Instruction-Following LLMs for Grammatical Error Correction: Analyzing Neutral-Anchored Instructional Sensitivity Across Editing Modes

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

Research on Instruction-Following LLMs for Grammatical Error Correction (GEC) addresses the lack of a principled characterization for how these models adjust edit decisions across different editing modes. The study defines three modes—Neutral, Minimal-Edit, and Fluency-Edit—and quantifies instructional sensitivity by measuring neutral-anchored performance shifts. Seven LLMs, including proprietary and open-weight models, were benchmarked using a unified zero-shot prompting schema on CoNLL-2014, BEA-2019, and JFLEG datasets. Findings indicate that the Minimal-Edit instruction mitigates over-editing and boosts precision, sometimes improving recall for strong models. Conversely, the Fluency-Edit instruction often encourages broader paraphrastic rewriting, which may improve perceived fluency but lowers GLEU scores, suggesting a metric-objective mismatch. Notably, Claude-Sonnet-4.5 demonstrated superior zero-shot capabilities, achieving F_0.5 scores of 67.05 on CoNLL-2014, 64.91 on BEA-2019, and a GLEU score of 66.09 on JFLEG, surpassing previous zero-shot and matching few-shot results.

Key takeaway

For NLP Engineers and AI Scientists optimizing Grammatical Error Correction (GEC) models, understanding instructional sensitivity is crucial. Your choice of editing mode instruction directly impacts model behavior, with "Minimal-Edit" improving precision and reducing over-editing. Be aware that "Fluency-Edit" may lead to broader rewrites that lower standard metrics like GLEU, necessitating a re-evaluation of your success criteria. Leverage models like Claude-Sonnet-4.5 for superior zero-shot GEC performance, but always align your prompting strategy with your specific correction objectives.

Key insights

Instructional sensitivity significantly impacts LLM behavior in Grammatical Error Correction across defined editing modes.

Principles

• Minimal-Edit instructions reduce GEC over-editing.
• Fluency-Edit instructions can lead to metric-objective mismatch.
• Zero-shot LLM performance varies greatly with instruction type.

Method

The study defines Neutral, Minimal-Edit, and Fluency-Edit modes, then measures neutral-anchored performance shifts to quantify instructional sensitivity across seven LLMs using a unified zero-shot prompting schema on GEC datasets.

In practice

• Employ Minimal-Edit instructions for precise GEC.
• Evaluate Fluency-Edit outputs beyond GLEU scores.
• Consider Claude-Sonnet-4.5 for strong zero-shot GEC.

Topics

• Grammatical Error Correction
• Instruction Following
• Large Language Models
• Zero-Shot Learning
• Model Evaluation
• Claude-Sonnet-4.5

Best for: AI Engineer, Machine Learning Engineer, Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• On the Limits of LLM Adaptability: Impact of Model-Internalized Priors on Annotation Task Performance — LLM annotation reliability hinges on definition alignment, not just text memorization, as prompt-based corrections are largely ineffective.
• Encode Errors: Representational Retrieval of In-Context Demonstrations for Multilingual Grammatical Error Correction — LLMs' internal states contain extractable grammatical error representations (GER) that improve in-context learning for GEC.
• GIFT: Guided Fine-Tuning and Transfer for Enhancing Instruction-Tuned Language Models — GIFT enhances instruction-tuned models by guiding base model adapter training with instruction model confidence signals.
• The easiest way to improve prompts — Zero-shot and few-shot prompting are core techniques for controlling language model output based on guidance provided.
• CoCoGEC: Counterfactual Generation for Robust Grammatical Error Correction — Training GEC models with context-altered counterfactuals significantly enhances their robustness against contextual perturbations.
• English is Not All You Need: Systematically Exploring the Role of Multilinguality in LLM Post-Training — Multilingual post-training significantly improves LLM performance across languages and tasks, outperforming English-only approaches.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.