Chain of Thought (CoT)
Details
- Full Name
-
Chain of Thought Prompting
Core Concepts:
- Step-by-Step Reasoning
-
Explicitly show intermediate reasoning steps before reaching a conclusion
- Reasoning Transparency
-
Make the thought process visible, not just the final answer
- Intermediate Representations
-
Break complex problems into smaller, manageable steps
- Error Reduction
-
Exposing reasoning allows detection of logical errors mid-process
- Complex Task Decomposition
-
Handle multi-step problems that cannot be solved in one jump
- Zero-Shot CoT
-
Simple prompt like "Let’s think step by step" to trigger CoT behavior
- Few-Shot CoT
-
Provide examples with reasoning chains to guide the model
- Self-Consistency
-
Generate multiple reasoning paths and select most consistent answer
- Key Proponents
-
Wei et al. (Google Research, 2022), "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models"
- Historical Context
-
Breakthrough in LLM prompting research, significantly improved performance on reasoning tasks (math, logic, common sense)
When to Use:
-
Complex reasoning problems (multi-step math, logic puzzles)
-
When you need to verify the reasoning process, not just the answer
-
Debugging incorrect LLM outputs by seeing where reasoning went wrong
-
Teaching or explaining complex topics where steps matter
-
Problems requiring planning or strategy
-
Any task where intermediate steps provide value
Related Research:
-
Tree of Thoughts (ToT): Extension allowing branching and backtracking
-
Self-Consistency: Sample multiple reasoning paths
-
Least-to-Most Prompting: Build up from simple to complex
Example Prompt Pattern:
Problem: [Complex question] Let's solve this step by step: 1. [First step] 2. [Second step] ... Therefore: [Conclusion]
Current Status:
-
The core finding (Wei et al., NeurIPS 2022) is stable, but its practical relevance is cutoff-bound: reasoning models (late 2024 onward) internalise chain-of-thought during training — OpenAI’s own docs now advise against "think step by step" prompts for those models
-
Treat CoT output as a useful but unreliable window into model reasoning: Turpin et al., "Language Models Don’t Always Say What They Think" (2023) and Anthropic’s 2025 follow-up show stated reasoning can be unfaithful to the actual decision factors — not a safety or audit guarantee