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

ICLR 2026 Conference Submission5807 Authors

Keywords: Educational Simulation, Multi-Agent Simulation, Computational Social Science

TL;DR: A flexible multi-agent simulation for styding educational social dynamics.

Abstract: The scientific study of educational social dynamics, such as bullying and peer pressure, is crucial for student well-being yet hindered by profound ethical and methodological barriers inherent in traditional research. While multi-agent simulations powered by Large Language Models (LLMs) provide an ethically viable alternative, they often fail to bridge the gap from believable narratives to rigorous experiments, plagued by two fundamental hurdles: a lack of psychologically plausible motivations (the Fidelity Challenge) and the absence of systematic methods for quantifying complex interactions (the Measurement Challenge). To overcome these obstacles, we introduce $\textbf{EduMirror}$, a multi-agent platform designed as a computational laboratory for the scientific study of educational social dynamics. EduMirror's framework integrates four key components: (1) A Systematic Scenario Design Workflow grounds simulations in established social science theory, ensuring construct validity. (2) To address the Fidelity Challenge, a unified Value-Driven Agent Architecture models agent motivation based on both individual psychological needs and Social Value Orientation (SVO). (3) To solve the Measurement Challenge, a Dual-Track Measurement Protocol employs specialized LLMs as a post-hoc Rater for observable behaviors and an in-situ Surveyor for internal states, transforming qualitative interactions into quantitative data. (4) Together, these components enable researchers to conduct controlled Intervention Experiments, branching simulations to systematically assess the causal impact of different strategies. We validate our platform through case studies on school bullying and group cooperation, demonstrating its capacity to generate theoretically-consistent and empirically-verifiable social phenomena, thereby establishing a robust methodology for in silico educational research.

Primary Area: datasets and benchmarks

Submission Number: 5807