Think before answering
Chain of thought prompting asks the model to write its reasoning steps before the final answer. On arithmetic, logic, and multi step tasks this often raises accuracy, because the model uses generated tokens as a scratchpad instead of guessing in one leap.
Two ways to trigger it
- Few shot chain of thought shows examples whose answers include worked reasoning, so the model imitates the style.
- Zero shot chain of thought simply appends a cue such as let us think step by step, which nudges the model to lay out steps without any example.
Why it works
Each generated step conditions the next, so the model spreads computation across many tokens. Hard problems that fail in a single forward pass become tractable when broken into smaller stated moves.
Costs and cautions
Reasoning text uses more tokens and time. The stated steps are a story the model tells, not a guaranteed trace of its true computation, so a confident wrong chain can still mislead. For final delivery you often hide or strip the reasoning.
Key idea
Chain of thought makes the model write intermediate steps so it spreads computation across tokens, lifting multi step accuracy at the cost of more tokens and reasoning that may not reflect true computation.