LLM Performance Assessment in Computer Science Graduate Entrance Exams

Large Language Models (LLMs) are increasingly utilized in educational settings, raising questions about their efficacy in standardized testing contexts. This study evaluates the performance of popular LLMs, including GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Flash, Llama-3.1-70B, Mistral Large 2, DeepSeek-V2, and Gemma-2-27B, in answering data structure and algorithm design questions from the Iranian university entrance exams for master’s programs in computer science and information technology. The research analyzes the accuracy of responses, the length of answers, required reading time, and vocabulary levels across the 2022,2023, and 2024 exams. Questions were categorized into five main areas: Algorithm Analysis and Complexity, Sorting and Searching, Graph Algorithms, Data Structures, and Advanced Topics and NP-completeness. The study compares the LLMs’ performance in both Persian and English contexts, providing a comprehensive evaluation of their capabilities and limitations in this domain. Results indicate that GPT-4o achieved the highest average accuracy ($\mathbf{7 5. 0 \%}$), followed by Claude 3.5 Sonnet (67.2%) and Mistral Large 2 (64.1%). The majority of the models performed better on English questions compared to Persian, with GPT-4o showing the largest performance gap (81.3% vs $\mathbf{6 8. 8 \%}$). The findings offer insights into the practical applications of LLMs in educational assessments and contribute to the ongoing discourse on the role of AI in academia, particularly in non-English speaking regions.