The Order Matters: Sequential Fine-Tuning of LLaMA for Coherent Automated Essay Scoring

In three linesStudy on sequential fine-tuning of LLaMA-3.1-8B for Automated Essay Scoring. Model trained progressively on discourse elements (lead → position → claim → evidence → conclusion) achieves F1=65% (evidence) and 87% (conclusion), outperforming LLaMA-70B. Shows that curriculum design aligned with text structure improves coherence and generalization.

Summary generated by Claude — human-verified