If you ask a human: "What is 37 multiplied by 48?" and demand they yell out the answer in half a second without using paper or mental scratchpads, they will probably guess wrong.
Yet people do this to AI models every day! They write: "Calculate the net profit across these 8 stores after discounts, tax, and inventory shrinkage. Give me the final number."
When you force an AI to answer in one shot, it has to guess the next token instantly. But when you allow it to think step-by-step, its accuracy skyrockets by up to 40%!
The Scratchpad Token Secret
Language models don't have a hidden calculator running in the background. Their thinking happens through the tokens they write.
Every token the AI outputs is added to its context window. When the AI writes out: "Step 1: Gross revenue = $100,000. Step 2: 10% discount = $10,000, leaving $90,000...", it creates a verified mathematical breadcrumb trail that anchors the final calculation.
The 3 Ways to Trigger Chain-of-Thought
- 1. The Zero-Shot Magic Phrase: "Let's work through this step by step to ensure total arithmetic accuracy."
- 2. Tagged Scratchpad Blocks: "First, perform all calculations inside a <scratchpad> block. Then, output the final number inside <final_result>."
- 3. Explicit Step Deconstruction: Break down the problem into sub-questions (e.g. "1. Calculate A. 2. Calculate B. 3. Subtract B from A.").