The adaptive learning taxonomy for responsible large language model integration in higher education.

Large Language Models (LLMs) are being rapidly deployed in higher education, yet institutions lack integrated, theoretically grounded frameworks to guide responsible adoption. While meta-analytic evidence reports large positive effects on academic performance (g = 0.867; Wang and Fan 2025), emerging research reveals a "cognitive paradox": performance gains may coincide with diminished metacognitive accuracy and self-regulation-a gap no existing framework comprehensively addresses across adaptation design, data governance, pedagogical targeting, and learner agency simultaneously. METHODOLOGY: Through a targeted narrative synthesis of 35 empirical studies published between January 2024 and February 2025, supplemented by selected post-window studies incorporated during peer review, and grounded in four learning theories-Connectivism, Distributed Cognition, Cognitive Load Theory, and Self-Regulated Learning-this paper develops the Adaptive Learning Taxonomy for Educational Decision-Making (ALT-ED). The framework structures institutional decision-making across four operational dimensions: Adaptation Trigger, Data Granularity, Pedagogical Locus, and Agency. KEY

findingsThree central findings emerge: (1) the cognitive paradox requires shifting the pedagogical locus from cognitive scaffolding to metacognitive development; (2) high-resolution data collection amplifies algorithmic bias, supporting a default of progressive data minimisation; and (3) the policy-practice chasm demands transparent, auditable design frameworks rather than prohibition-based approaches. RECOMMENDATIONS: Institutions should prioritise a metacognitive pedagogical locus to foster "learning to learn" skills, employ progressive data granularity to balance personalisation with student privacy, and default to learner-negotiated agency as a structural safeguard against algorithmic determinism. LIMITATIONS: The framework has not yet been empirically validated through prospective institutional implementation. Additional limitations include reliance on predominantly short-term quasi-experimental evidence, the absence of standardised metrics for real-time metacognitive efficiency, and a literature base that is overwhelmingly English-language and Western-centric. Future research should focus on cross-institutional pilot validations, longitudinal studies of cognitive entrainment, and the development of standardised instruments for measuring metacognitive efficiency in AI-augmented environments.