Simulating Human-Like Learning Dynamics with LLM-Empowered Agents

ACL ARR 2026 January Submission7557 Authors

Keywords: Large Language Models, Multi-agent Simulation, Human Learning Process, Shortcut Learning, Educational Psychology

Abstract: Simulating human learning behavior with deep learning methods has become a major research focus in both psychology and intelligent systems. However, existing approaches often rely on static surveys or rule-based modeling, limiting their ability to capture long-term learning dynamics and interpretable cognitive processes. We propose **LearnerAgent**, a multi-agent framework based on Large Language Models (LLMs) that simulates a realistic classroom environment to study human-like learning dynamics over time. LearnerAgent instantiates psychologically grounded learner profiles, including Deep, Surface, and Lazy Learners, along with a persona-free General Learner probing the base LLM’s default behavior. A year-long simulation with weekly learning, monthly strategy decisions, periodic exams, and peer interaction enables longitudinal analysis. Our key findings are as follows: 1) Learners exhibit stable behavioral and cognitive patterns consistent with their profiles. 2) The Deep Learner achieves the most sustained cognitive growth, while the others struggle with trap questions due to reliance on surface-level heuristics. 3) Learners’ self-concept evolves realistically, with the General Learner developing inflated self-efficacy. 4) Crucially, the base LLM emerges as a "diligent but brittle surface learner"—appearing competent through effort yet lacking robust, transferable understanding. Extensive simulations show the effectiveness of LearnerAgent and offer insights into LLMs' behavior.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: AI/LLM Agents, Interpretability and Analysis of Models for NLP, Ethics, Bias, and Fairness

Contribution Types: Model analysis & interpretability

Languages Studied: English

Submission Number: 7557