INTELLECTUALIZING VIRTUAL LEARNING ENVIRONMENTS WITH ARTIFICIAL INTELLIGENCE AGENTS: CURRENT TRENDS, METHODS AND CHALLENGES

Abstract

The paper provides a structured review of current approaches to intelligent virtual learning environments (VLEs) based on the interaction of artificial intelligence (AI) agents. The aim of the study is to summarize methods for designing agent-based components of VLEs and to identify practical limitations and research gaps that hinder the scalable deployment of personalized learning. The review is based on publications indexed in leading bibliographic and digital libraries (Scopus, Web of Science, IEEE Xplore, ACM Digital Library, Google Scholar) with emphasis on interaction models agent-student, agent-agent, and agent-instructor/administrator. The paper discusses classical machine learning techniques used for recommendations and learning outcome prediction (k-nearest neighbors, kNN; support vector machines, SVM; Naive Bayes), clustering algorithms for behavioral pattern discovery (DBSCAN), and approaches to building and optimizing individual learning pathways. In addition, conversational support tools (chatbots), reinforcement learning (RL) for adaptive selection of pedagogical interventions, and large language models (LLMs) as the core of generative agents (tutors, facilitators, and assessment assistants) are analyzed. The analysis highlights key risks that arise when VLEs adopt generative agents: limited transparency of decisions, hallucinations, domain-specific biases, and potential leakage of sensitive educational data. The study argues that integrating explainable AI (XAI) - in particular, local explanations for recommendations, justified assessment decisions, and traceability of agents' reasoning - improves user trust and the controllability of the learning process. The paper also emphasizes the need to embed ethical principles (privacy, fairness, safety, academic integrity) and quality assurance procedures for LLM outputs (fact-checking, context constraints, and citation policies). The review highlights the potential of multi-agent architectures in which agent roles are separated by function (knowledge diagnosis, personalization, feedback generation, moderation) and interaction is optimized via shared knowledge-exchange protocols and performance metrics. The practical contribution of the paper is a set of recommendations for selecting VLE intelligence methods according to learning goals and transparency requirements, as well as outlining directions for further research on bias mitigation and increasing the reliability of generative agents.