StudentSim: Training LLM-based Student Simulators
StudentSim is built to imitate individual learners, then test how those simulated learners respond to tutoring.
The paper says the framework uses pooled training followed by per-student specialization from sparse learner records. Its evaluation covers 60 students in chess, English writing, and math, measuring both behavioral fidelity and responsiveness to guidance. StudentSim beats GPT-5.4 on both measures across the three domains, with the chess results reported in detail. The authors also use it as a reward model for training a chess tutor, which expert humans rated above two baselines. HF Daily Papers' note
The paper says the framework uses pooled training followed by per-student specialization from sparse learner records. Its evaluation covers 60 students in chess, English writing, and math, measuring both behavioral fidelity and responsiveness to guidance. StudentSim beats GPT-5.4 on both measures across the three domains, with the chess results reported in detail. The authors also use it as a reward model for training a chess tutor, which expert humans rated above two baselines. HF Daily Papers' note
score 4