A tutoring app uses Claude to solve multi-step word problems, and it often gets intermediate arithmetic wrong. Which prompt change most directly targets this failure?
Show answer and explanation
Answer: B
Chain-of-thought prompting asks the model to reason through intermediate steps before committing to an answer. For multi-step arithmetic and word problems this tends to reduce compounding errors, because each step is conditioned on the ones before it. Higher temperature adds randomness that usually hurts exact computation, stop sequences only truncate output, and removing the problem details makes reasoning worse rather than better.
Why the other options are wrong
- A. Higher temperature increases sampling variability. For exact arithmetic that tends to add errors rather than uncover better solution paths, so it is the wrong lever.
- C. A stop sequence ends generation when a chosen string appears. Cutting output after the first number removes the working the model needs and does nothing for the calculations.
- D. Trimming the prompt to the final question removes the problem details and the reasoning scaffold. Shorter zero-shot prompts often reduce accuracy on multi-step tasks.