Answers are grounded in this paper and cite their sources.
Marking handwritten, descriptive exam answers is still largely manual.An evaluator reads every script, judges whether the student really grasped the concept, and writes a mark.Two evaluators seldom agree exactly, and a single evaluator can drift during a long marking session.This paper presents SAGE (Smart Answer-sheet Grading Engine), a platform that combines optical character recognition, sentence embeddings, retrieval-augmented generation (RAG) and large-language-model (LLM) scoring in one workflow.An OCR pipeline digitizes scanned or handwritten scripts, the text is converted to dense vectors, and those are matched against an index built from reference and syllabus material.The LLM then scores each answer against the passages retrieved for it, rather than leaning only on what it remembers from training.We want to move past keyword matching and pattern classification, which earlier work finds brittle on descriptive answers, and arrive at scoring that can explain a mark in plain language.We cover the architecture, the scoring pipeline, the module design, the related work behind our choice of a retrieval-based ap-proach, and an evaluation protocol built on OCR accuracy, agreement with human graders and retrieval precision.The evaluation dataset has not been built and no experiments have been run, so the numbers reported here are targets, not measured results.
No takes yet. Share an insight, caveat, or question.
M et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: