Chat ≠ tutoring
Fluent dialogue can give answers too early, hallucinate facts, and ignore mastery. A tutoring framework must control pedagogical moves and update a learner model.
Research · PhD topic
LLM Personalized Tutoring
An intelligent LLM-based personalized tutoring framework for adaptive education — with learner modelling, pedagogical control, and teacher oversight.
A defendable doctorate sits between classical intelligent tutoring systems (strong control, brittle authoring) and free-form LLM chatbots (fluent talk, weak adaptivity). The research object is a framework: diagnose learner state → choose a pedagogical action → generate constrained dialogue → update the learner model — with a human teacher able to oversee and escalate.
LLM+
Dialogue is not enough — adaptivity needs a learner model
HITL
Teacher / facilitator oversight stays in the loop
1 domain
Lock a curriculum testbed (lab Learning Hub or professional training)
3–4 yrs
Doctoral plan with yearly checkpoints
Working title
An Intelligent LLM-Based Personalized Tutoring Framework for Adaptive Education
Keep “framework”, “personalized”, and “adaptive” in the title: they are the claim. Subtitles may emphasise learner modelling, pedagogy, or domain grounding once a track is locked.
Curriculum testbeds
Pick one primary testbed; optionally ground content in aviation or IT training.
| Setting | What it is | Why it matters | PhD leverage |
|---|---|---|---|
| NEXORA-iQ Learning Hub | AI paths, Windows Server, Linux — live class companions | Open educational setting the lab already runs | Fastest path to a prototype and ethics-friendly pilots |
| Professional / technical training | Aviation ops or IT infrastructure upskilling modules | Uses your industry background as authentic content | Stronger external validity if you can access learners |
| Formal education partner | School / university course with a human teacher | Curriculum ownership and assessment already exist | Best if a supervisor or partner provides classroom access |
Research gap notes
Classical ITS meets LLMs
Intelligent tutoring systems already solved parts of adaptivity with brittle authoring. LLMs reverse the problem: fluent generation, weak control. The doctorate owns that control layer.
Lab advantage
NEXORA-iQ Learning Hub and daily live classes give a real curriculum and learners — use them as the first testbed, optionally grounded in aviation/IT training content.
Choose one PhD emphasis
The framework title can host all five as chapters, but examiners expect one primary contribution. Lock an emphasis before the main study.
An Intelligent LLM-Based Personalized Tutoring Framework for Adaptive Education
Claim
Personalization is a closed loop: diagnose learner state, select a pedagogical action, generate dialogue with an LLM under constraints, then update the learner model — not a free-form chatbot on course notes.
Research questions
- How should an LLM tutor represent learner mastery, misconceptions, and affect so adaptation is inspectable?
- Which policy (hint, example, quiz, remediating path, escalate to human) improves learning gains versus static LLM chat?
- How can teachers or facilitators oversee, override, and correct the tutor without rewriting every turn?
Method
Design a reference framework (learner model + pedagogical policy + constrained generation + oversight UI). Implement on one curriculum testbed. Controlled study: framework tutor vs vanilla LLM chat vs materials-only. Pre/post knowledge tests + interaction logs + teacher review.
Contribution
A reusable tutoring framework and evaluation protocol for adaptive LLM education — architecture + evidence, not a single app demo.
Avoid
Wrapping ChatGPT around PDFs and calling it “personalized”. Examiners want the adaptation loop and learning outcomes.
Fit
Recommended default: the working title is the thesis claim.
Inspectable Learner Models for LLM Tutors in Adaptive Education
Claim
Without an explicit learner model, an LLM cannot personalize reliably — it only improvises. The thesis object is a mastery/misconception model that steers generation and can be audited.
Research questions
- Which learner-state features (knowledge components, error types, confidence) best predict next-step success in the chosen domain?
- How should the model be updated from dialogue and short assessments without opaque fine-tuning alone?
- Can learners and teachers understand and contest the model’s view of “what I know”?
Method
Define knowledge components for one syllabus. Combine short quizzes with dialogue signals. Compare Bayesian / KT-style updates vs LLM-inferred state. Measure prediction of post-test and teacher agreement with model snapshots.
Contribution
An inspectable learner-modelling method tailored to LLM tutoring, with open logging schema.
Avoid
End-to-end fine-tuning with no interpretable state and no learning-gain study.
Fit
Choose if you prefer modelling and measurement over full system architecture.
Pedagogical Control Policies for Adaptive LLM Tutoring
Claim
The scarce resource is not fluency — it is when to explain, when to withhold the answer, and when to escalate. Policy, not prose quality, is the personalization engine.
Research questions
- Which scaffolding strategies transfer from classical ITS into LLM dialogue without collapsing into answer-giving?
- How should the policy balance challenge and support for novices vs intermediates on the same syllabus?
- When does escalation to a human tutor outperform more LLM turns?
Method
Encode a pedagogical policy layer above the LLM (rules + learned policy optional). Ablate policies in tutors. Learning gains, hint abuse, time-on-task, and escalation rates as outcomes.
Contribution
A pedagogical policy design for LLM tutors with empirical comparison of strategies.
Avoid
Prompt-only “be a Socratic tutor” with no policy state or ablation.
Fit
Strong if you enjoy instructional design and Learning Hub class practice.
Domain-Grounded LLM Tutors for Professional and Technical Adaptive Learning
Claim
Hallucinated teaching is worse than no tutor. Grounding (curriculum RAG, verified examples, tool checks) plus personalization is required for aviation/IT-style technical education.
Research questions
- How should retrieval and tool checks constrain tutor utterances on technical syllabi?
- Does personalization still help when generation is strictly grounded?
- How do domain experts rate correctness and pedagogical usefulness versus a general LLM?
Method
Build a grounded tutor over one technical curriculum (e.g. Linux/Windows admin or aviation-ops concepts). Expert review of sampled dialogues + learner study. Error taxonomy for hallucinations and pedagogical faults.
Contribution
A grounding + personalization stack for technical tutoring with measured factual and learning outcomes.
Avoid
Generic chat over the open web with unverifiable answers.
Fit
Best bridge to aviation + IT infrastructure content expertise.
Evaluating Personalized LLM Tutors: Learning Gains, Over-Reliance, and Oversight
Claim
Engagement and thumbs-up are not enough. Adaptive tutors must show learning gains, bounded over-reliance, and workable teacher oversight.
Research questions
- Which metric suite predicts real learning better than chat satisfaction?
- How does personalization affect over-reliance and help-seeking compared with static materials?
- What oversight workflow lets a teacher audit and correct tutor behaviour at acceptable cost?
Method
Open evaluation harness: pre/post tests, delayed retention, log-based behavioural metrics, teacher audit time. Run on one testbed; publish materials for replication.
Contribution
An evaluation and oversight protocol other labs can reuse — including NEXORA-iQ classes.
Avoid
Leaderboard-style LLM benchmarks with no learners.
Fit
Methods-heavy path if access to many learners is limited at first.
Doctoral plan
Four-year shape with near-term lab checkpoints. Adjust dates to your university calendar; keep the sequence.
Frame and curriculum lock
- Confirm the working title and one disciplinary home
- Literature map: ITS, AIED, LLM tutors, knowledge tracing, scaffolding
- Lock testbed syllabus (Learning Hub track and/or professional module)
- Ethics / consent plan for learner studies
- Lab two-page proposal published; full outline checkpoint
Framework prototype
- Implement learner model + policy + constrained LLM generation + teacher oversight
- Pilot with small cohort (lab members / class volunteers)
- Refine instruments (pre/post, logs, teacher audit)
- Methods chapter draft; lab hours demo
Main study
- Controlled comparison: framework vs vanilla LLM vs materials-only
- Analyse learning gains, misconceptions, over-reliance, oversight cost
- Expert review sample for grounded correctness (if technical domain)
- Lab working paper + results draft
Thesis and venue
- Thesis write-up and revisions
- Venue paper (AIED, CHI/Learning@Scale, computers & education, or NLP+education)
- Open framework pieces for NEXORA-iQ Learning Hub where allowed
- Viva / defence preparation
Near-term milestones
| When | Milestone | Track |
|---|---|---|
| Sep 2026 | Topic locked: LLM personalized tutoring framework | Decision |
| Nov 2026 | Choose track emphasis + curriculum testbed | Research |
| Dec 2026 | Literature map + full outline checkpoint | Research |
| Mar 2027 | Framework prototype v1 + pilot protocol | Build |
| Sep 2027 | Pilot complete; instruments fixed | Study |
| Jun 2028 | Main study data collected | Study |
| Dec 2028 | Lab working paper + thesis core chapters | Write |
| 2029–2030 | Thesis submission and venue paper | Submit |
Proposal summary
This proposal frames personalized LLM tutoring as an adaptive framework with an explicit learner model, pedagogical policy, constrained generation, and teacher oversight. The candidate implements the framework on one curriculum testbed and evaluates learning gains against vanilla LLM chat and materials-only baselines.
Problem
LLM chatbots are being dropped into education without a personalization loop. They answer too readily, hallucinate on technical content, and leave teachers without audit or escalation. Classical ITS had control but not scalable dialogue. The gap is a framework that joins both.
Aim
Design, implement, and evaluate an intelligent LLM-based personalized tutoring framework for adaptive education that improves learning outcomes while remaining inspectable and overridable by a human teacher.
Objectives
- Specify the framework: learner model, policy, generation constraints, oversight.
- Implement on one syllabus (Learning Hub and/or professional module).
- Run a controlled study of learning gains and over-reliance.
- Publish open evaluation materials in the lab series.
Core reading anchors
- Intelligent tutoring systems (VanLehn and successors). Classical loop: assess → decide → tutor → update. Re-state this with LLMs.
- Knowledge tracing / learner modelling. Mastery over knowledge components; keep it inspectable beside the LLM.
- Scaffolding and cognitive apprenticeship. When to hint, model, or fade support — policy, not vibes.
- LLM-based education / AI tutors (recent AIED & CHI). Position against chat wrappers; cite learning-outcome studies, not demos only.
- Human-in-the-loop teaching. Teacher oversight, correction, and escalation — ties to responsible AI practice.