Research map / Multi-agent systems and games / LLM agents
EduMirror: Modeling Educational Social Dynamics with Value-driven Multi-agent Simulation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces EduMirror, a Concordia-based LLM multi-agent simulator with value-driven (psychological needs + SVO) agents, dual-track LLM Rater/Surveyor measurement, and intervention branching for studying educational social dynamics such as bullying and cooperation.
Problem
Studying educational social dynamics is hard: observational studies lack causal power and experiments are ethically constrained, while existing LLM-based social simulations have weak psychological grounding and poor measurement of latent psychological states.
Approach
Value-driven LLM agents whose behavior is guided by an individual psychological-needs system (Maslow/PERMA-based unmet-need gaps) and a social value orientation angle computed from self vs other need-satisfaction signals, paired with a dual-track measurement protocol (LLM Rater for behaviors, LLM Surveyor for psychometric latent states) and counterfactual intervention branching.
Key results
- In the preschool scalability test, EduMirror achieved the highest average score of 4.80, 4.18, and 4.03 for 5, 15, and 30 agents, versus the next best baseline LLMob at 4.25, 3.60, and 3.83.
- In a human study with 152 respondents, participants scored below 30% accuracy in distinguishing real from simulated bullying cases for six of the groups.
- In bullying interventions, supportive-cooperative teacher strategy produced the largest improvement across psychological need dimensions versus ignoring, punitive, and individual-support strategies.
Surprising findings
- Authoritative-punitive intervention had limited or negative effects on self-esteem and meaning in the bullying simulation.
Limitations
- Only one LLM and culturally specific scenarios were tested, so generalizability across models, cultures, and age groups is untested.
- Simulations involve small groups; scaling to whole-school network effects is an open technical challenge.
- Individual and social value systems are parallel configurations rather than jointly coupled.
Experimental setup
Datasets: ten real bullying cases from online news and interviews (for human evaluation)
Metrics: win rate naturalness coherence plausibility developmental typicality human identification accuracy SVO angle alignment
Baselines: ReAct BabyAGI LLMob JAG-Concordia D2A
Builds on: Concordia D2A Maslow's Hierarchy of Needs Social Value Orientation
Closest papers by meaning
- Simulating Society Requires Simulating Thought (NeurIPS 2025 · similarity 0.86)
- iLLuMinaTE: An LLM-XAI Framework Leveraging Social Science Explanation Theories Towards Actionable Student Performance Feedback (AAAI 2025 · similarity 0.86)
- BEAGLE: Behavior-Enforced Agent for Grounded Learner Emulation (NeurIPS 2026 · similarity 0.85)
- Investigating Prosocial Behavior Theory in LLM Agents Under Policy-Induced Inequities (AAAI 2026 · similarity 0.85)
- Position: LLM Social Simulations Are a Promising Research Method (ICML 2025 · similarity 0.84)
- TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.