GPT-5 Summarization Improves Automated Essay Scoring Accuracy for Long-Form Student Writing
A new study proposes using GPT-5 variants to summarize long essays before automated grading, overcoming transformer input-length limits and improving scoring reliability. The method, tested on the ASAP 2.0 dataset, could enable faster, more accurate feedback in educational assessment.
Researchers have proposed a generative AI-assisted summarization framework to improve automated essay scoring (AES) for long-form student writing. The study, released on arXiv, addresses a key limitation of current transformer-based AES models: their fixed input-length constraints, which can cause information loss when processing lengthy essays.
Using the ASAP 2.0 dataset, the team generated controlled-length summaries with three GPT-5 variants—GPT-5, GPT-5 mini, and GPT-5 nano—and used those summaries as inputs for downstream AES models. This approach aims to preserve the original essay's key content while fitting within model token limits, thereby maintaining scoring reliability.
The findings suggest that summarization before scoring can make AES more practical for real-world educational settings, where essays often exceed typical transformer input windows. By reducing information loss, the method could help schools provide faster, more consistent feedback on student writing without sacrificing accuracy.
For educators and students interested in AI grading tools, platforms like Gradescope and Turnitin already incorporate some AI-assisted features. Free summarization tools such as Smarterly or TLDR This offer a hands-on way to see how text summarization works in practice.