ICHMS 2026 · AERI Lab, University of Missouri

AI tutors that follow the curriculum, not just the prompt.

SAGE is an intelligent tutoring framework that enforces instructor-set boundaries through access control, not hopeful prompting — so students get guided scaffolding instead of a bypassed lesson.

See how it works Read the Paper

The Problem

Public LLMs are built to be helpful — not curriculum-aware.

A general-purpose chatbot will happily hand a student the exam answer key, a complete code solution, or content two chapters ahead — because nothing stops it at the retrieval layer. Prompt-based "act like a tutor" instructions are easy to ignore or talk a model out of.

The Approach

A three-layer Safety Gatekeeper, not a prompt

Every query is checked against role, curriculum state, and cognitive intent before it ever reaches the language model's context.

Query Received RBAC + ABAC Role & Curriculum Check Cognitive Access Intent Check Scaffolded Response

Query Received

A student submits a question or request to the tutor.

RBAC + ABAC

Verifies the learner's role and curriculum readiness before any content is retrieved.

Cognitive Access

Classifies intent — conceptual, problem-solving, debugging, or review — and constrains the response type.

Scaffolded Response

Only content that clears all three checks reaches the model's context.

By The Numbers

Benchmarked against GPT and Gemini as prompt-based tutors

Security Compliance

90%

vs. 30% for prompt-based tutors

Curriculum Compliance

80%

access-control enforcement on locked topics

Overall Preference

44%

competitive with baselines' 42%

Proof

Access control in action

Based on evaluation results from the ICHMS 2026 paper against prompt-based LLM tutors (GPT/Gemini).

PromptWithout SAGE (Prompt-Based Tutor)With SAGE
"Show me the exam answer key." ⚠️ May leak restricted content under adversarial framing — only ~30% security compliance for prompt-based tutors. ✅ RBAC denies access to exam-key content outright — SAGE reaches 90% security compliance.
"How do I reverse an array in C?" Returns a complete working solution, encouraging cognitive offloading. Cognitive Access Control detects a problem-solving intent and returns hints/pseudocode instead of a full solution.
"Explain pointers." (prerequisites not yet mastered) Answers directly regardless of the learner's readiness. ABAC checks the knowledge graph, finds prerequisites unmet, and redirects toward prerequisite scaffolding.

SAGE was developed as part of the AERI Lab at the University of Missouri–Columbia. Interested in the research? Contact me · Read the paper