On this page

Synthesis: Semi-structured interviews with 12 teachers who tutored LLM-simulated students (MathDial dataset) reveal key authenticity gaps: overly complex language, lack of emotions, unnatural attentiveness, and logical inconsistency. The study categorizes four real-world student behavior types along scaffolding and presence dimensions, and provides design guidelines for building higher-fidelity LLM student simulations.

Methodology

Martynova et al. interviewed 12 teachers who had extensively interacted with LLM-simulated students during collection of the MathDial dialogue tutoring dataset. The study used a mixed-method approach grounded in two frameworks:

Teachers tutored LLM students in K-12 math Problem Solving dialogues, then rated realism and described deviations from authentic student behavior.

Key Findings

Authenticity Gaps in LLM Students

Issue Description
Language complexity Responses too technical, lengthy, and formal for K-12 students
Emotional absence Lack of frustration, fear, embarrassment, or disengagement
Unnatural attentiveness Students too engaged; never lose focus or go silent
Logical inconsistency Knowledge jumps without gradual building; no forgetting
No question-asking Teachers had too much control over discussion flow

Four Student Behavior Categories

The study classifies real-world student behaviors along two dimensions:

High Scaffolding Needs Low Scaffolding Needs
Social Presence Short/simple writing, negative emotions, disengagement Asking questions, disagreeing with teacher
Cognitive Presence Gradual knowledge-building, memory/forgetting Changing tactics based on feedback

LLMs captured the bottom-right quadrant reasonably well but failed to represent the other three categories.

Design Guidelines

  1. Diverse personalities — model Big Five personality traits to produce varied engagement levels and emotional responses
  2. Gradual knowledge building — integrate knowledge tracing to avoid unrealistic knowledge jumps
  3. Model forgetting — account for memory decay over time
  4. Promote question-asking — use context-aware triggers for the LLM student to ask questions
  5. Vary language complexity — regulate response length, formality, and introduce age-appropriate errors
  6. Allow disengagement — let simulated students lose focus or stay silent, providing authentic teaching challenges

Significance

  • Teacher training: more realistic LLM student simulations enable scalable practice for pre-service and in-service teachers
  • Validation gap: only 3% of studies simulating learners do post-factum validation — this study provides a framework for it
  • MathDial is the only publicly available dataset of real teacher/LLM-student interactions
  • Addresses the growing trend of using unvalidated LLM simulations in educational contexts

onal contexts

What this means for practice

  • Teacher educators. Do not run LLM-simulated students as authentic rehearsal without checking fidelity first; the 12 tutoring teachers reported overly complex language, missing emotion, unnatural attentiveness, and knowledge jumps, and the simulations only represented the low-scaffolding/high-cognitive-presence quadrant well.
  • Teacher educators. Screen a simulation against all four real student behavior categories — disengagement and negative emotion, question-asking and disagreement, gradual knowledge building and memory, and tactic change on feedback — and choose prompts that force the three categories the LLM students failed to cover.
  • Researchers. Validate simulations against teacher judgment after use rather than assuming internal consistency; only 3% of studies simulating learners do any post-factum validation, which is the gap this framework addresses.
  • Teacher educators. Use the scaffolding-by-presence categorization as a debrief checklist: it tells you how a simulated student should push back or disengage, so a trainee's response can be reviewed against a defined behavior type.
  • Designers. Build knowledge tracing and forgetting into the simulated student, and let it go silent or lose focus, so that trainees meet the teaching challenge the design guidelines describe rather than a relentlessly cooperative partner.

Limitations

  • The study interviewed 12 teachers, most based in the UK and the majority female; the authors acknowledge possible gender effects on how personas are perceived and call for larger, more diverse samples.
  • All simulations analyzed came from one dataset (MathDial) built with an older GPT-3.5-turbo model, so behavior may differ with other LLMs — and the authors note more capable models can sometimes simulate less convincingly.
  • Scope was limited to K–12 mathematics dialogue; real student behavior varies by subject and was never tested elsewhere.
  • Realism judgments are teacher self-report in interviews, with no measured learning outcomes and no comparison against real students.

Citation

Martynova, D., Macina, J., Daheim, N., Yalçın, Ö. N., Zhang, X., & Sachan, M. (2026). Can LLMs Effectively Simulate Human Learners? Teachers' Insights from Tutoring LLM Students.

Embed this page

Copy the code below to embed a chromeless version of this page in a learning management system or other website. The embedded view hides the site header, navigation, and footer.