A Design‑Science Proof‑of‑Concept for Constraining LLM Pedagogical Decisions Using Knowledge Graphs

Source abstract

Abstract Artificial intelligence–driven educational systems increasingly rely on Large Language Models (LLMs) to deliver adaptive recommendations, personalized learning pathways, and automated pedagogical support. However, despite their linguistic fluency and contextual reasoning capabilities, LLMs remain vulnerable to pedagogical hallucinations, including logically invalid prerequisite recommendations, over-prescriptive interventions, and non-verifiable instructional guidance. These limitations pose substantial risks for educational trustworthiness, learner autonomy, and explainable decision-making within gamified learning environments. This study investigates whether symbolic grounding through Knowledge Graphs (KGs) can constrain the generative action space of LLMs to produce logically verifiable and pedagogically valid outputs. Adopting a Design Science Research (DSR) paradigm, the study develops a controlled proof-of-concept artifact that integrates explicit pedagogical semantic triples representing learner states, occupational prerequisites, and educational interventions. A quasi-experimental comparative evaluation was conducted across five state-of-the-art LLMs (GPT-5.3, Sonnet 4.6, Gemini 3.1 Pro, Mistral Large 2, and DeepSeek) under ungrounded and KG-grounded conditions. The findings demonstrate that KG grounding systematically eliminates over-prescriptive recommendations, enforces minimal intervention compliance, and improves logical action validity and explicit reasoning traceability across all evaluated models. Rather than functioning as post-generation hallucination filtering, the proposed approach proactively restricts invalid pedagogical actions through symbolic constraint structures embedded within the reasoning process itself. The study contributes empirical evidence that neuro-symbolic integration can improve explainability, determinism, and trustworthiness in educational AI systems while preserving adaptive personalization capabilities. Although deliberately bounded as a proof-of-concept rather than a production-scale educational platform, the findings establish foundational support for knowledge-grounded generative AI architectures in high-stakes domains requiring transparent and logically constrained decision-making.