Human-AI Co-authoring for Personalized Practice With Multilayer Graph

This paper presents a proof-of-concept for a scalable approach to personalized learning that generates practice questions aligned to intended practice targets, controlling for question complexity. The focus is to provide learner-centered scaffolding in preparation for assessments by generating questions that align learner topic proficiency with targeted complexity. Historical assessment items are transformed into a topic-linked representation by multi-labeling each question with topics from a predefined syllabus. The labeling is produced by employing a previously built prompt-engineered GPT-5 model configuration with retrieval augmented generation. Topic labels are combined with student performance traces to derive personalized topic-level proficiency proxies. We then build a multilayer graph based on conceptual hierarchy and historical co-assessment. Querying the multilayer graph retrieves coherent topic sets around a student's priority topics based on learning needs, supporting content selection that is both topically consistent and conceptually coherent. A prompt-engineered GPT-5 model then generates questions for an inferred SOLO taxonomy level based on a learner's profile (proficiency and residue), conditioned on the retrieved topic sets. Outputs are evaluated for SOLO alignment and profile calibration through expert review, showing a high level of agreement. This methodology offers a pathway to operationalizing human-AI collaboration for adaptive practice, enabling scalable, varied revision for learners.