Grading at scale: privacy-preserving automated short answer grading with small language models on CPUs

Automated Short Answer Grading (ASAG) is invaluable for scaling grading in an overloaded education system. However, current solutions offer a trade-off: proprietary Large Language Models (LLMs) like GPT-4o raise privacy concerns, while self-hosting requires expensive, hard-to-obtain GPU resources. To resolve this dilemma, this paper explores the use of Low-Rank Adaptation (LoRA) fine-tuned Small Language Models (SLMs, ≈1B parameters) against a zero-shot proprietary baseline (GPT-4o). GRAS, a new