ParaStudent: Generating and Evaluating Realistic Student Code by Teaching LLMs to Struggle
Mihran Miroyan, Rose Niousha, Joseph E. Gonzalez, Gireeja Ranade, Narges Norouzi
2025-07-22
Summary
This paper talks about ParaStudent, a system that trains large language models to generate and evaluate computer code in a way that mimics how real students write and improve their code, including making mistakes and fixing them.
What's the problem?
The problem is that existing AI code generators often produce perfect or too clean code, which doesn’t reflect how students actually learn and write code with errors and gradual progress, making it hard to study and improve educational tools.
What's the solution?
The authors taught the AI models to 'struggle' like students by intentionally generating code with common errors and showing step-by-step improvements. They also created evaluations that look at the style, mistakes, and how code evolves, making the generated student code more realistic.
Why it matters?
This matters because it helps create better educational software that can understand and respond to real student coding behaviors, improving programming learning and teaching methods.
Abstract
ParaStudent evaluates LLM-based code generation to mimic real student progress, capturing error patterns, incremental improvements, and stylistic variations through fine-tuning and multi-dimensional evaluation.