NEMESIS: A neuro-symbolic framework for metacognitive scaffolding in agentic tutoring systems, architecture and simulation study
Year: 2026
Type: conference-paper
Abstract: Many AI tutors personalize subject instruction but do little to help learners plan, monitor, and evaluate their own learning. We present NEMESIS, a neuro-symbolic framework with a Metacognitive Knowledge Graph based on Zimmerman's self-regulated learning model, a transformer estimator with gated multimodal fusion and cold-start priors, a symbolic planner, and three agents coordinated through a shared StateGraph. The simulation used five synthetic learner archetypes and 100 paired-seed runs per condition. NEMESIS reached the target SRL trajectory in 13.94 steps, compared with 14.52 for a phase-rule ablation and 15.05 for a trigger-matched random policy; the no-scaffolding control had no converged run within the 50-step horizon. None of the five within-archetype comparisons with the phase-rule ablation remained significant after Holm correction, so the 0.58-step aggregate difference is descriptive. Mean convergence remained between 13.81 and 14.59 under tested Dirichlet perturbations and between 14.02 and 14.56 under parametric transition mismatch. These Monte Carlo results describe planner behavior in the designed model; learner effectiveness requires the proposed human-subjects validation.
Source: EPJ Web of Conferences
Author: Shatrughna Upadhyay
Institution: Intuit (United States)
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: diamond
OpenAlex ID: W7201858331