Empowering English Writing Assessment and Feedback through Large Language Models: A Hybrid Framework Integrating Local Fine-Tuning and Prompt Engineering

This study investigates the application potential of Large Language Models (LLMs) in English writing assessment and feedback. Experiments were conducted to compare a fine-tuned local small-scale LLM with the web-based DeepSeek-R1 model, optimized via prompt engineering. Results indicate that for composition scoring, the fine-tuned local model achieved moderate consistency with human grading, validating its feasibility. Conversely, regarding instructional feedback, while the local model underperformed, the DeepSeek-R1 model combined with optimized prompt engineering yielded significant improvements, demonstrating that high-quality feedback is attainable through prompt engineering. Consequently, provided privacy risks are controlled, teachers and students can leverage such LLMs to effectively enhance English writing instruction and learning outcomes.