This paper examines knowledge distillation, a technique where a large, powerful model teaches a smaller model to perform a task. The authors use DeepSeek-R1, a large reasoning model, as the teacher, and Qwen2.5-7B, a compact student model, as the learner. The training data comes from past problems in the John O'Bryan Mathematics Competition at Northern Kentucky University.

The study is framed as a case study, meaning it focuses on a specific instance rather than a broad survey. Because the abstract does not include numerical results, the main contribution at this stage is the experimental setup: pairing a state-of-the-art reasoning model with a much smaller student and evaluating on challenging competition mathematics.

As the only source, this paper offers no comparison with other studies. Its significance lies in testing whether distillation can make compact models competitive on hard reasoning tasks, which would have practical implications for deploying capable models with limited computational resources.