the prompt (few-shot) or simply by adding phrases like think step by step (zero-shot). Subsequent techniques like tree-of-thought and self-consistency extend the approach by exploring multiple reasoning paths. Chain of thought has become a standard technique in prompt engineering and is built into many AI application frameworks.

   In plain English: Getting an AI to 'show its work' — reasoning through a problem step by step instead of jumping straight to an answer, which makes it much more accurate.

2. Chain-of-thought (CoT) prompting reveals how sensitive LLMs are to the structure of their own output. By generating reasoning tokens before the answer, the model conditions its final response on its own intermediate reasoning — essentially 'thinking out loud' improves the answer.

   Example: 'Adding chain-of-thought to our math tutoring prompt improved accuracy from 60% to 92%. The model needs to show its work, just like students.'

   Source: mechanism explanation

Related Terms

• Prompt Engineering
• GPT
• Vibe Coding
• Agentic AI
• system prompt
• scaffolding
• Few-Shot Learning
• Chain of Thought Prompting
• Knowledge Graph
• Anchor Token