Enhancing Fairness and Efficiency in Subjective Assessment through LLM-Based Automated Grading

Abstract

In recent years, the demand for fairness, speed, and transparency in grading has catalyzed interest in automated systems, particularly for subjective, theory-based assessments. Unlike objective tests, these examinations require nuanced understanding and contextual reasoning, traditionally making them dependent on human graders. However, human grading is often affected by inconsistencies, biases, and fatigue-induced errors. This work presents a system that leverages Large Language Models (LLMs) as grading agents for automating the evaluation of handwritten, theory-based exam scripts. Methods: The methodology employs a modular system architecture in which uploaded scripts are digitized, interpreted using vision-based models, and subsequently graded by domain-specific LLM agents. The system is implemented using FastAPI for the backend, Celery and RabbitMQ for asynchronous task handling, Redis for log streaming and task status management, and Next.js for the frontend interface. For mathematics scripts, a Math Agent is used to evaluate exam responses through context-aware reasoning. Results: Preliminary evaluation indicates that the system can grade an eight-question script within three minutes, significantly faster than the approximately fifteen minutes required by a human grader. This demonstrates that LLM-based grading systems can scale efficiently while reducing human bias and fatigue. Discussion and Conclusion: The project provides a foundation for broader integration of LLMs into educational assessment, while acknowledging limitations in current open-source vision models and inference latency. Future improvements may include fine-tuning and offline model support to enhance speed and reliability.