Chain-of-Thought - Chain-of-Thought Reasoning

Basic Information

• Type: LLM Reasoning Technique / Prompt Engineering Paradigm
• Paper: "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models" (2022)
• Authors: Jason Wei et al. (Google Research)
• Publication: NeurIPS 2022
• Citations: Extremely high, becoming foundational work in the field of LLM reasoning

Paradigm Description

Chain-of-Thought (CoT) is a prompting technique that significantly enhances the complex reasoning capabilities of large language models by generating intermediate reasoning steps. By including step-by-step reasoning examples in the prompt (Few-shot CoT) or simply adding "Let's think step by step" (Zero-shot CoT), LLMs can decompose complex problems into manageable intermediate steps for reasoning.

Core Variants

• Zero-shot CoT: Simply adding "Let's think step by step" activates reasoning
• Manual CoT (Few-shot): Manually crafting step-by-step reasoning examples as prompts
• Automatic CoT (Auto-CoT): Automatically generating reasoning chain examples
• Multimodal CoT: Cross-modal (text + image) chain-of-thought reasoning

Technical Evolution (2025-2026)

• Effect Controversy: Latest research shows CoT effectiveness varies by model type and task
• Non-reasoning models: Average improvement is small and increases answer variability
• Reasoning models: Only marginal gains but increases time cost by 20-80%
• Generalization Questioned: 2025 research indicates CoT reasoning is a "fragile mirage" outside the training distribution
• Hallucination Trade-off: CoT reduces hallucination frequency but obscures detection signals
• Diminishing Value: Wharton research report indicates the value of CoT prompts is diminishing

Advanced Reasoning Paradigm Evolution

• Chain-of-Thought: Basic linear reasoning chain
• Tree-of-Thought: Tree structure exploring multiple reasoning paths
• Graph-of-Thought: Graph-structured reasoning network
• Training-Driven Methods: RLHF, process reward models, self-teaching reasoning
• Multi-Agent Systems: Achieving reasoning through decomposition and collaborative error correction

Applicable Scenarios

• Mathematical reasoning and arithmetic problems
• Commonsense reasoning tasks
• Symbolic reasoning tasks
• Complex multi-step problem solving

Relationship with OpenClaw Ecosystem

Chain-of-Thought is the foundational technology for the reasoning capabilities of OpenClaw personal AI agents. Agents should automatically enable chain-of-thought reasoning when handling complex tasks, breaking down large problems into smaller steps. However, based on the latest research, OpenClaw should dynamically choose whether to enable CoT based on task type and model capabilities, avoiding unnecessary delays and costs on simple tasks.