StudentSim

StudentSim is a research framework open-sourced on GitHub by members of Microsoft Research, released alongside a paper of the same name. The problem it tackles: finding out which guidance helps which student is slow and costly when you have to keep testing on real learners. Its answer is to train a "student simulator" for each student. A base version is first trained on records pooled across many students, then trained further on one student's own records, so the simulator carries that student's skill level and typical mistakes and changes its answer after tutoring the way that student did. The training material is real learning records from three domains: chess, English writing by second-language learners, and middle-school math. The chess data comes from real players' games on Lichess, and the writing and math data come from academic datasets you have to request separately. The base model is Qwen3-4B-Instruct-2507, and each student's simulator is a small LoRA adapter. For chess, the repo also shows how to use a simulator as a practice student and train an AI tutor that writes guidance, using reinforcement learning. What sets it apart is that it tests two things at once: whether the simulator behaves like the student (behavioral fidelity), and whether it corrects itself after tutoring the way that student would (guidance responsiveness).

Who it's for:researchers working on AI tutoring, teaching strategies and learner behavior. There's no ready-made product for students or teachers, so general readers will get limited direct use from it; it's more useful as a look at what this kind of research does.

Before you start: this is research code. It needs Python 3.10 or later plus dependencies such as PyTorch, Transformers and vLLM depending on what you run, training needs suitable compute, and you download the base model from Qwen yourself. Every step that calls a language model needs your own Azure OpenAI endpoint and key, which costs API fees. The training data comes from real people. The chess records come from Lichess's public CC0 games from May 2025, and the data notes say student IDs are random, with no way to map them back to accounts on the site. For writing and math, you request the data from its providers, accept their agreements and add it yourself; the repo doesn't redistribute it. The chess data notes say the upload hasn't landed yet and the download path is still a placeholder, so you can't yet reproduce it by following the README. Microsoft's Transparency Note says users must source data legally and ethically, including securing rights, consent or de-identification; that outputs are simulated responses and must not be used to assess, rank or make decisions about an identifiable learner; and that it isn't recommended for commercial or real-world use without further testing, or for high-risk decisions. The code is MIT licensed.

Based on the project's public documentation. Not tested by AI Inspiration World.