Comparative Evaluation of ChatGPT, Gemini, and DeepSeek in Educational Problem Solving

Abstract

This study compares the performance of three large language models; ChatGPT, Gemini, and DeepSeek, on a set of programming-related educational problems from the Aizu Online Judge (AOJ) platform. The evaluation focuses on problem-solving accuracy and code characteristics, with additional comparisons to human Java submissions to contextualize model performance. Metrics include CPU time, memory usage, and code size, enabling a detailed analysis of solution quality and efficiency. Results indicate that ChatGPT consistently achieves the most efficient solutions while maintaining high accuracy, often matching the fastest human submissions. Gemini and DeepSeek also demonstrate strong accuracy but tend to produce less optimized code in computationally demanding cases. These findings contribute to understanding how current LLMs can address structured problem-solving tasks within educational environments.

Ähnliche Arbeiten