This paper presents a comprehensive literature survey on Optical Character Recognition (OCR)-based systems for the automated evaluation of handwritten answer sheets, emphasizing the integration of modern deep learning and lan- guage understanding techniques. The review consolidates re- search spanning handwritten text recognition (HTR), post-OCR correction, writer-adaptive learning, and multimodal assessment that combines textual, mathematical, and diagrammatic inputs. Recent developments leveraging large vision–language models (VLMs) such as GPT-4V are analyzed for their potential to perform semantic comparison and rubric-aware grading of handwritten solutions. The survey also examines transformer- based architectures, meta-learning frameworks for unseen writer adaptation, and hybrid OCR pipelines integrating CNN, RNN, and attention mechanisms. Key datasets, benchmark results, and performance trends across diverse educational and language settings are discussed. In addition, the paper identifies major challenges such as OCR noise propagation, reasoning inconsis- tencies in large language models, and domain-specific calibration requirements for STEM assessments. By synthesizing current progress and limitations, this work aims to provide a structured foundation for developing future end-to-end, multimodal, and semantically aware AI-driven evaluation systems for scalable and reliable academic assessment. Index Terms—OCR, Handwritten Text Recognition, Answer Sheet Evaluation, Semantic Correction, Transformers, CTC, Meta-learning.
Joseph et al. (Mon,) studied this question.