Overview
Most AI tools for education answer questions, while instructors rarely see what students actually misunderstand. ALIVE closes the loop. An instructor-governed AI delivers lectures, asks formative questions, and evaluates spoken or free-text answers. It gives feedback, hands demonstrated weaknesses to a one-to-one tutor, and tracks recurring misconceptions so the instructor can revise the course.
Consequential course changes stay under human control. The instructor approves content, can override the system, and decides whether student data may be used.
This is a new system and should not be confused with the earlier ALIVE Avatar-Lecture Interactive Video Engine preprint, which is listed under the MENTOR project.
🔁 The Closed Instructional Loop
🏗️ System
- Shared model and services: the classroom and the tutor share a course-adapted 31B multimodal model, cloned-voice text-to-speech, and a retrieval service. All of it runs on institution-operated hardware, so student data stays on site.
- Graph-structured retrieval over a 102-document course corpus.
- LoRA adaptation on 37,236 gated course-derived pairs.
- Per-student mastery state, six question types, and deterministic misconception tracking.
📊 Results
| Component | Comparison | Gain |
|---|---|---|
| Graph-structured retrieval | Relevant vs. instruction-matched irrelevant context | +0.241 gold-token recall |
| LoRA course adaptation | Held-out evaluation set 1 | +0.056 recall |
| LoRA course adaptation | Held-out evaluation set 2 | +0.102 recall |
🛡️ Governance by Design
- Approval gates before generated content reaches students.
- Audit trails for system decisions and content changes.
- Instructor overrides at every stage.
- An explicit data-use switch that controls whether student interactions may inform course revision.
🚀 Key Contributions
- A deployed closed-loop AI education system connecting lecture delivery, assessment, tutoring, and course revision.
- Real deployment across two courses, 48 lectures, and 2,067 slides on institution-operated infrastructure.
- Measured gains from graph-structured retrieval and course-specific LoRA adaptation.
- Human governance built into the architecture: consequential course changes stay under instructor control.
📝 Citation
@unpublished{islam2026closing,
title = {Closing the Instructional Feedback Loop with Instructor-Governed AI},
author = {Islam, Md Zabirul and Wang, Ge},
note = {Manuscript in preparation},
year = {2026}
}