Chain-of-Thought Emergence › Worked Example: Analyzing Chain-of-Thought Behavior

Let's examine in detail how chain-of-thought affects model outputs on a specific problem that requires multi-step reasoning. This concrete example will illustrate the principles we've discussed and reveal both the power and the mechanics of chain-of-thought prompting.

Problem: A bookstore sells notebooks for $3 each. Maria has $20. She wants to buy as many notebooks as she can and still have at least $5 left for lunch. How many notebooks can Maria buy?

This problem requires understanding a constraint (keeping money for lunch), computing the available budget, and then performing division. All these pieces must be correctly integrated. Let's see how different prompting approaches handle it.

Standard Prompting Response (hypothetical from smaller model):

6 notebooks

This answer is incorrect. The model appears to have divided 20 by 3, getting approximately 6.67, and rounded down without considering the lunch money constraint. The constraint was either ignored or lost in the compressed computation that produced the direct answer. We cannot know from this response why the model got it wrong. Did it not understand the constraint? Did it understand but fail to apply it? The opacity of direct answering makes diagnosis impossible.

Chain-of-Thought Response (from larger model):

Maria has $20 total. She needs to keep at least $5 for lunch. So she can spend at most $20 - $5 = $15 on notebooks. Each notebook costs $3. $15 ÷ $3 = 5 notebooks. Maria can buy 5 notebooks.

The chain-of-thought response correctly identifies the constraint (keeping $5 for lunch), computes the available budget ($15), and performs the division to get the correct answer (5 notebooks). Each step is visible and connected to the next, so a reader can verify the calculation.

What makes this work? The explicit intermediate steps force the model to stand for each part of the problem distinctly. The constraint that might be lost in direct computation is explicitly stated in the second line and used in the third. Each arithmetic operation is isolated and visible, making errors easier to catch and correct. The model's "thinking" is exposed to us and, perhaps more importantly, to itself. Each written step becomes part of the context for generating the next step.

This example also illustrates the debugging advantage of chain-of-thought. If a model produced an incorrect CoT response, we could examine the chain to identify where reasoning went wrong. Was the constraint understood but misapplied? Was the subtraction correct but the division wrong? This transparency is impossible with direct answers, and it is useful for understanding model behavior and improving prompts.