New Dataset Links Student Writing and Feedback to Benchmark AI Assessment Tools

Researchers have introduced Exposía, a novel dataset designed to bridge the gap between academic writing instruction and computational assessment. The dataset captures the iterative nature of scholarly writing by pairing student research proposals with comments and reviews from both peers and instructors, along with detailed human-assigned evaluation scores.

What Exposía Includes#

The dataset draws from an introductory scientific writing course in computer science and documents the full writing cycle: initial drafts, feedback from reviewers, and subsequent revisions. Beyond the raw writing samples, Exposía provides a pedagogically grounded rubric for scoring both the quality of student proposals and the quality of peer feedback itself—treating review as a learnable skill.

Testing Large Language Models#

The research team used Exposía to evaluate how well current LLMs can automate two distinct tasks: scoring student proposals and rating the feedback that students give one another. Their findings reveal meaningful differences across models:

• Closed-source models outperformed open-weight alternatives consistently across both tasks, though the specific best-performing model varied depending on whether the goal was assessing proposals or reviews
• Prompting strategy matters significantly: asking LLMs to score multiple writing dimensions together proved more effective than evaluating individual aspects in isolation

Implications for Classrooms#

The results underscore both the promise and limitations of deploying modern AI in educational settings. While LLMs show capability in these tasks, the gap between proprietary and freely available models highlights ongoing challenges for institutions seeking to integrate accessible AI tools into writing instruction without relying on commercial platforms.