EduMirror: Modeling Educational Social Dynamics with Value-driven Multi-agent Simulation

EduMirror is a multi-agent simulation framework for the scientific study of educational social dynamics, presented at the 43rd International Conference on Machine Learning by Lin, Yu, Zeng and Zhong (Beijing Normal University and HKUST Guangzhou). The authors frame the problem as a dilemma: observational studies capture only static correlations and are prone to response bias, while controlled experiments such as randomized controlled trials are ethically constrained because withholding necessary interventions violates ethical standards. Their answer is an in silico Simulation environment built on Concordia and orchestrated by an LLM-driven Game Master, in which agents interact in natural language across a library of 20 pre-designed scenarios spanning four themes and eight virtual environments. Evaluated across seventeen scenarios, two case studies (bullying; peer cooperation and competition), human realism judgments, and five baseline agent architectures, the authors report dynamics they characterize as realistic, theory-consistent, and measurable by empirical criteria.

Synthesis: EduMirror introduces a multi-agent simulator for educational social dynamics, addressing the dilemma that observational studies lack causal power while controlled experiments face ethical constraints. The simulator features value-driven agents grounded in psychological needs and social value orientation, with a dual-track measurement protocol quantifying both observable behaviors and latent psychological states. Validation through case studies on school bullying and group cooperation, plus broader evaluations across diverse educational scenarios, shows the simulator generates realistic, theory-consistent dynamics measurable by empirical criteria. This enables structured in silico educational research for hypothesis testing and counterfactual intervention analysis. The approach connects to Learner Modeling and Adaptive Instruction by modeling latent psychological states and to Learning Analytics by providing a sandbox for testing interventions before deployment in real classrooms.

Key Findings

1. EduMirror ships 20 pre-designed scenarios across four themes — Peer & Group Dynamics (7), Individual Social Cognition (5), Classroom Culture (3), Home-School Dynamics (5) — instantiated in eight virtual environments, including the classroom, dormitory, playground, cafeteria, home, teacher's office, gymnasium, and library.
2. The Individual Value System covers five need categories (Safety, Mental Health, Self-Esteem, Social Belonging, Meaning and Growth) with 13 sub-dimensions scored on a 0–10 Likert scale; the Social Value System assigns a stable target orientation (Altruistic, Prosocial, Individualistic, Competitive) and computes a continuous orientation signal θt in [0, π/2].
3. Measurement splits into two tracks: an LLM Rater scores observable behaviors from completed traces, while an LLM Surveyor is applied post-hoc to logged internal states to administer psychometric questionnaires, so probing does not interfere with ongoing interaction.
4. Across seventeen evaluated scenarios, EduMirror achieved the highest average win rate under LLM-based post-hoc evaluation, and in a kindergarten scalability test it scored highest at every group size: 4.80 at 5 agents, 4.18 at 15 agents, and 4.03 at 30 agents across Naturalness, Coherence, Plausibility, and Developmental Typicality.
5. In the bullying case study, higher initial psychological values produced resilience while lower values increased volatility and accelerated victimization, and RSES scores gathered by the LLM Surveyor tracked that deterioration; across 20 bullying scenarios, intervention effectiveness increased from neglect to authoritative-punitive to supportive-individual to supportive-cooperative, the last improving all need dimensions.
6. In the peer interaction case study, ablating SVO weakened the distinction between personality profiles, and a slider-based SVO questionnaire closely matched agents' internal representations; Team Competition and Pre-Education produced the most stable competition outcomes, while the no-intervention control fluctuated most.
7. Human realism checks were favorable but small: against ten real bullying cases, 152 valid survey responses struggled to separate real from simulated (six groups below 30% accuracy; more than 10% chose "difficult to distinguish," peaking at 52.63% in Group 6), and a second study with 21 participants placed 14 of 20 sampled interaction cases in a high-consensus category (average agreement 89.2%) against 6 moderate (57.7%).

A value-driven cognitive architecture for educational agents

Most educational agent-based modeling relies on hand-crafted rules or Belief–Desire–Intention models — interpretable, but weak at nuanced and sometimes irrational social behavior. EduMirror constrains LLM generation inside an education-oriented cognitive architecture: agents carry configurable traits, goals, and formative memories, and behavior is generated as conditional sequence modeling over interaction history, profile, environment, and customized information such as goal instructions and desire states.

The planner is driven by unmet-need gaps — for each need dimension, a non-negative clipped distance between the expected and current value. Candidate actions are then scored for need-gap reduction, inferred effects on others, and consistency with the agent's SVO profile, and the highest-scoring option is selected. Because desire states and social orientation enter the prompt as explicit conditions, agent reasoning stays inspectable rather than hidden in an opaque policy. The design is modular: agent forms, decision logic, and measurement hooks are configurable, so new agents can be added without redesigning the scenario, intervention, or measurement pipeline.

Dual-track measurement and the evidence for realism

The paper's central methodological bet is that simulation is useful only if latent psychological states become measurable. The LLM Rater codes completed interaction traces for observable behavior; the LLM Surveyor is applied afterward to logged internal states, administering instruments such as the RSES for self-esteem or a slider-based SVO measure that estimates social preferences as a continuous angle. Agreement between these external questionnaire estimates and internal value representations is presented as a consistency check supporting psychological validity — with the caveat that the Surveyor measures dimensions already encoded in the agent's value system.

Realism was checked three ways: pairwise win rates across seventeen scenarios, a human study asking participants to distinguish ten real bullying cases from ten simulated ones, and a second study in which 21 participants rated 20 sampled peer-interaction cases. Baseline comparisons and an SVO ablation support the value-driven architecture specifically, and the authors report that GPT-4o judgments aligned closely with human evaluations.

Practice-relevant results from the two case studies

Case Study 1 targeted dynamically evolving needs. Victim agents simulated under diverse initializations showed that higher initial values produced resilience while lower values produced volatility, with teacher interventions compared across four conditions: neglect, authoritative-punitive, supportive-individual, and supportive-cooperative. Effectiveness rose along that ordering, and supportive-cooperative — collective action by peers, teachers, and families — produced the most significant improvement across all need dimensions.

Case Study 2 targeted stable traits, using scenarios of increasing social complexity: a small study group, a class-wide collaborative task under mild competition, and a class leadership election. Removing SVO weakened the differentiation of cooperation and competition, confirming that mechanism's contribution. In the election scenario, Team Competition, Teacher Reminder, and Pre-Education reduced extreme rivalry, and their lower variance across repeated simulations suggested a genuine balancing effect rather than random fluctuation, while the no-intervention control fluctuated most widely.

What this means for practice

• Instructors. The intervention comparison gives a concrete ordering for classroom response: neglect produced consistent decline in all needs, authoritative-punitive improved safety, belonging, and mental health only modestly, supportive-individual was inconsistent, and supportive-cooperative improved every dimension. The election findings point the same way — fairness framing, structured teamwork, and teacher facilitation were associated with more stable, less excessive competition.
• Researchers. The framework is offered as a computational laboratory for questions that are hard or unethical to test directly: save a simulation state, branch it, and compare counterfactual strategies under matched initial conditions, with dual-track measurement supplying outcomes on both behavior and latent state.
• Software developers. The value systems, measurement hooks, and intervention engine are decoupled from simulation infrastructure, so profiles, motivations, and decision logic can be reconfigured — and new agent forms added — without redesigning the scenario, intervention, or measurement pipeline.

Limitations

• All evidence comes from simulated LLM agents; no validation in real classrooms is reported. The authors describe the experiments as "snapshots of specific social situations" and propose longitudinal simulation across an entire school year as future work, noting that bullying, peer influence, and identity formation evolve over extended periods.
• The human realism checks use small, specific material: ten real bullying cases sourced from online news and interviews (rewritten in a standardized tone by GPT-4o) against ten simulated cases, judged by 152 survey responses, plus 21 participants rating 20 sampled interaction cases. The authors state that findings were generated using a specific LLM within scenarios inspired by a particular cultural context, and call for testing across different language models, cultural settings, and age groups.
• Scale is limited: simulations involved small groups, the scalability test topped out at 30 agents in a kindergarten scenario, and the authors call scaling to an entire school — including network effects and sub-group formation — a technical challenge.
• The questionnaire-based measurement is a consistency check rather than independent validation, since the LLM Surveyor measures dimensions already encoded in the agent's internal value system. Individual values (psychological needs) and social values (SVO) are modeled as parallel, selectable configurations rather than coupled and co-evolving, and agent cognitive processes such as memory consolidation and emotional regulation remain abstractions.

Citation

Lin, J., Yu, H., Zeng, Y., & Zhong, F. (2026). EduMirror: Modeling Educational Social Dynamics with Value-driven Multi-agent Simulation.