Micro-Corpus LLM Synergy for English Writing Instruction: A Controlled Experiment

Abstract

Even though Large Language Models (LLMs) demonstrate high efficiency and effectiveness in text generation, their tendencies toward hallucination and limited controllability raise concerns in educational contexts, where transparency and reliability are essential. Responding to claims that traditional corpora are becoming obsolete, this study re-examines the pedagogical value of corpora in the era of LLMs through a focused review of recent literature and a classroom-based randomized controlled trial. Specifically, representative studies published between 2020 and 2024 are systematically reviewed using a three-dimensional analytical framework comprising Efficiency, Controllability, and Educational Adaptability. In addition, a controlled experiment (N = 60) in a university English writing course compares a pure LLM-based instructional model with a corpus-augmented LLM model using Retrieval-Augmented Generation (RAG). The results show that the corpus-augmented model significantly reduced immediate post-test errors (9.7) compared to the LLM-only group (28.3), improved delayed retention, encouraged greater learner questioning behaviour, and reduced teachers’ preparation time by approximately 30%. These findings demonstrate that small-scale, task-specific pedagogical corpora play a critical role in enhancing the controllability and instructional effectiveness of LLMs. The study proposes a micro-corpus–LLM synergy framework and provides an open micro-corpus to support classroom implementation and replication.