GEN - AI Autograding for Academic Evaluation

Abstract

Academic institutions face challenges in providing rapid, consistent, and accurate grading across various assignments. This paper introduces a multi-faceted GenAI-based automatic grading system incorporating four microservices: Programming Assignments, Diagram-based Assignments, Technical Subject-related Essays, and English Essays. Developed using the MERN stack with a TypeScript front-end, the system integrates large language models (LLMs) for automatic, discipline-specific evaluation. Each microservice individually assesses criteria such as code correctness, diagrammatic accuracy, technical precision, and linguistic depth, providing numerical grades and feedback. Programming assignments utilize collaborative LLMs for Java code evaluation, while the diagram assessment employs Generative AI, zero-shot classification (DeBERTa-v3-large), vision-language modeling (Qwen2.5-VL-7B-Instruct), and SpaCy-based semantic analysis. Essay evaluations leverage NLP techniques and GPT-3.5 models for detailed grammar, vocabulary, and coherence analysis. Pilot tests demonstrated significant improvements in grading speed, consistency, and reduced bias, enabling scalable, context-aware grading through aggregated dashboard results.