VisionTutor: An Adaptive Tutoring Platform for Real-Time Progressive Learning

Abstract

Learning using AI has progressed at a high speed, but tutoring systems are not multimodally perceptive in real time, pedagogically grounded, or responsive to context—particularly in the fields of STEM. VisionTutor is a real-time adaptive AI tutoring system proposed in this manuscript that facilitates live screen viewing, speech interaction, and multimodal input comprehension using Gemini 2.5 Pro. VisionTutor offers real-time feedback through a canvas-based visual environment and conversational tutoring, including coding and learning mathematics. One of the innovations is the Cognitive Learning Scoring Model, which is trained on 5,000 simulated learner-system interactions from a DistilBERT-based regression pipeline. It predicts learner engagement and effectiveness along parameters such as promptness, tool use, problem-solving approach, and AI-behavior. The evaluation has an R² measure of 0.9856, reflecting high predictive reliability. The system also applies learning science frameworks such as ICAP and SRL, mapping technical behavior indicators into educative constructs. The findings of the present research show that VisionTutor offers personalized assistance and also offers understandable performance analytics, which are beneficial both for students and instructors. This research offers a basis for real-time learning analytics, extendable feedback processes, and a more fundamental integration of multimodal artificial intelligence into learning environments.