Skip to main content

Chain of Thought

The LLM is prompted to show its step-by-step reasoning before producing a final answer. Instead of jumping directly to a conclusion, the model works through intermediate steps — breaking down the problem, solving sub-parts, and synthesizing into a final result. This is single-pass reasoning with no tool use or external interaction.

Structure​

The entire chain happens in one LLM call. The model generates the reasoning trace and the answer in a single forward pass. No external tools, no loops, no iteration.

How It Works​

  1. Elicit reasoning — prompt includes instructions like "think step by step" or "show your work"
  2. Decompose — model breaks the problem into smaller parts
  3. Solve sequentially — each sub-part is addressed in order
  4. Verify — model checks intermediate results for consistency
  5. Conclude — final answer is synthesized from the reasoning chain

Variants:

  • Zero-shot CoT — just add "think step by step" to the prompt
  • Few-shot CoT — provide examples of step-by-step reasoning
  • Extended thinking — model uses a dedicated reasoning block before responding (Anthropic, OpenAI o-series)

Key Characteristics​

  • More accurate — especially for math, logic, and multi-step problems
  • Transparent — reasoning is visible and auditable
  • Self-correcting — model catches its own mistakes during the chain
  • Higher token usage — longer responses due to reasoning trace
  • Single pass — no tool use, no external grounding, no iteration

When to Use​

  • Mathematical or logical reasoning
  • Multi-step problems where errors compound
  • Explainability matters — you need to see how the answer was reached
  • Complex decision-making with multiple factors
  • Tasks where the model tends to jump to wrong conclusions without reasoning