Long‑term personalized tutoring requires language models to turn a learner’s evolving history into suitable instructional support. Existing LLM‑based tutors use memory and retrieval for personalization, yet they often lack an explicit pedagogical link between historical evidence and teaching actions.
RegulaTutor addresses this gap by grounding the system in self‑regulated learning (SRL) research and adopting a cognitive, affective, metacognitive, and motivational (CAMM) framework. The framework derives learner interpretations from historical evidence and generates teaching policies that retain references to their supporting evidence, updating them as new observations arrive.
The workflow proceeds in stages: first, collect historical evidence; next, apply the CAMM framework to produce interpretations and policies; finally, deliver policies to the learner while logging execution records.
To enable systematic refinement, RegulaTutor introduces PairRoute, an offline procedure that examines intermediate execution logs to separate failures into harness, model, and mixed categories. PairRoute proposes targeted harness revisions, then builds stage‑level supervision for remaining model failures under the chosen harness, ensuring verifiable model improvements.
The study focuses on refining historical‑evidence acquisition, the foundation of personalized instruction, allowing outputs to be checked against learning logs and reference answers. Experiments show RegulaTutor improves historical‑evidence accuracy, learner diagnosis, and judged teaching quality across model configurations. PairRoute results demonstrate a clear endpoint gain in standalone‑model historical‑evidence accuracy along the refinement trajectory.
These findings highlight RegulaTutor’s potential to support lifelong personalized education through evolving instructional policies and targeted system enhancements.
Review