Date of Award

5-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

Committee Chair/Advisor

Andrew Robb

Committee Member

Dr. Andrew Duchowski

Committee Member

Dr. Abolfazl Razi

Committee Member

Dr. Matias Volonte

Abstract

Large language models (LLMs) are increasingly used in higher education, yet there is still limited empirical evidence on how effectively they support meaningful personalization for different learners. This dissertation investigates LLM-based personalization from three connected perspectives: students’ perceptions and use of LLMs, the feasibility of inferring learner profiles from text-based interaction, and the effects of aligning LLM-generated responses with those profiles during problem solving.

First, the dissertation examines how student characteristics shape perceptions of LLMs in academic settings. Using survey data from 666 undergraduate students and structural equation modeling, the study identifies distinct dimensions of perceived helpfulness, dependence, and accuracy assessment, as well as course-specific usefulness. Students generally viewed LLMs as helpful for comprehension and coursework, but these perceptions varied with usage intensity, subscription status, personality traits, and academic context, indicating that learner-related factors shape how LLMs are valued and relied upon.

Second, the dissertation evaluates whether contemporary LLMs can infer learning-style profiles from naturalistic text interactions. An AI-powered educational chatbot was used to collect 1,650 problem-solving exchanges from 153 university students. Zero-shot and few-shot prompting were then applied to GPT-4o, o3-mini, Gemini 1.5 Flash, and Gemini 2.0 Flash to classify students on the four dimensions of the Felder--Silverman Index of Learning Styles. The results show that current LLMs can infer some learning-style dimensions only unevenly: performance was strongest for sensing--intuitive and, in some settings, sequential--global, but predictions were often biased toward dominant learning-style categories, consistent with broader concerns about bias reproduction in LLM-based classification. Few-shot prompting improved visual--verbal classification, yet it did not reliably reduce this bias and in some cases appeared to amplify it, while active--reflective remained difficult across models.

Third, the dissertation investigates whether aligning LLM-generated explanations with learners’ learning-style profiles changes engagement and cognitive load during AI-supported problem solving. In a within-subject eye-tracking study, aligned responses did not increase dwell time on the solution area, but they were associated with smaller solution-area saccades and higher inter-subject correlation in the problem and solution areas, indicating more coordinated visual processing. Subgroup analyses showed the clearest gaze-based alignment effects for sequential and visual learners. Pupil-based generalized additive mixed models further showed that non-aligned responses were associated with higher fitted $\Delta \log(\mathrm{LF}/\mathrm{HF})$ across the solution, problem, text-box, and pooled all-AOI analyses, with the largest AOI-specific difference observed in the text-box region. Taken together, these findings suggest that alignment primarily reorganized attention and reduced physiological processing demand rather than simply increasing time on task.

Overall, the dissertation treats LLM personalization as a multi-stage problem involving adoption, learner modeling, and interaction design. Its main contribution is process-level evidence that profile-informed LLM responses can reduce friction during AI-supported learning. At the same time, the findings do not by themselves establish improved learning outcomes or definitive support for the strongest form of the meshing hypothesis, underscoring the need for future work linking these process measures to retention, transfer, and long-term learning.

Author ORCID Identifier

0000-0001-8736-1690

Download

Available for download on Monday, May 31, 2027

Share

COinS