Teaching Language Models How to Code Like Learners: Conversational Serialization for Student Simulation

Researchers propose a method for training open-source language models to simulate how programming students learn and debug code, using authentic student data serialized into conversational formats. This approach addresses privacy and cost concerns with proprietary models while demonstrating improved performance in replicating student problem-solving behavior compared to existing baselines.

This research addresses a critical gap in educational AI by developing artificial learner models trained on real student behavior rather than relying on expensive proprietary systems. The innovation lies in converting temporal debugging sequences into conversational dialogues, allowing models to internalize the iterative nature of student learning—how learners respond to test failures, error messages, and feedback loops. This mirrors authentic educational experiences more closely than static code-only training approaches.

The broader context reflects growing concerns about educational institutions' dependence on closed-source language models, which create vendor lock-in and privacy risks around sensitive student data. By training smaller, open-weight models (4B and 8B parameters) on real programming submissions, the authors demonstrate that scale and proprietary access aren't prerequisites for effective educational simulation. This democratizes access to robust learner-simulation tools.

The market implications extend beyond academia. Educational technology platforms, tutoring systems, and programming assessment tools could adopt these methods to evaluate pedagogical strategies at scale without expensive API calls or privacy compromises. Open-source implementations reduce barriers for smaller ed-tech companies competing against well-funded incumbents with proprietary model access.

Looking ahead, this framework could inspire similar serialization approaches for other domains requiring temporal behavioral data—healthcare, customer support, professional development. The release of code and methodology positions this work as infrastructure for the emerging field of synthetic learner simulation. Adoption hinges on how effectively these models can generalize across different programming contexts and educational datasets.

• →Open-weight models trained on authentic student data outperform prompted proprietary LLMs for educational simulation tasks
• →Converting debugging sequences into conversational formats enables models to learn iterative problem-solving patterns
• →Approach reduces privacy risks and costs associated with relying on closed-source commercial language models
• →Smaller models (4B-8B parameters) achieve functional alignment comparable to larger systems when trained on domain-specific educational data
• →Open-source release enables reproducibility and broader adoption across educational technology platforms