Adaptive AI-Enhanced Tutor System for STEM Education

Thesis Abstract

This study investigates the design, implementation, and evaluation of an Adaptive AI-Enhanced Tutor System (AIAETS) tailored for STEM education to address persistent achievement gaps and uneven access to high-quality instructional support in secondary and tertiary settings. Despite advances in intelligent tutoring, existing systems often fail to adapt to diverse learner profiles, particularly in complex STEM domains where conceptual reasoning and procedural fluency must co-evolve. The aim is to develop and validate an adaptive tutoring architecture that synergizes real-time diagnostic assessment, multimodal feedback, and personalized learning pathways to improve conceptual understanding, procedural mastery, and sustained engagement. Specific objectives include (i) developing a Bayesian knowledge tracing–based diagnostic engine integrated with a reinforcement learning (RL) policy tuner to personalize content sequencing, hints, and formative assessments; (ii) implementing a multimodal feedback module leveraging natural language processing, computer vision-based concept recognition, and symbolic reasoning to scaffold problem-solving strategies; (iii) evaluating learning gains in conceptual understanding, procedural fluency, and motivation across diverse STEM topics (algebra, physics, and introductory programming) and learner groups; (iv) examining the system’s impact on self-regulated learning behaviors and time-on-task; and (v) assessing usability, interpretability, and equity considerations in deployment. The research adopts a mixed-methods design, combining a randomized controlled trial (n = 320 participants across three universities) with longitudinal classroom observations (n = 12 classes, 360 hours) and qualitative interviews (n = 40). The population comprises secondary and early undergraduate learners enrolled in STEM courses with identified learning gaps. Data collection instruments include pre- and post-tests aligned with STEM taxonomies (conceptual knowledge and procedural fluency), system interaction logs, motivational scales, standardized attitude surveys, and semi-structured interview protocols. The primary instruments for validation are the STEM Mastery Assessment (SMA) and the Motivation for Learning STEM Scale (MLSS). Validity and reliability will be established through content validation with subject-matter experts, pilot testing (n = 40), Cronbach’s alpha for internal consistency (>0.80), and test-retest reliability analyses. Data analysis will proceed in stages (i) quantitative analysis using ANCOVA to compare post-test gains controlling for baseline, multilevel modeling to account for nested data (students within classes), and regression analyses to identify predictors of improvement; (ii) time-series analysis of log data to assess engagement trajectories; (iii) Bayesian knowledge tracing to estimate latent knowledge states and model update rates; (iv) reinforcement learning evaluation to measure policy effectiveness in content sequencing and hint provision; (v) qualitative analysis employing thematic analysis to extract learner experiences, consented with triangulation against quantitative results. A conceptual model drawing on Cognitive Load Theory, Self-Determination Theory, and the Constructivist learning framework will guide interpretation, with specific attention to how adaptive scaffolding modulates intrinsic motivation and cognitive load. The study will also incorporate equity-focused analysis to examine differential effects across gender, language proficiency, and prior achievement. Expected findings include (a) statistically significant gains in SMA scores and procedural fluency for the intervention group versus control, (b) improved engagement metrics and reduced cognitive load as evidenced by interaction patterns and time-on-task indicators, (c) robust diagnostic accuracy of the Bayesian tracing component and effective policy adaptation through RL tuning, and (d) positive qualitative feedback regarding perceived autonomy, usefulness of feedback, and perceived fairness in adaptation. The research contributes to knowledge by bridging adaptive tutoring, AI-driven assessment, and STEM education in a unified architectural framework, offering empirical evidence on the efficacy of integrated diagnostic, generative feedback, and content sequencing modules in real-world classrooms. It advances theoretical understanding of how adaptive tutoring interfaces influence motivation and cognitive processing in STEM domains and informs design guidelines for scalable, equitable AI-assisted instruction. The study concludes that an Adaptive AI-Enhanced Tutor System can deliver personalized, scalable, and pedagogically sound support that enhances STEM learning outcomes while sustaining motivation and reducing cognitive overload. Recommendations include refining interpretability features for teacher oversight, extending the system to multidisciplinary STEM curricula, exploring long-term retention effects, and expanding deployment to diverse institutional contexts to validate generalizability.