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

In the world of education, grading a stack of essays is a heavy, time-consuming task. For decades, computers have tried to help with this, but they often struggle to match the nuance of a human teacher. Recently, a new generation of powerful computer programs, known as large language models, has shown a remarkable ability to read and understand human writing. These systems can be instructed to act as examiners, assigning scores to student work based on specific rules. However, a significant problem remains: these programs are expensive to run. The more detailed the instructions given to the computer, the more it costs to get an answer. Furthermore, the best way to ask the computer a question is not always obvious; it changes depending on the specific essay or the version of the software being used. Educators and testing organizations face a difficult choice: pay a high price to run every essay through every possible set of instructions to find the best one, or risk using a cheaper, less accurate method that might miss the mark.

A team of researchers at Carleton University has proposed a new way to solve this dilemma. Instead of trying to guess the perfect set of instructions beforehand, they built a system that learns the best approach while it works. They treated the different ways of asking the computer to grade an essay like a set of slot machines, each with a different chance of paying out a good result. As the system graded essays, it kept track of which method produced scores that matched human teachers most closely and which ones were the most efficient. Over time, the system naturally stopped wasting effort on the poor methods and focused almost entirely on the ones that worked best. This approach allowed them to find the most effective grading strategy while using far fewer computer resources than traditional methods.

The researchers tested this idea using a collection of 787 real essays from the IELTS writing exam, a major international English proficiency test. They set up four distinct ways for the computer to grade the papers. Two methods asked the computer to give a single, overall score immediately, while the other two asked it to break the essay down into four specific categories—such as how well the writer answered the prompt and how organized the ideas were—before calculating a final score. For half of these methods, the computer was also shown examples of good and bad essays to help it understand what to look for; for the other half, it had to rely on its own knowledge. The system then began grading the essays, using a learning algorithm to decide which of the four methods to use for each new paper.

The results were clear and immediate. The system quickly learned that the most effective strategy was the one that broke the essay down into specific categories and provided examples of what a good or bad essay looked like. This method consistently produced scores that aligned most closely with the human examiners. In contrast, the methods that skipped the examples or tried to give a single score without breaking it down performed significantly worse. The learning algorithm was so effective that after a short period of testing, it stopped choosing the weaker methods almost entirely, dedicating over 70 percent of its efforts to the best-performing approach.

Perhaps the most surprising discovery came when the researchers looked at the cost of these instructions. Conventional wisdom suggests that giving a computer more detailed rules and explanations about how to grade should lead to better results. However, the study found the opposite. When the researchers removed the long, detailed descriptions of the grading rules from the instructions, the computer actually performed better. The simpler instructions, which relied on the computer's existing understanding of writing rather than a manual of rules, led to more accurate scores and used fewer computer resources. This suggests that these powerful computer programs have already learned the nuances of academic writing during their training and do not need to be told every single detail again.

The efficiency gains from this adaptive approach were substantial. By letting the system choose the best method on the fly, the researchers reduced the number of times they had to call the computer for help by nearly 79 percent compared to a traditional method that tried every possible option on every essay. This massive reduction in computer calls translated directly into a 73 percent drop in the total number of words the computer had to process, and a 70 percent reduction in the estimated cost of the experiment. The system achieved the same level of accuracy as the exhaustive, expensive method but did so with a fraction of the effort.

This work represents a shift in how automated grading can be handled. Instead of treating the choice of instructions as a fixed setting that must be decided before the work begins, the researchers showed that it can be a dynamic process that improves as it goes. While the study was limited to a specific set of essays and one type of computer program, the findings offer a promising path forward for educational technology. By balancing the need for accuracy with the reality of operational costs, this method provides a way to make high-quality automated grading accessible and sustainable for large-scale testing programs. The study suggests that the future of automated assessment lies not in building more complex instructions, but in building smarter systems that know how to learn which instructions work best in the moment.