acceptodds
Under review as a conference paper at ICLR 2027

StudentSim: Training LLM-based Student Simulators

Abstract

AI tutors are most useful when they adapt to each student's strengths, weaknesses, and preferred guidance, but feedback about which guidance works for whom is sparse, slow, and costly to collect. Student simulators can provide this feedback at machine timescales. Existing approaches capture only part of what tutor training requires: state-tracking models fit learner behavior but cannot respond to natural-language guidance, while prompted LLMs respond fluently but do not reliably reproduce an individual student's competence and characteristic errors. We present StudentSim, a training framework that turns sparse per-student records into individualized simulators through pooled domain training followed by per-student specialization. StudentSim combines structured behavioral grounding with a language-model interface, enabling each simulator to mirror a student's responses and revise them under tutor guidance. We also introduce StudentSimEval, a standardized protocol over 60 students in chess, second-language English writing, and mathematics. It evaluates behavioral fidelity (F↑), agreement between the simulator's responses and the real student's responses, and guidance responsiveness (R↑), whether the simulator reaches the response targeted by guidance. Across all three domains, StudentSim outperforms behavior-prediction and prompted-LLM baselines on both metrics. In chess, StudentSim reaches F = 0.51 and R = 0.91, compared with 0.23 and 0.72 for GPT-5.4 and 0.45 and 0.27 for Maia2. To demonstrate downstream use for AI tutor improvement, we use a trained StudentSim as the reward in tutor reinforcement learning. Expert chess players rate the resulting tutor as more accurate, better-guided, and more personalized than a no-RL baseline and a tutor trained with GPT-5.4 simulator feedback.

Primary area:
foundation or frontier models, including LLMs
Keywords:
user simulator, AI for education

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_xomDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper presents StudentSim, a framework for training individualized LLM-based student simulators through pooled domain training followed by per-student specialization. It introduces StudentSimEval, covering 60 students across chess, second-language English writing, and mathematics. Evaluation measures behavioral fidelity to recorded student responses and guidance responsiveness against canonical corrections. The paper reports improvements over prompted LLMs and domain-specific references. A chess proof of concept additionally uses the simulator within a composite reward for tutor reinforcement learning, with expert evaluations favoring the resulting tutor.

Critical strengths
  1. The paper addresses the challenge of simulating learner behavior while supporting natural-language instructional interactions. Evaluating both behavioral fidelity and instructional interaction provides a more comprehensive assessment of student simulation.
  2. The experiments span three domains, use chronological held-out records, and report results across three training runs. The inclusion of indirect guidance settings and blinded expert evaluation further strengthens the empirical evidence.
Critical weaknesses
  1. The RR metric measures agreement with canonical corrections rather than actual student responses to guidance. High FF and RR scores may therefore reflect general problem-solving ability rather than faithful, guidance-sensitive student simulation. Controls with absent, shuffled, or misleading guidance are needed to distinguish these possibilities.

  2. Appendix C.6 reports a substantial discrepancy (R=0.565R=0.565 vs. 0.9520.952) across greedy inference runs using the same mathematics checkpoint. The explanation based on bf16 nondeterminism remains unverified, and the reported stability on the full evaluation does not resolve this discrepancy. Controlled repeated inference experiments are needed to establish the reliability of these results.

Questions for authors
  1. What causes the same-checkpoint discrepancy (R=0.565R=0.565 vs. 0.9520.952)? Can the authors report repeated evaluations on both the subset and full mathematics set, including per-record consistency and sensitivity to numerical precision and batching?

  2. How do the pooled Stage-1 and personalized Stage-2 models compare on the same held-out records? How does RR change when guidance is removed or shuffled?

  3. In Table 25, multiplying a negative reward by a gate in (0,1](0,1] makes the reward less negative as the gate decreases. Does the implementation address this issue? If not, how frequently does it occur, and does correcting it affect downstream results?

Reviewer_4bdzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The main contribution of the paper is a two-stage training pipeline for building individual student simulators. It addresses the problem that knowledge tracing models capture student behavior but cannot generate text, so they cannot respond to tutor guidance. A LoRA adapter is first trained on records pooled across students and then specialized on each student’s own records. The authors propose two metrics for evaluating student simulators, behavioral fidelity and guidance responsiveness, and a benchmark of 60 students across chess, L2 English writing and math. They close the loop by using the student simulator as a reward for RL training of a chess tutor (a verifiable domain), and expert chess players rate the resulting tutor higher than the baselines. The paper is well written, and it is the first to train generative student simulators on real learner data and use them end-to-end for tutor training. However, some important details are hidden in the appendix, and the evaluation is mostly on chess. The baselines are weak, and single-step data does not support the claims about multi-turn student behavior. I would be willing to increase my scores if the baselines are improved, the multi-turn claims are reduced and the missing results and ablations are added.

Critical strengths
  • Important problem: realistic student simulators can unlock scalable training and benchmarking of AI tutoring models before deployment to real learners.
  • The paper closes the loop: instead of evaluating the simulator only in isolation, it uses the simulator to train a downstream tutor that expert chess players then evaluate.
  • Clearly written and grounded in real student records across three diverse domains (paired with templated or LLM-generated tutor guidance).
Critical weaknesses
  • Behavioral fidelity has to be hand-designed for each domain’s response space: move accuracy in chess, LanguageTool error density in L2 and 4-way multiple choice in math. This makes it harder to extend to new domains. Moreover, the L2 metric only matches error rates per category, not whether the simulator makes the same errors as the student.
  • The training data are not truly multi-turn. Every interaction has three steps: student answer → tutor guidance → student correction. How do you make sure the simulator generalizes beyond one tutor-student loop, for example to conversations where the tutor scaffolds over multiple reasoning steps?
    • The guidance is also not real tutoring data. For chess and math it is generated by an LLM, and for L2 it is templated (F.2), using the same generation pipeline for both training and testing.
    • Control experiments with no guidance or shuffled/irrelevant guidance are needed to verify that measures instruction-following rather than simply switching to "canonical solver mode" on the second turn.
  • Computational cost: there is one LoRA adapter per student. Training can be cheap, but how does storing, serving, and continually updating separate adapters for thousands of students compare in cost and latency to prompted or in-context simulators?
  • The baselines are weak and no baseline is fine-tuned on student data.
    • Most importantly, the Stage-1 model without per-student specialization is missing, important to directly show the per-student part is needed.
    • If the argument for two-stage training is that there is little data per student (L207–208), what about synthetic data generation seeded from real datasets? For example, LLM trained to predict exact student textual answers from knowledge tracing data.
  • The RL experiment uses the pooled Stage-1 simulator, not the per-student simulators (L442–444). It is not clear why the student simulator in the RL chess setup has extra heads, a board-state perception head and a personalization head. I would rather expect it to simulate several different student types, e.g., using the per-student simulators. Also, only the StudentSim condition gets these heads. The gains in Table 4 may come from the heads (heads measuring very similar measures as the final Table 4) rather than from the simulator.
Questions for authors

Question for authors:

  • L2 and math include only 15 students each, and the margins are small (e.g., math F 0.64 vs. 0.61). Please report per-student results with confidence intervals.
  • Inter-annotator agreement for the human evaluation should be reported
  • What is the performance of the Stage-1 pooled simulator (without Stage-2 specialization) across all three domains?
  • Idea: I encourage the authors to discuss and compare to general user models such as UserLM [1], which could serve as another student baseline.

[1] (UserLM) Flipping the Dialogue: Training and Evaluating User Language Models, ICLR 2026

Reviewer_kGzHOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
3(+2)
Presentation
2
Soundness
3(+2)
Read review
Paper summary

The paper trains a per-student simulator: a LoRA is trained on pooled records from a domain, then the same adapter is further trained on one student's records. F and R are evaluated on 60 real students in chess, second-language writing, and mathematics.

Critical strengths
  1. Sixty real students across three domains with frozen per-student splits is a genuine artifact.
  2. There is a downstream application: an RL experiment that trains a tutor with simulator feedback.
Critical weaknesses

There is a serious problem with the writing: the paper reads more like a system report than a scientific study. What it lacks is analysis of the results. For example, why do F and R take the values they do? The main text does not explain this.

  1. Most of R does not come from modelling the student. In Table 3, prompted GPT-4o reaches chess R = 0.7655 against the method's 0.9067, and untrained Qwen reaches 0.6132 on math against 0.9181. Neither baseline has a per-student adapter, so much of R is explained by reading the endpoint out of the guidance.

  2. The method itself is not new. Training Stage 1 and then continuing the same adapter in Stage 2 is the standard per-user adapter recipe. The paper cites MAML and MeLU but compares against no personalization method. The baselines are a population model (Maia2), prompted GPT-4o and GPT-5.4, and untrained Qwen, so the claim that specialization helps holds only against models without specialization.

Questions for authors
  1. Can you report per-student results with confidence intervals?
  2. What is the per-mode R for the baselines?
  3. What does Stage-1-only give in each domain?
10 comments

Thank you for your review. We are glad you recognize the sixty real students across three domains, each with a frozen per-student split, and the downstream application, "an RL experiment that trains a tutor with simulator feedback". These are the two ends the paper connects: a simulator learned from each student's own records, and its use as feedback for training an AI tutor. We have uploaded a revised manuscript, in which changes from the submitted version are marked in purple. Turning to the comments you raise, we list in the table below where each is addressed, in the version you reviewed or in this revision; the details follow.

CommentVersion you reviewedAdded in this revision
Analysis of F and RSection 1, p. 2; Sections 5.2–5.3, pp. 6–8
W1. Guidance and RSection 5.3, p. 7; Table 3
W2. Method and baselinesSection 4, p. 5; Section 5.1, pp. 5–6
Q1. Per-student resultsAppendix H.3
Q2. Per-mode RAppendix C.7, Tables 8–10
Q3. Stage-1-onlyThis reply
Ethics flagEthics Statement, p. 10

Analysis of F and R

Why F and R take the values they do has been analyzed in the main text since the original submission. For example, Section 1 (p. 2) explains why the models differ in all three domains: "prompted GPT-5.4 responds to guidance but poorly reproduces individual behavior, while Maia2 models human moves but cannot consume natural-language guidance. StudentSim is strong on both axes, and the same pattern holds in L2 and mathematics." Sections 5.2 and 5.3 (pp. 6–8) state what a higher F and R mean in each domain and trace these differences on held-out chess cases (Figures 4 and 5). Appendix C.7 gives R under each guidance mode for every model that takes guidance.

W1. Guidance and R

The comparison in your comment already shows that R is not explained by reading the endpoint out of the guidance. The baselines you cite read the same guidance as StudentSim, and untrained Qwen is the same base model as StudentSim, so if reading the endpoint explained R, they would reach the same R. StudentSim is higher in every domain (Table 3), and the gap is largest under the guidance types that do not state the endpoint (Section 5.3, p. 7; Appendix C.7).

W2. Method and baselines

We study a new and important problem: simulating how students learn from tutoring, so that the simulator can help AI tutors improve. Thus, StudentSim is new in its training target and in the training mechanisms built for that target. A student simulator must simulate both a student's current knowledge and how that student's action would change after tutoring, a part of learning dynamics, which calls for a novel training design different from typical personalization. Reaching this target with sparse per-student records is itself a persistent difficulty in student modeling: StudentSim is, as Reviewer 4bdz notes, "the first to train generative student simulators on real learner data", and its two-stage training is one part of the method framework that makes this possible. Prior work has found that learning performance data "tend to be highly sparse (80%~90% missing observations) in most real-world applications", and that this sparsity "presents challenges to using learner models to effectively predict future performance" (Zhang et al., 2025); a survey of knowledge tracing concludes that "there remains a need for further improvement in addressing this issue comprehensively" (Shen et al., 2024). Additionally, StudentSim includes other components for addressing the core problems of student simulation, described in the main text since the original submission. For example, to keep the simulator from answering above the level of the student it models, a known failure of LLM-simulated students (Yuan et al., 2026; Section 4, p. 5): "A structured state summarizes prior behavior, … Training aligns expressed competence with the state evidence …"; and to simulate how the student responds to a tutor at different levels of support (Wood et al., 1976; Section 5.1, p. 5): "Effective tutoring adjusts the level of support to what the learner needs at that moment, … Each domain therefore includes guidance types that span that range of support …"

Every baseline is already conditioned on each student, in the form that each baseline can accept, as described in the main text since the original submission (Section 5.1, p. 6): GPT-4o and GPT-5.4 are "prompted with each student's profile, problem, and guidance", untrained Qwen is "prompted identically", and Maia2 is "conditioned on FEN and player ELO". The comparison is therefore among models that all condition on each student, which differ in how they do so. MeLU, a cold-start recommender, and MAML, a general meta-learning algorithm, are cited in Section 4 (p. 4) only for two facts about sparse per-user records: such records are common in interaction-based systems, and training a model for each student on so few records overfits. Neither belongs to the line of work on student simulation, and neither generates a student's response to a tutor's guidance, which F and R evaluate, so neither can serve as a baseline.

Q1. Per-student results

We have added per-student F and R for every model, each with a 95% interval over that student's own held-out records (Appendix H.3, Tables 29–34). Against GPT-5.4, StudentSim is higher for 53 of the 60 students on F and 54 of 60 on R.

Q2. Per-mode R for the baselines

This has been reported since the original submission: Appendix C.7 (Tables 8–10) gives R under each guidance mode for every baseline that takes guidance, in all three domains.

Q3. Stage-1-only

Personalization in StudentSim begins in Stage 1: in both stages every input carries the student's structured state (Section 4; Appendix B), so the Stage-1-only model is already a per-student simulator. It is already higher on F than every baseline in Table 2, in every domain: 0.5163 in chess, 0.5195 in L2 and 0.6364 in math, against at most 0.4535, 0.5141 and 0.6121. Student records grow over time, and Stage 2 is the interface that takes in these new records and keeps the simulator fitted to the student. Its effect grows with the student's own records: in one L2 training run (seed 42), Stage 2 with 0%, 25% and 100% of a learner's records in chronological order gives F of 0.5165, 0.5233 and 0.5628.

Ethics flag

The Ethics Statement (p. 10) covers the terms of use of the three data sources, no redistribution of restricted records, pseudonyms in reported examples, the risks of per-student models and the scope we set for them, and IRB approval with informed consent for the expert study. Please specify the concern behind the flag, so that we can respond to it.

In this revision, we have added per-student results (Q1); the other points are addressed at the locations given above. We remain available to clarify any of these points.

References

  • Shen, S., et al. (2024). A Survey of Knowledge Tracing: Models, Variants, and Applications. arXiv:2105.15106.
  • Wood, D., et al. (1976). The Role of Tutoring in Problem Solving. Journal of Child Psychology and Psychiatry.
  • Yuan, Z., et al. (2026). Towards Valid Student Simulation with Large Language Models. arXiv:2601.05473.
  • Zhang, L., et al. (2025). Data Augmentation for Sparse Multidimensional Learning Performance Data Using Generative AI. arXiv:2409.15631.
Reviewer_fPQpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes StudentSim, a framework for building personalized student simulators using a two-stage training approach. It first trains a base model on pooled student data and then adapts it to individual students through LoRA fine-tuning. The authors introduce StudentSimEval to evaluate behavioral fidelity and guidance responsiveness across chess, English writing, and mathematics. They also explore using the trained simulator as a reward model for improving an AI tutor through reinforcement learning.

Critical strengths
  1. The paper addresses an interesting and practical problem. Considering both behavioral fidelity and guidance responsiveness is useful for evaluating student simulators in interactive teaching scenarios.

  2. The experiments cover three different domains and use real student records. The proposed method consistently outperforms the selected baselines on both evaluation metrics.

  3. The downstream tutor RL experiment is a nice addition. It shows how student simulators could be used beyond behavior prediction and provides some evidence of their potential value for tutor training.

Critical weaknesses
  1. The methodological novelty is somewhat limited. The main approach is pooled training followed by student-specific LoRA fine-tuning, which is a fairly standard personalization strategy. The paper would benefit from a clearer explanation of what is technically new beyond applying this strategy to student simulation.

  2. The evaluation of guidance responsiveness is not entirely convincing. The R metric measures whether the simulator reaches the correct answer after receiving guidance, rather than whether it responds as the actual student would. This makes it difficult to distinguish student-specific responses from general instruction-following or correction ability, especially since the guidance in chess and mathematics is generated toward predefined answers.

  3. The experiments do not fully establish the benefit of student-specific training. The baselines mainly consist of prompted LLMs and non-personalized models. A direct comparison between the pooled Stage-1 model and the personalized Stage-2 models across all three domains is missing. Such a comparison would help clarify how much improvement actually comes from personalization.

  4. The downstream RL results are difficult to interpret. The RL experiment uses the pooled simulator rather than the personalized models. In addition, the StudentSim reward includes extra personalization and perception heads that are not used in the GPT-5.4 baseline. It is therefore unclear how much of the improvement comes from the simulator itself versus these additional reward components.

  5. There are some concerns about experimental stability. Appendix C.6 reports substantially different responsiveness scores (0.565 and 0.952) when evaluating the same mathematics checkpoint. This discrepancy deserves further investigation, particularly given the relatively small performance gaps on some tasks.

Questions for authors
  1. Could the authors report the performance of the Stage-1 model without student-specific adaptation across all three domains?

  2. How does guidance responsiveness change when the guidance is removed or replaced with irrelevant guidance?

  3. Could the authors provide an ablation of the additional reward heads in the tutor RL experiment?

  4. What explains the large difference between the two inference results reported in Appendix C.6?

Comments on the paper

1 comment

Dear Reviewers,

The authors have posted their responses to the reviews. Please go through their rebuttal at your earliest convenience, and engage in a discussion or leave any follow-up comments.

The discussion period ends on November 18.

Thanks,

Area Chair

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.