Search and Filter

When Rubrics Change: Cross-Rubric Generalization For Critical Thinking Essay Scoring

Date

AI Overview

Why it matters

Automated scoring systems usually assume fixed scoring criteria, but educators frequently revise rubrics or add new ones. Obtaining expert human annotations for every new rubric can be expensive.

Study at a glance
  • Question: Can language models generalize to score student essays using previously unseen critical thinking rubrics?
  • Setting and data: 500 argumentative essays by 6th–12th grade U.S. students from the PERSUADE 2.0 corpus, evaluated across six critical thinking rubrics.
  • Method: Model evaluation comparing fine-tuning with intermediate trait representations and varied essay supervision against zero-shot prompting baselines.
  • AI tools: Llama 3.1 8B Instruct, GPT-5, GPT-5-mini, GPT-4.1, ChatGPT (GPT-5.4)
Key findings
  • Rubric-agnostic traits improved scoring performance by 5.0% macro F1 when neither the target rubrics nor the target essays were seen during training.
  • Increasing target-essay supervision during training consistently improved scoring performance across rubrics, with human gold labels providing the strongest gains.
  • The best fine-tuned open-source Llama model outperformed zero-shot GPT-5-mini prompting by 2.1% in macro F1 and trailed GPT-5 by only 1.9%.

Summary drafted with AI from the full paper and may contain errors. See the full paper for details.

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.