A LoRA-adapted open-weight Gemma-3-27B model fine-tuned on 120 essays outperforms a fine-tuned GPT-3.5 baseline on TOEFL essay scoring across all reported metrics¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? On the same 360-essay held-out test set, the fine-tuned Gemma model achieves RMSE 0.474, QWK 0.828, and 90.56% agreement within ±0.5, surpassing the fine-tuned GPT-3.5 baseline (RMSE 0.573, QWK 0.78, agreement 84.72%). → John Maurice Gayed 2026

Evidence¶

John Maurice Gayed 2026¶

John Maurice Gayed. (2026). AiAWE: An Open-Source LLM Automated Writing Evaluation System Using LoRA-Adapted Instruction-Tuned Models. A PREPRINT. https://arxiv.org/abs/2606.12801 doi:10.48550/arxiv.2606.12801

q2 · i? · design · r2

Evaluation on the 360-essay held-out test set of the ETS TOEFL Independent Writing corpus, with both open-weight models fine-tuned on the same 120 essays under identical LoRA hyperparameters. The Gemma model shows "RMSE of 0.474 represents a 17.2% reduction relative to the GPT-3.5 baseline (0.573)" and QWK 0.828 vs 0.78.

"The fine-tuned Gemma model achieves the best performance across all metrics. Its RMSE of 0.474 represents a 17.2% reduction relative to the GPT-3.5 baseline (0.573), and its QWK of 0.828 exceeds the 0.78 reported by Wang and Gayed [2024] for the same GPT-3.5 fine-tuning approach on the same dataset."

Discussion¶

Related Claims¶

• Under identical LoRA hyperparameters and training data, the 27B Gemma model outperforms the 70B LLaMA model on every computed essay-scoring metric — related
• LLaMA's LoRA adaptation compresses its prediction range to roughly 2.0–4.0, while Gemma maintains a broader distribution closer to the ground truth — related
• Both fine-tuned models show regression toward the mean in score-wise bias, and error rises monotonically above score 3.0, with high scores hardest to predict — related
• LoRA fine-tuning hyperparameters are not model-agnostic: increasing rank above 64 destroyed LLaMA's feedback generation while Gemma remained robust — related
• LoRA adaptation adds only 0.6M trainable parameters on an 86.4M frozen backbone, keeping methods lightweight relative to LLM-based scoring — related
• Adaptation through prompt injection avoids fine-tuning, so the system works with any OpenAI-compatible LLM and its logic is transparent to educators — related
• Epi2Diff achieves the best performance across all four item-difficulty test sets and all reported metrics, outperforming SLM fine-tuning, LLM prompting, and supervised LLM adaptation baselines — related
• On benchmarks, GPT-3.5-turbo outperformed the fine-tuned Phi model in knowledge and story completion, while PhinetunEd scored best in story generation — related
• Supervised fine-tuning on real score–essay pairs substantially outperforms prompting for proficiency-conditioned writing simulation — related
• Fine-tuned GPT-4o achieves state-of-the-art scoring of TOEFL argumentative essays with QWK 0.84 and RMSE 0.44 against official scores — related
• Self-hosted open-weight LLMs offer privacy, reproducibility, and environmental advantages over commercial LLM interfaces in educational ITSs — related