文章背景与核心概要
随着大语言模型（LLM）在教育领域的广泛应用，自动作文评分（AES）正变得越来越普及。然而，复杂的提示词策略（如多步推理、包含校准示例等）往往伴随着高昂的计算成本和API调用开销，这给教育科技（EdTech）平台带来了沉重的经济负担。
本文提出了一种全新的、具备成本意识的自动化作文评分框架。该框架将不同的提示词策略视为多臂强盗算法（MAB）控制器中的“臂”，使系统能够在推理过程中自适应地选择最高效且准确的提示词配置。实验表明，该方法在雅思写作Task 2作文评分中达到了与穷举网格搜索相当的准确率，同时将大模型调用次数减少了78.4%，为平衡运营成本与评估效度提供了重要指导。
Learning to Grade Efficiently: A Bandit-Driven Prompt-Selection Framework for Low-Cost LLM Essay Scoring
Authors: Olga Manakina, Igor Bogdanov
Date: August 24, 2026
Subject: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
DOI: 10.48550/arXiv.2608.23814
Learning to Grade Efficiently: A Bandit-Driven Prompt-Selection Framework for Low-Cost LLM Essay Scoring
Authors: Olga Manakina, Igor Bogdanov
Date: August 24, 2026
Subject: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
DOI: 10.48550/arXiv.2608.23814
Summary
This paper introduces a novel, cost-aware framework for Automated Essay Scoring (AES) using Large Language Models (LLMs). By treating different prompting strategies as "arms" in a Multi-Armed Bandit (MAB) controller, the authors enable the system to adaptively select the most efficient and accurate prompting configuration during inference.
The framework successfully shifts prompt selection from a static, offline hyperparameter optimization problem to an efficient, online learning task. Experimental results on IELTS Writing Task 2 essays demonstrate that the MAB approach achieves scoring accuracy comparable to exhaustive grid search while reducing LLM calls by 78.4%. The study also provides the first cost-reliability learning curves, offering essential guidance for EdTech platforms balancing operational costs with assessment validity.
Summary
This paper introduces a novel, cost-aware framework for Automated Essay Scoring (AES) using Large Language Models (LLMs). By treating different prompting strategies as "arms" in a Multi-Armed Bandit (MAB) controller, the authors enable the system to adaptively select the most efficient and accurate prompting configuration during inference.
The framework successfully shifts prompt selection from a static, offline hyperparameter optimization problem to an efficient, online learning task. Experimental results on IELTS Writing Task 2 essays demonstrate that the MAB approach achieves scoring accuracy comparable to exhaustive grid search while reducing LLM calls by 78.4%. The study also provides the first cost-reliability learning curves, offering essential guidance for EdTech platforms balancing operational costs with assessment validity.
Key Contributions
- Adaptive Prompting: Implements an online control mechanism to dynamically select the best grading recipe (e.g., multi-step vs. single-step, with or without calibration examples).
- Efficiency Gains: Achieves a 78.4% reduction in LLM calls compared to traditional grid search methods.
- Cost-Reliability Analysis: Establishes the first learning curves that map token usage and latency against scoring agreement metrics.
- Practical Application: Provides actionable insights for educational technology platforms to optimize the trade-off between computational costs and grading accuracy.
Key Contributions
- Adaptive Prompting: Implements an online control mechanism to dynamically select the best grading recipe (e.g., multi-step vs. single-step, with or without calibration examples).
- Efficiency Gains: Achieves a 78.4% reduction in LLM calls compared to traditional grid search methods.
- Cost-Reliability Analysis: Establishes the first learning curves that map token usage and latency against scoring agreement metrics.
- Practical Application: Provides actionable insights for educational technology platforms to optimize the trade-off between computational costs and grading accuracy.
Access & Resources
Access & Resources
Metadata
- Comments: Accepted as a presentation at the EDM 2025 Workshop on Educational Data Mining in Writing and Literacy Instruction.
- MSC Classes: 68T05, 68T50, 68W27
- ACM Classes: I.2.6; I.2.7; K.3.1
Metadata
- Comments: Accepted as a presentation at the EDM 2025 Workshop on Educational Data Mining in Writing and Literacy Instruction.
- MSC Classes: 68T05, 68T50, 68W27
- ACM Classes: I.2.6; I.2.7; K.3.1