[Paper Review] ChatGPT Participates in a Computer Science Exam

Sebastian Bordt, Ulrike von Luxburg|arXiv (Cornell University)|Mar 8, 2023

Artificial Intelligence in Healthcare and Education31 citations

TL;DR

The paper reports an experimental evaluation where ChatGPT participated in an undergraduate CS exam, narrowly passing (20.5/40) with GPT-4 showing a 17% improvement to about 24 points, suggesting limitations in true understanding despite strong performance.

ABSTRACT

We asked ChatGPT to participate in an undergraduate computer science exam on ''Algorithms and Data Structures''. The program was evaluated on the entire exam as posed to the students. We hand-copied its answers onto an exam sheet, which was subsequently graded in a blind setup alongside those of 200 participating students. We find that ChatGPT narrowly passed the exam, obtaining 20.5 out of 40 points. This impressive performance indicates that ChatGPT can indeed succeed in challenging tasks like university exams. At the same time, the questions in our exam are structurally similar to those of other exams, solved homework problems, and teaching materials that can be found online and might have been part of ChatGPT's training data. Therefore, it would be inadequate to conclude from this experiment that ChatGPT has any understanding of computer science. We also assess the improvements brought by GPT-4. We find that GPT-4 would have obtained about 17\% more exam points than GPT-3.5, reaching the performance of the average student. The transcripts of our conversations with ChatGPT are available at \url{https://github.com/tml-tuebingen/chatgpt-algorithm-exam}, and the entire graded exam is in the appendix of this paper.

Motivation & Objective

• Evaluate whether ChatGPT can perform in a real undergraduate computer science exam in a blind grading setup alongside human students.
• Assess differences between GPT-3.5 and GPT-4 performance on a standardized algorithm and data structures exam.
• Explore limitations and interpretability of ChatGPT’s exam answers in a controlled educational context.
• Discuss implications for understanding AI capabilities in educational benchmarks.

Proposed method

• Use a standard undergraduate Algorithms and Data Structures exam posed in LaTeX text form.
• Pose 19 questions to ChatGPT across 19 conversations using simple prompts and no chain-of-thought prompting.
• Hand-copy ChatGPT’s answers onto an exam sheet to enable blind grading alongside 200 student papers.
• Have 10 teaching assistants grade the mixed set of papers in a single room under blind conditions.
• Compare GPT-3.5 results to GPT-4 results by repeating the process with the GPT-4 base model and non-blind grading.

Experimental results

Research questions

• RQ1 Can a modern language model pass a standardized undergraduate CS exam under blind grading conditions?
• RQ2 How does GPT-4 performance compare to GPT-3.5 on the same exam?
• RQ3 Which types of questions (e.g., proofs, pseudo-code, graphs) are most challenging for the model?
• RQ4 Does performance on this exam reflect genuine understanding or reliance on training data?
• RQ5 What are the limitations and potential implications for AI-assisted education?

Key findings

• ChatGPT earned 20.5 out of 40 points and narrowly passed the exam under blind grading for GPT-3.5.
• GPT-4 would have obtained about 17% more points than GPT-3.5, reaching the performance of the average student on this exam.
• GPT-4 showed improved performance on harder multiple-choice questions and certain graph-related tasks, but struggled with graph drawing and non-pseudo-code structured output.
• Some strengths were seen in pseudo-code and dynamic programming problems, while several questions involving standard algorithm explanations or graph structures remained challenging.
• Overall, ChatGPT’s success does not imply human-like understanding of computer science; questions resemble online teaching materials and solved problems that may have appeared in training data.