Метод ітеративної мультиагентної верифікації оцінок у віртуальних навчальних середовищах на основі пояснюваного штучного інтелекту

• Datum: 2026
• Ключові слова: keywords
• Hauptverfasser: Колодій, Роман, Виклюк, Ярослав
• Format: Artikel
• Sprache: Englisch
• Veröffentlicht: Kamianets-Podilskyi National Ivan Ohiienko University 2026
• Online Zugang: https://mcm-tech.kpnu.edu.ua/article/view/354691
• Назва журналу: Mathematical and computer modelling. Series: Technical sciences

The paper develops a method of iterative multi-agent verification of grades that ensures transparency and reliability of automated open-response assessment in virtual learning environments (VLE). The relevance of the problem of opaque decision-making by large language models (LLMs) and their tendency to generate factually incorrect statements in educational assessment tasks is substantiated. A formal model of the VLE assessment subsystem as a multi-agent system comprising three specialized agents (evaluator agent, verifier agent, and explainer agent) is proposed. For each agent, input-to-output mapping functions are defined. The MultiAgentGrading algorithm implementing a four-phase assessment procedure is developed: initial generation using chain-of-thought strategy (ante-hoc component), critical analysis by the verifier (post-hoc component), iterative refinement, and pedagogical aggregation of the result. The method combines built-in and post-hoc explainability mechanisms in a unified agent interaction cycle, enabling minimization of hallucination risks and enhancement of assessment reproducibility. Convergence conditions for the iterative process and a safeguard mechanism against infinite loops are defined. The transition from a linear «more explanations means more trust» paradigm to a calibrated trust concept is justified, where user confidence aligns with the model’s actual capabilities.