Challenging Deepseek-R1 with Serbian High School Math Competition Problems

Abstract

This paper investigates the application of the large language model DeepSeekR1 in solving problems from the Serbian National athematics Competition for high school students (2023/2024), with the aim of examining its ability to understand, reason, and generate accurate mathematical solutions. The analysis covers all grade levels and both competition categories (A and B), comprising a total of 36 problems across various mathematical domains, including algebra, geometry, combinatorics, number theory, and logic. All problems were presented to the model in their original textual form, using LaTeX syntax to ensure accurate representation of mathematical expressions. The results obtained by the model were compared to official grading criteria and to the average scores achieved by student participants. In several cases, DeepSeek-R1 reached scores that would have qualified for official awards, especially in the higher grades of category B. The model showed stronger performance on algebraic problems and those with a more formal structure, while it encountered difficulties with logic-based problems and less standard formulations. Most errors were due to occasional misinterpretation of the problem statements or the omission of key reasoning steps. This research provides a realistic assessment of the capabilities of a contemporary large language model in solving complex mathematical problems. It also highlights possible directions for its integration as a support tool in the teaching and learning of mathematics.