AI teachers are already capable of completing many individual teaching tasks.
A math problem? It can break down the steps. An English essay? It can point out grammatical errors. A chess game? It can suggest the next move. In many scenarios, the model’s output is already as helpful as a patient tutor.
However, teaching effectiveness cannot be measured solely from the teacher’s side. Whether students correct their mistakes after listening, whether they simply follow prompts, and whether the same explanation produces different results for different students—all of these occur on the student’s side.
An AI teacher also needs to repeatedly observe student responses to learn personalized instruction.
In reality, these responses primarily come from human studies. Each time the teaching strategy is adjusted, students must be organized to participate, interactions must be collected, and evaluations must be conducted manually.
AI models can be updated quickly, but the acquisition of educational feedback is limited by the duration of human study experiments.
The AI tutor requires student feedback to improve its teaching, but collecting genuine feedback is costly. StudentSim aims to convert real student records into reusable proxy feedback for the training phase.
Recent research on StudentSim, starting from this contradiction.
The research team aims to enhance the AI teacher so that it can understand students' strengths and weaknesses and provide tailored guidance for each individual. During development, the team focused on the student simulator: Can real student data be used to train an AI student that is evaluable and reusable, allowing the AI teacher to receive more feedback during training?
Paper link: https://arxiv.org/abs/2609.01591
Code link: https://github.com/microsoft/StudentSim
Hugging Face Paper homepage: https://huggingface.co/papers/2609.01591
This is not an isolated case in today's AI agent research. User simulators are becoming a popular research direction, applied in scenarios such as customer service, e-commerce, medical consultations, and web navigation. Researchers construct simulated users to expose agents to more interactions before deployment, testing their strategies, dialogue techniques, and robustness.
User simulations in educational scenarios are more complex. Students have relatively stable knowledge boundaries, habitual errors, and learning trajectories. For the same problem, some may miscalculate signs, others may misunderstand concepts, some can deduce the answer after a hint, while others require more direct guidance.
The digital twins of human students recently discussed are also centered around these types of needs.
Modeling students is not a new topic.
From Bayesian knowledge tracing in 1995 to deep knowledge tracing and then attention-based knowledge tracing, this direction has been explored by predecessors for nearly 30 years and is highly mature in modeling student behavior.
Their common limitation is that the models only accept structured inputs such as questions, states, and ability scores, with no interface for receiving natural language instructions. Regardless of what the teacher says, these models cannot interactively adjust their performance as a student would.
It may seem convenient to directly have a large model play the role of a student—simply telling it, “You are an elementary school student with weak math skills,” and the model immediately adopts the voice of a young child, showing hesitation and uncertainty.
However, the character's tone and level of understanding are not aligned.
A model can respond in a tone suitable for an elementary school student saying, "I don't really understand," yet follow up with calculus-level reasoning. It appears to be playing the role of a student, but its actual responses are still governed by its own knowledge base, which may far exceed the level of knowledge it is restricted to display.
During AI tutor training, this cognitive level bias can cause issues. The tutor receives responses not from a student at their current ability level, but from a high-capability model masked by a persona.
Precisely because of this, relying solely on a capability description as a condition creates a very fragile pathway for large models. This approach has been widely adopted in educational contexts over the past two years—used in conversational tutoring corpora, multi-agent classrooms, and learner data generation—while at the same time, a growing body of research has emerged to evaluate its validity, raising concerns from three perspectives: architecture, fidelity benchmarks, and teacher user experience.
In education, evaluating the effectiveness of guidance cannot rely solely on a student’s performance when working independently; it must also consider how far the student can progress with support.
Vygotsky’s zone of proximal development addresses exactly this: the gap between what a student can do independently and what they can achieve with teacher support reflects the space where instruction can effectively intervene.
This idea was later operationalized through dynamic assessment: rather than simply recording whether a student answered correctly, it also observes how their performance changes after providing prompts.
On the student simulator, this concept corresponds to two types of abilities.
First, the simulator must replicate the student’s state when answering independently, including their ability limits, error patterns, and common choices. Second, the simulator must adapt in response to teacher guidance, demonstrating how the student might improve with assistance.
StudentSim
StudentSim defines these two types of abilities as behavioral fidelity and guidance responsiveness, and uses them to train and evaluate personalized student simulators.
The student simulator must answer two questions simultaneously: Does it behave like the target student when answering independently, and does it demonstrate corresponding changes after receiving guidance from the teacher?
The training process consists of two steps.
Step one: Aggregate records from multiple students within a domain to train a general student behavior model. This phase learns common errors, answer formats, and correction pathways after guidance.
In the second step, continue training using each student’s individual data to obtain a personalized simulator. Since individual student data is limited, direct training is prone to overfitting; by first learning group-level patterns and then adapting to individual data, we can more stably derive a simulator for each student.
Training process
The research team also constructed StudentSimEval, which evaluates 60 real students across three domains: chess, second-language English writing, and basic mathematics.
In chess, the simulator predicts a player’s move and adjusts the position after coaching. In second-language writing, the simulator generates essays reflecting the learner’s error patterns and revises passages based on teacher feedback. In mathematics, the simulator predicts student answers and corrects them after instruction.
Although these tasks appear very different on the surface, the evaluation goal remains consistent: first, assess whether the simulator behaves like the student themselves, and then determine whether it can respond to guidance. All methods are evaluated using the same student records and compared on the same held-out test records.
In the chess results, StudentSim achieved an F of 0.5150 and an R of 0.9067; GPT-5.4 achieved 0.2316 and 0.7186, respectively; Maia2 achieved 0.4535 and 0.2721, respectively. In the second-language writing and mathematics experiments, StudentSim also outperformed the corresponding baselines on both metrics.
Two-dimensional results from chess evaluation. GPT-5.4 provides better adherence to instructions but has lower behavioral fidelity; Maia2 more closely matches human player moves but lacks a natural language instruction interface; StudentSim achieves high scores on both metrics.
This set of results shows a clear division of labor. GPT-5.4 can read instructions but lacks fidelity when simulating specific students. Maia2 more closely resembles human chess moves but cannot absorb natural language coaching cues. StudentSim, trained on real learning records and adapted for each individual student, improves both in modeling individual behavior and responding to guidance.
However, the student simulator itself is not the goal; the endpoint is that its functionality must be applicable in the real world and usable to improve teacher models.
Previously, training AI tutors relied on reward signals from expert-annotated high-quality tutoring dialogues in some cases, and from general large models acting as evaluators using scoring scales in others. Some efforts have attempted to connect rewards to simulated students, but these used LLM students engaged in cognitively inauthentic role-playing, rather than simulators trained on real learner data. StudentSim takes a different approach on this point.
The paper further integrates StudentSim into the AI tutor's reinforcement learning pipeline.
The experimental scenario is chess. The system first retrieves actual erroneous moves made by real students, the AI tutor generates guidance, and the StudentSim, after reading the guidance, provides a corrected move.
Stockfish is used to calculate the change in quality between the corrected move and the original erroneous move, and this score is fed back to the tutor as an RL signal. The control group includes a tutor without RL and a tutor using GPT-5.4 as a student simulator reward.
Tutor generation guidance: The frozen AI student simulator provides corrected answers, and the system updates the tutor based on the quality change before and after the correction.
Three tutor groups use the same base model, SFT starting point, and GRPO settings, but differ in reward sources.
After blind evaluation by chess assessors, the tutor trained with StudentSim reward ranked first in accuracy, quality of guidance, and personalized scoring.
Experts blindly evaluated three dimensions under randomized presentation order: accuracy based on the percentage of responses free from misleading factual errors, and guidance quality and personalization scored on a scale of 1 to 5. The tutor trained with StudentSim reward ranked first in all three dimensions.
More notably, a set of reverse experiments showed that the tutor trained using GPT-5.4 as a student simulator performed worse than StudentSim and even fell below the baseline with no RL at all, due to a significantly higher rate of severe factual errors. An inaccurate simulator misleads the teacher: a simulator with cognitive abilities mismatched to real students but exceptionally strong reasoning skills can generate plausible answers even to wildly incorrect instructions, thereby assigning random rewards to faulty guidance—causing RL to optimize toward chaos (or even error).
StudentSim does not replace real student experiments in educational research. Instead, it transforms real student data into simulated feedback that can be repeatedly accessed during the training phase, enabling the AI tutor to complete more rounds of screening and optimization before entering human studies.
For AI teacher systems requiring personalized feedback, these student simulators follow the recent trend of user simulators—the idea that anyone can be simulated to create digital twins of humans—offering a transformative engineering approach:
Learn how students make mistakes, learn how students change under guidance, and provide machine learning timescale feedback signals for the reinforcement learning of the teacher model.
Reference: https://arxiv.org/abs/2609.01591
This article is from the WeChat public account "New Intelligence Yuan," authored by New Intelligence Yuan; edited by LRST.