Simulating Students' Java Programming Errors with Large Language Models

This paper investigates the use of large language models (LLMs) to simulate realistic Java programming errors, leveraging a dataset of over 74,000 student submissions. Five LLMs were evaluated using three prompting strategies—Input-Output, Chain-of-Thought, and iterative Self-Refine—focusing on the diversity and alignment of generated errors with authentic student mistakes, with Claude Sonnet 4 demonstrating the most balanced performance. The findings indicate that while LLMs can produce diverse error patterns, their effectiveness varies based on the complexity of programming tasks, suggesting important considerations for their application in educational tools and learning analytics.