Search papers, labs, and topics across Lattice.
This study investigates cross-rubric generalization in automated essay scoring (AES) by training a Large Language Model (LLM) on essays labeled with one set of rubrics and evaluating it on previously unseen rubrics targeting different essay aspects. The researchers introduce a fine-tuning framework that utilizes rubric-agnostic intermediate representations, termed traits, alongside target-essay supervision, leading to a significant 5.0% improvement in macro F1 scores in challenging conditions where both rubrics and essays are unseen during training. Notably, their best-performing Llama-based model surpasses the performance of GPT-5-mini by 2.1% in macro F1, demonstrating the efficacy of trait-based structures in enhancing generalization capabilities in AES tasks.
Training on one rubric while scoring with another can boost AES performance by over 5%, revealing the potential of trait-based representations.
Automated essay scoring (AES) research has largely focused on cross-prompt generalization, where essays from unseen prompts are scored while the scoring criteria are typically held constant. In practice, however, educators may revise or even introduce new rubrics in their scoring task, to evaluate different aspects of essays. We study cross-rubric generalization: training on essays labeled under one set of rubrics and evaluating on previously unseen rubrics, which target different aspects of the essay. We use a Large Language Model (LLM) fine-tuning framework with two components: rubric-agnostic intermediate representations, called traits, and target-essay supervision under seen rubrics during training. On an AES dataset augmented with multiple rubric-defined labels of student critical thinking skills, we find that traits improve macro F1 by 5.0% over a baseline without traits in the hardest setting, where both target rubrics and target essays are unseen during training. We further find that increasing target-essay supervision improves performance, with our best fine-tuned open-source Llama-based model outperforming GPT-5-mini prompting by 2.1% macro F1 and trailing GPT-5 by 1.9%. These results show that trait-based intermediate structure and controlled supervision improve generalization to unseen rubrics.
