Edit-level Majority Voting Mitigates Over-Correction in LLM-based Grammatical Error Correction

Abstract: Grammatical error correction using LLMs often suffers from the over-correction issue. To mitigate this, we propose a training-free inference method that performs edit-level majority voting over multiple candidates generated by a single model, without requiring model modifications or additional training. Across nine benchmarks covering English, Czech, German, Ukrainian, Korean, Hindi, and Romanian, the proposed method outperforms both greedy and MBR decoding in most cases. Moreover, it yields stable correction quality regardless of the instruction prompts used. We release two repository supporting GEC datasets loading and LLM inference.

• The paper proposes an edit-level majority voting method that aggregates recurrent edits from k candidates to mitigate over-correction.
• It demonstrates significant precision improvements, including a 14-point F0.5 gain on English datasets, without additional training.
• The method offers a training-free, prompt-robust solution suitable for rapid deployment in low-resource and multilingual domains.

Edit-Level Majority Voting for Over-Correction Mitigation in LLM-based Grammatical Error Correction

Problem Overview

LLMs have advanced the state-of-the-art in grammatical error correction (GEC), but exhibit a pronounced tendency toward over-correction: making unnecessary edits to already correct text, often for fluency or stylistic reasons rather than minimal grammatical remediation. This tendency contravenes the minimal edit paradigm exemplified by the F0.5 metric, which rewards precision—edits should be made only where necessary, preserving grammatical spans otherwise. Fine-tuned LLMs and dedicated model architectures can reduce over-correction, but at the cost of additional training, computational overhead, and maintenance complexity. Data scarcity across languages and domains limits the feasibility of such solutions.

Methodology

The paper proposes edit-level majority voting as a training-free inference mechanism. The process involves generating $k$ candidates for a GEC input via nucleus (top-p) sampling from a single LLM, extracting the sets of edits from each candidate, and retaining only those edits whose frequency across hypotheses meets a predefined threshold $T$. The key hypothesis is that valid, minimal edits will recur consistently across sampled outputs, while spurious or fluency-driven changes will not.

Formally:

• Let $S$ be the input sentence, and $\{H_1, \ldots, H_k\}$ be the corrected outputs.
• Edits $\{E_1, \ldots, E_k\}$ are extracted (e.g., via ERRANT).
• Aggregate edits across candidates, count occurrences, and accept only those with $count(e) \ge T$.
• Apply the selected high-confidence edits to $S$.

This procedure does not modify the LLM or require additional training, and is robust to prompt variations and model architecture.

Experimental Evaluation

English Datasets

Experiments were conducted on CWEB-G (low error density, over-correction sensitivity), BEA-2019 (moderate errors), and JFLEG (fluency-focused corrections). Models included Llama-3.1-8B-Instruct, Qwen3-8B, gemma-2-9b-it, and baseline GECToR/T5 models. The method yielded substantial gains, notably:

• Llama-3.1-8B-Instruct on CWEB-G: F0.5 improved by up to 14 points over greedy decoding.
• All evaluated LLMs showed increased precision at the cost of modest recall, effectively suppressing over-correction, especially in low-error domains.
• Performance approached or, in some cases, matched fine-tuned GEC baselines without retraining.

Edit-level majority voting consistently outperformed greedy and Minimum Bayes Risk (MBR) decoding, which, due to sentence-level utility and candidacy selection, cannot reliably eliminate pervasive over-corrections.

Multilingual and Low-Resource Datasets

Cross-linguistic evaluation covered Czech (AKCES-GEC), German (Falko-Merlin), Ukrainian (UNLP-2023), Korean (Kor-learner), Hindi (Hi-GEC), and Romanian (RONACC). The method proved effective where source texts required minimal edits. Improvements correlated with datasets having lower edit density; substantial gains were seen for Hi-GEC and UNLP-2023, but diminished returns appeared in high-density error domains (e.g., Falko-Merlin, RONACC).

Anomalies (e.g., AKCES-GEC's improvement despite high edit densities) indicate possible metric extraction idiosyncrasies or linguistic factors warranting further investigation.

Computational Trade-offs and Hyperparameters

Performance increases with candidate pool size ($k$) but plateaus beyond $k=8$, suggesting diminishing marginal returns relative to inference time. The threshold $T$ directly modulates the minimality of edits: higher $T$0 suppresses nearly all but the most confident corrections, while lower $T$1 risks allowing under-correction.

Robustness to Instruction Prompts

The method yields stable correction quality and output format across prompt templates and instructions, as evidenced by reduced score variance compared to greedy inference. This advantage obviates manual or heuristic removal of irrelevant text (e.g., "Sure! Here"), streamlining post-processing.

Theoretical and Practical Implications

Edit-level majority voting exposes strong self-consistency in LLM-generated corrections: valid edits surface repeatedly under stochastic sampling, while fluency or irrelevant outputs are idiosyncratic. This suggests a generalizable principle for deploying LLMs to other edit-based tasks, such as text simplification or machine translation post-editing, without retraining or model modification.

The method’s training-free nature makes it especially suitable for:

• Rapid prototyping and deployment in low-resource languages/domains;
• Scenarios with inconsistent or unreliable training data;
• Applications requiring strict minimal-edit constraints.

Despite its effectiveness against over-correction, under-correction remains unaddressed; lower thresholds could mitigate this, but risk reintroducing spurious edits. Ensemble methods beyond majority voting (e.g., correction acceptability discrimination, integer programming) and evaluation via reference-free metrics (SOME, IMPARA) constitute avenues for future research.

Conclusion

Edit-level majority voting is a practical, robust mechanism for reducing over-correction in LLM-based GEC, enabling significant precision improvements and competitive performance with fine-tuned models, particularly in domains requiring minimal edits. Its output stability, domain/language generality, and training-free deployment offer compelling advantages. The approach invites further exploration for edit-based tasks, nuanced metric development, and broader language coverage (2605.13624).