Learning to Grade Efficiently: A Bandit-Driven Prompt-Selection Framework for Low-Cost LLM Essay Scoring

Synthesis: Large language models demonstrate strong capability in automated essay scoring, but contemporary approaches typically employ fixed prompt selection that fails to address operational cost concerns and evolving optimal configurations. Manakina and Bogdanov propose a cost-aware multi-armed bandit controller that treats each prompt type as an arm, enabling adaptive selection of optimal prompting strategies during inference. On IELTS Writing Task 2 essays, the framework achieves comparable scoring accuracy to exhaustive grid search while reducing Large Language Models (LLMs) calls by 78.4%. The multi-step assessment approach with calibration examples achieves the highest accuracy, and the work produces the first cost-reliability learning curves for essay scoring.

Key Findings

1. A multi-armed bandit controller adaptively selects optimal prompting strategies during inference for essay scoring.
2. The framework achieves comparable scoring accuracy to exhaustive grid search while reducing LLM calls by 78.4%.
3. Four grading recipes (multi-step vs single-step, with vs without calibration examples) were implemented; the multi-step approach with examples achieves the highest accuracy.
4. Token usage and latency are tracked alongside agreement metrics to produce the first cost-reliability learning curves for essay scoring.
5. The work transforms prompt selection in automated essay scoring from an offline hyperparameter optimization problem into an efficient online learning task.

What this means for practice

• Designers. Put prompt selection under an online controller instead of fixing a template: treat each grading recipe as an arm in a multi-armed bandit and update it on agreement between predicted and examiner scores, which cut Large Language Models (LLMs) calls by 78.4% while holding scoring agreement.
• Designers. Budget tokens and latency as first-class constraints — track token usage, latency, and agreement together, because few-shot prompts with rubrics and rationales can exceed the context window of mid-tier models.
• Designers. Do not assume more rubric detail improves scoring: the simplified prompt without detailed rubric explanations matched or beat the rubric-heavy variant (QWK 0.485, MAE 0.965) at fewer tokens.
• Designers. Use the multi-step recipe calibrated with annotated high- and low-scoring example essays where accuracy matters most, since it achieved the highest accuracy of the four recipes tested.
• Researchers. Extend evaluation beyond a single benchmark and a single model before treating the framework as generalizable to other assessment contexts.

Limitations

• Evaluation uses one corpus — the Kaggle IELTS Writing Scored Essays Dataset, 787 Academic Task 2 compositions with official band scores of 1–9 — so results may not transfer to ASAP, TOEFL11, or other essay types.
• Only one model, Google Gemini Flash 2.5, was tested; the authors state they plan to extend to GPT-4 and Llama-3 70B to assess generalizability.
• The epsilon-greedy controller held a constant exploration rate (ε = 0.2) throughout, and the authors acknowledge the implementation could be refined with adaptive exploration strategies.
• The authors describe the work as a work-in-progress with preliminary experiments and an arm space of four recipes; the released cost-reliability curves rest on that narrow prompt space rather than a broader or operational deployment.

Citation

Manakina & Bogdanov (2026). Learning to Grade Efficiently: A Bandit-Driven Prompt-Selection Framework for Low-Cost LLM Essay Scoring.