Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Fine-tuning Small Language Models for Cross-Prompt Arabic Essay Scoring: A Benchmarking Study

AbstractAutomated Essay Scoring (AES) has been studied since the early days of computational methods. While recent advances incorporating large language models (LLMs) have significantly improved performance, most research has focused on English. In this paper, we present the first benchmark of Arabic cross-prompt AES using Arabic-centric LLMs. We evaluate ALLaM 7B, Jais 8B, and Fanar 9B under zero-shot, few-shot, and fine-tuning settings, and compare them against established AES systems and frontier LLMs. Fine-tuning substantially improves performance across all models, with Fanar achieving state-of-the-art results and outperforming the strongest baseline by 8 points on LAILA dataset. We further show that fine-tuned Arabic-centric LLMs generalize effectively in cross-dataset setting (LAILA → ZAEBUC). Finally, we analyze the best-performing model, examining error patterns, trait interactions, and the effects of multi-trait autoregressive scoring.

More from EMNLP 2026

All EMNLP 2026 papers →