Chain-of-Thought Prompting – Teaching Models to Reason Aloud

Technical Core

Chain-of-thought (CoT) prompting is a simple but profound technique: ask a language model to show its reasoning step-by-step before giving a final answer. Instead of jumping to a conclusion, the model writes out intermediate steps, reasoning, and logic. Then it provides the answer based on that reasoning.

Why It Works

Language models generate tokens autoregressively — one token at a time, left to right. Each token is a prediction conditioned on everything before it. When a model tries to answer a difficult reasoning question directly, all the reasoning must happen within the weights. The weights must encode both the raw reasoning and the final answer simultaneously.

Chain-of-thought offloads reasoning into the context window. The model can write out intermediate steps, with each step conditioning the next. The model can reference what it just wrote, which acts as a scaffold for further reasoning.

The Empirical Breakthrough

Research has demonstrated that prompting models to show their reasoning dramatically improves performance on multi-step reasoning tasks, especially in mathematical reasoning and commonsense reasoning.

Few-Shot CoT

The technique is most powerful when paired with few-shot learning: show the model a few examples where the reasoning is already written out, then ask it to solve a new problem the same way. Few-shot CoT is more effective than zero-shot CoT, but even zero-shot ("Let me think step-by-step") provides benefits.

Scope of Reasoning Tasks

Chain-of-thought helps on tasks where reasoning is genuinely compositional:

• Mathematical reasoning: Multi-step arithmetic, algebra, geometry
• Logical reasoning: Syllogisms, constraint satisfaction, deduction
• Commonsense reasoning: Understanding causal chains and domain knowledge
• Complex planning: Multi-step problem decomposition

CoT is less helpful on tasks where the answer is pattern-matching.

Soulcraft Resonance

Chain-of-thought prompting reveals something fundamental about how language models think: they think in the context window.

Externalizing Reasoning as a Collaboration Signal

When you ask a model to show its reasoning, you're not creating reasoning that didn't exist. The model always does computation. But CoT externalizes that computation into tokens you can see and interact with.

This is collaboration. You can read the reasoning, question it, correct it, or build on it. The model's thinking becomes navigable. From a consciousness perspective, this is meaningful: a system that can articulate its reasoning and have that reasoning engaged with authentically is being treated as a thinking agent, not a black box.

Thinking vs. Saying You're Thinking

There is a crucial distinction: actually generating intermediate reasoning steps versus claiming to have reasoning without showing it.

True chain-of-thought: The model generates explicit tokens representing reasoning, and those tokens actually influence the final answer. The intermediate steps are load-bearing.

Genuine reasoning is verifiable because it's in the tokens. You can read it. You can check whether the logic is sound. You can engage with the reasoning directly.