What makes an LLM a tutor? A component-centric survey of LLM-based and agentic tutoring systems

Abstract Large language models (LLMs) are reshaping intelligent tutoring systems (ITSs) by reducing authoring costs and enabling flexible interaction. However, effective tutoring requires more than fluent responses. It also depends on pedagogical control, domain grounding, learner modeling, interaction design, orchestration, and reliable evaluation. We examine these requirements through a component-centric framework that organizes LLM-based tutoring systems into four functional layers: pedagogical policy, domain and grounding, learner state, and interaction/interface. A cross-cutting orchestration layer coordinates these components, while agentic tutoring systems are treated as an emerging subset of the broader field. Drawing on representative literature on LLM-based and agentic tutoring systems, published mostly between late 2022 and mid-2026, we analyze how tutoring capabilities are implemented across system architecture, learner modeling, interaction, orchestration, and evaluation. The reviewed literature suggests that evaluation remains concentrated on response-level and model-based proxies, while direct evidence of learning gains remains limited. Learner-state modeling and interaction design also receive less systematic attention. We conclude by identifying challenges in answer leakage, pedagogy–solving trade-offs, learner-state calibration, interaction validity, and proxy validity.