Simulating Human Like Learning Dynamics with LLM-Empowered Agents

By Yu Yuan, Lili Zhao, Wei Chen, Guangting Zheng, Kai Zhang, Mengdi Zhang, Qi Liu

Published: 7 Aug, 2025 | arXiv:2508.05622

Overview

This paper introduces LearnerAgent, a simulation framework built on large language models to investigate how learning behaviours evolve over time. The researchers designed a year-long simulated learning environment where agents modeled as different learner types engage in structured learning activities. These activities include weekly knowledge acquisition, strategy updates, testing sessions, and peer interactions.

Why It Matters

Simulating learning behavior with language models opens new ways to examine cognition, metacognition, and self-regulation in artificial agents. Traditional cognitive modeling methods are often rigid or limited to short-term tasks. 

This work allows for a more continuous and context-rich analysis of how AI agents develop understanding over time. It also provides tools to evaluate the cognitive strengths and weaknesses of large language models using human-inspired frameworks.

Core Method

The LearnerAgent framework simulates the learning process using weekly iterations across twelve months. In each cycle, the agents engage in the following activities:

• Content Study: Reviewing new materials in psychology and machine learning.

• Strategy Reflection: Choosing or adjusting study strategies based on past performance.

• Testing: Answering domain-specific questions, including "trap" questions that test for shallow understanding.

• Peer Interaction: Discussing concepts with other simulated agents to model collaborative learning.

The agents are powered by a large language model and are prompted in ways that reflect the behavior of different learner types. For example, the Deep Learner prioritizes understanding and critical thinking, while the Surface Learner focuses on memorization and short-term performance. The Lazy Learner reduces effort and avoids engagement unless forced. The General Learner does not follow a specific style but adapts freely.

Key Findings

• Cognitive growth was seen most clearly in the Deep Learner, who showed steady improvement and generalization of concepts across time.

• Surface Learners performed well on standard tasks but were exposed by trap questions that required deeper understanding.

• General Learners developed strong self-efficacy beliefs even when their actual performance was limited, showing a disconnect between confidence and capability.

• The base language model without any persona guidance tended to behave like a diligent Surface Learner, showing efficiency in learning tasks but limited depth or transfer of knowledge.

Limitations and Considerations

• The simulated environment simplifies many real-world learning factors such as emotional state, external motivation, and diverse instructional strategies.

• The learning behaviors are dependent on the capabilities and biases of the underlying language model, which may not generalize across model families.

• The personas were designed manually and may not fully capture the complexity of real human learning styles.

• While the study simulates progression over time, it remains a scripted environment that may not fully reflect how learners evolve in open and dynamic systems.

Path Forward

This work opens several avenues for future research and practical applications:

• Incorporate more complex psychological traits such as motivation, curiosity, or persistence into the agent design.

• Expand testing to include real-world educational tasks and evaluate how simulated learners compare with human students.

• Use LearnerAgent as a diagnostic tool to evaluate or fine-tune language models by observing their ability to simulate realistic learning trajectories.

• Apply this framework to build adaptive tutoring systems that can simulate and respond to diverse learner profiles in real-time.