New Llama-based model improves essay scoring generalization to unseen rubrics

Researchers have developed a new framework for automated essay scoring (AES) that improves generalization to unseen scoring rubrics. By using rubric-agnostic intermediate representations called 'traits' and controlled supervision, their fine-tuned Llama-based model achieved a 5.0% improvement in macro F1 score compared to a baseline without traits in the most challenging scenario. This approach also showed competitive performance against proprietary models, with the best open-source model outperforming GPT-5-mini by 2.1% macro F1 and trailing GPT-5 by only 1.9%. AI

IMPACT Enhances the ability of AI systems to evaluate essays across different grading criteria, potentially improving educational tools.

RANK_REASON Academic paper detailing a new method for automated essay scoring. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• GPT-5
• GPT-5 mini
• Hugging Face
• llama
• Nischal Ashok Kumar

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →