When the arithmetic plan goes wrong

A September 15 preprint by Zhongdi Qu and Carla P. Gomes examines why irrelevant wording can disrupt AI solutions to school mathematics problems.

The authors describe four internal stages: recognizing the problem structure, choosing operations, assigning numbers and calculating. Their experiments locate distractor-related failure in operation planning, rather than simply in arithmetic.

These are preprint findings about the studied models, not a universal account of mathematical reasoning. For AI tutoring, the study suggests a useful evaluation question: does a solution survive the addition of information that should not matter?

References

• Original preprint
• Paper PDF