Bandit Framework Optimizes LLM Essay Scoring, Cuts Costs

This research introduces a novel method for optimizing automated essay scoring (AES) using Large Language Models (LLMs), specifically addressing the high operational costs associated with frequent LLM calls. Traditional AES systems often rely on fixed prompting strategies, which can be inefficient and fail to adapt to evolving optimal configurations. The proposed framework redefines prompt selection as a multi-armed bandit (MAB) problem. Each prompt type is treated as an "arm," allowing the system to adaptively learn and select the most effective and cost-efficient prompting strategy during inference. This online learning approach contrasts with conventional offline hyperparameter optimization. Experiments conducted on IELTS Writing Task 2 essays demonstrated that this MAB framework achieved scoring accuracy comparable to an exhaustive grid search, but with a remarkable 78.4% reduction in LLM calls. The study also explored different grading recipes, finding that a multi-step approach with calibration examples yielded the highest accuracy. By tracking token usage and latency, the research provides valuable cost-reliability curves, offering practical insights for educational technology platforms balancing assessment validity with operational expenses.