All learning paths Learn See Try Make Grade a maths answer step by step What this lesson is about Mark the first step that is wrong, say why, and leave the steps after it alone.
A student shows you five steps of a speed sum and the last number matches the textbook. What do you do?
A Check the last line against the textbook, and pass it because the two match. B Mark the FIRST wrong step, not the wrong final answer. C Point out every place the working slips, so he sees the whole list at once.
Check my recall Why this matters A 2023 paper compared feedback on the final answer with feedback on every step. Step feedback trained the better model. Your review works the same way.
You can grade a written solution step by step and name the one step where it first goes wrong.
A 2023 paper compared feedback on the final answer with feedback on every step. Step feedback trained the better model. Your review works the same way.
Mark the FIRST wrong step, not the wrong final answer. A right final answer reached through a wrong step is still wrong. Write the step number, the rule it broke, and the step it should have been. Task A five-step solution to a speed and distance question ends with the right number.
Weak approach Checking the arithmetic in all five steps, finding every sum correct, and marking it right. A sum can be correct inside a step that should never have been taken.
Stronger approach Marking step two: hours were added to kilometres per hour. Steps three to five only carry that mistake forward.
Why the stronger approach works The final answer was right by luck. Naming the first wrong step is the part the author can act on.
Try a changed situation A mason's estimate for a boundary wall reaches the figure you expected. But the brick count came from the wrong wall length. What do you say?
Make something yourself Grade one AI-written maths solution of five steps or more. Name the first wrong step, the rule it broke, and the step that should have been there instead.
Lesson 9 of 9 on Reasoning and maths for AI work next Follow a small program line by line Part 3 ends in a work sample you can send The rules this work sample is read against, with a worked example
Read the primary or official source: arXiv · Let's Verify Step by Step, Lightman and others, 2023