Chapter 9 · Talking to models: prompting and context2022
Chain of thought
Asking a model to work step by step often improves answers to multi-step problems.
Chapter 9 · Talking to models: prompting and context2022
Asking a model to work step by step often improves answers to multi-step problems.
Small steps, each easy to get right.
Writing the steps lets the model use its own words as a notepad.
You met next-token prediction. The model has no hidden scratch paper. Whatever it writes becomes part of the context for the next token.
Choose a puzzle. With one step the toy solver must leap. Add steps until it gets the right answer.
A bus has 45 people. At stop 1, 12 get off and 7 get on. At stop 2, 20 get off and 5 get on. How many are on the bus now?
Answer: 27 ✗ Wrong. The right answer is 25.
Teaching rule: the toy solver can do 2 operations per step. More than that and it slips. Real models are less tidy.
Why can writing steps help a model?
When does chain of thought help most?
Are written steps always right?