AIAny
Icon for item

StudentSim: Training LLM-based Student Simulators

Turns sparse per-student records into individualized simulators that both reproduce a student’s responses and update them under tutor guidance using pooled LLM pretraining followed by per-student specialization; releases StudentSimEval and reference simulators across chess, L2 writing, and math.

Introduction

Most tutor-training pipelines need two signals from learners: a realistic starting competence and a reliable way that competence changes under instruction. StudentSim reframes student simulation around this dual objective and shows that neither existing state-tracking models nor prompt-roleplayed LLMs satisfy both at once. The key move is a two-stage training recipe—pool domain data to learn general revision and mistake patterns, then specialize to each student’s sparse records—so each simulator mirrors a particular learner yet remains teachable.

Key Findings
  • Two concrete evaluation axes: behavioral fidelity (how well a simulator matches a student’s own responses) and guidance responsiveness (how readily the simulator updates toward tutor guidance). Framing simulation as these two metrics makes the target optimizable.
  • The pooled-then-specialize pipeline yields per-student simulators that outperform a strong prompted LLM baseline on both axes across three domains (chess, L2 English writing, mathematics), demonstrating that specialization is necessary to capture individual competence while retaining responsiveness.
  • A StudentSim-trained simulator used as a reward model for tutor reinforcement learning produces tutors rated by experts as more accurate, better-guided, and more personalized than tutors optimized against other simulators, showing practical utility beyond benchmarking.
Who it's for and tradeoffs

Great fit if you train or evaluate AI tutors and need scalable proxy learners that reflect individual mistakes and teachability. Use StudentSim when per-student personalization matters but real-student collection is too slow or costly. Look elsewhere if you only need population-level behavior (state-tracking models are simpler) or if legal/privacy constraints prevent using de-identified learner records, since the method depends on per-student training data. The framework emphasizes comparability (fixed per-student splits, shared metrics) over producing a single universal simulator.

Information

  • Websitearxiv.org
  • OrganizationsMicrosoft Research, University of Illinois Urbana-Champaign
  • AuthorsKe Yang, Chenglong Wang, Michel Galley, Chandan Singh, Jeevana Priya Inala, ChengXiang Zhai, Jianfeng Gao
  • Published date2026/09/01

More Items

Studies looping shared transformer layers in Mixture-of-Experts models under matched budgets and proposes SMELT: loop the middle half twice while matching per-token FLOPs, non-embedding parameters, and KV cache. Shows 6.8–18.0% training-FLOPs savings on the compute-optimal frontier, stronger downstream gains on code and long-context tasks.

Analyzes on-policy distillation for LLM fine-tuning, shows teacher token-level supervision is often noisy and not the main driver of gains, and introduces OPSA, a supervision-free, entropy-adaptive method that suppresses low-probability tokens to improve downstream accuracy.

A zero-data self-evolution framework that co-trains a Challenger, Solver, and Judge so LLMs can iteratively improve on both verifiable and unverifiable tasks without human labels. Uses role-asymmetry and subtask-amplification preference pairs to train the Judge and sustain improvement.