GRACE-Agent: a multi-stage LLM agent framework for grammatical error correction in English academic writing with reduced false positives

Grammatical Error Correction (GEC) in English academic writing is critically important for non-native language learners; however, existing Large Language Models (LLMs) commonly exhibit over-correction in academic contexts, resulting in elevated False Positive Rates (FPR) that severely undermine their trustworthiness in educational practice. This paper proposes GRACE-Agent (Grammar-Aware Contextual Evaluation Agent), a multi-stage LLM inference pipeline tailored for English academic writing, which suppresses false positives at the semantic, register, and probabilistic levels by integrating three functional modules: a Contextual Verifier, an Academic Register Filter, and Multi-Pass Self-Consistency Reasoning. We independently construct GRACE-Corpus, a fine-grained annotated dataset comprising 4,300 English academic texts covering 8 grammatical error categories, and design a multi-expert rater validity verification scheme centered on Cohen’s \(\kappa \) $\kappa$ coefficient. Experimental results demonstrate that GRACE-Agent achieves an F0.5 of 84.1% on the test set, surpassing the best baseline by 9.6 percentage points; the FPR is reduced to 9.2%, representing a 61.7% reduction compared to the ChatGPT zero-shot method; the macro-average \(\kappa \) $\kappa$ with expert raters reaches 0.80, attaining the level of “near-perfect agreement.” Ablation experiments validate the independent contribution of each module, with the Multi-Pass Self-Consistency Reasoning module contributing the most. Limitations of this study include a focus on sentence-level grammatical errors, reliance on a closed-source GPT-4o inference engine, and dataset coverage primarily limited to general academic writing in science, engineering, and the humanities. This study provides a validated framework for the trustworthy deployment of LLMs in academic writing assistance.