Edit-level Majority Voting in GEC

• The paper demonstrates that performing majority voting on atomic edits mitigates over-correction by filtering out unstable, non-recurring corrections.
• The method employs a seven-step workflow to generate, extract, count, and reconstruct corrections from consensus edits, avoiding full-sentence selection.
• Empirical results across multiple benchmarks indicate significant precision improvements, especially for tasks requiring minimal and precise corrections.

Edit-level Majority Voting is a training-free inference method for large-language-model-based grammatical error correction (GEC) that replaces whole-sentence selection with voting over extracted edits. Multiple corrected candidates are sampled from a single LLM, each candidate is converted into an edit set relative to the source sentence, and only edits that recur often enough across samples are retained; the final corrected sentence is then reconstructed from those accepted edits. The method was introduced to mitigate over-correction, a persistent failure mode in LLM-based GEC in which models rewrite correct text for fluency, normalize dialectal forms, or append prompt-following boilerplate. By operating at the level of atomic edits rather than complete hypotheses, it treats repeated edits as a proxy for stable corrections and unstable edits as likely noise (Goto et al., 13 May 2026).

1. Problem setting and core definition

In GEC, the target behavior is to make the smallest possible changes needed to correct an erroneous sentence. The paper motivating edit-level majority voting argues that LLM outputs often violate this minimality requirement: a single sampled correction can mix genuine grammatical fixes with fluency-driven paraphrases and irrelevant insertions such as “Okay, the corrected sentence is…”. Because GEC evaluation emphasizes precision, including metrics such as $F_{0.5}$, unnecessary edits are disproportionately harmful. This makes over-correction a central inference-time problem rather than merely a stylistic artifact (Goto et al., 13 May 2026).

The “edit-level” qualifier refers to representing each candidate correction not as a monolithic sentence but as a set of edits extracted from the pair consisting of the erroneous source sentence $S$ and a candidate correction $H_i$. The paper uses edit extractors such as ERRANT or JP-ERRANT. An edit is an atomic transformation needed to map $S$ to $H_i$. Examples explicitly given include ["go" -> "goes"], ["the" -> ] for deletion, [-> "everyday"] for insertion, and extraneous insertions such as [-> "Okay, the ..."]. The voting procedure is therefore not over sentences, tokens, or model logits, but over these extracted edit objects.

A central implication of this formulation is that the final output need not coincide with any sampled hypothesis. The reconstructed sentence can combine edits that were distributed across different samples while excluding low-frequency edits that appeared only once. This is the main structural distinction between edit-level majority voting and sentence-level selection methods.

2. Formalization and inference procedure

The method begins with a single source sentence $S$ and a prompting template for GEC. The same LLM is sampled repeatedly with nucleus sampling to obtain $k$ corrected candidates,

$\{H_1, H_2, \ldots, H_k\}.$

Each candidate is converted into an edit set,

$\{E_1, E_2, \ldots, E_k\}.$

The union of all extracted edits is then formed, and each edit is assigned a count equal to the number of candidate edit sets in which it appears. The accepted edit set consists of those edits whose count is at least a threshold $T$, where $S$0. The final corrected sentence is obtained by applying the accepted edits to the original source sentence (Goto et al., 13 May 2026).

The paper presents the procedure as a seven-step workflow: input sentence and prompt; generate $S$1 corrected sentences; extract edits with ERRANT or JP-ERRANT; form the union of edits; count edit frequencies; retain edits with frequency at least $S$2; reconstruct the output from the retained edits. In the reported experiments, the sampling configuration is fixed to $S$3, top-$S$4, and temperature $S$5. Diversity therefore comes from stochastic sampling from one model rather than from ensembling multiple models.

This single-model design is a defining property. The paper explicitly distinguishes the approach from prior majority-voting ensemble methods by noting that it does not require multiple models, model modification, or additional training. It interprets the procedure as a form of edit-level self-consistency: edits that recur across independent samples are treated as reflecting a more stable model decision than edits that appear only sporadically.

3. Mechanism for mitigating over-correction

The method’s rationale is that genuine grammatical fixes are relatively constrained and therefore recur across samples, whereas fluency rewrites, lexical substitutions, and prompt-following artifacts vary more from one sample to another. Requiring cross-sample agreement at the edit level therefore raises the acceptance bar for changes to the source sentence. In the paper’s framing, frequent edits are more likely to be true corrections, while infrequent edits are more likely to be stochastic side effects (Goto et al., 13 May 2026).

This mechanism differs sharply from sentence-level majority voting. If selection is performed over full hypotheses, the system must choose one entire candidate, even when each candidate contains both useful and spurious edits. Sentence-level selection cannot remove over-corrected fragments if those fragments are entangled with otherwise correct hypotheses. Edit-level majority voting can instead keep one correct edit from one candidate, another correct edit from a different candidate, and reject extraneous insertions or paraphrases that fail to recur.

The paper also contrasts the method with Minimum Bayes Risk (MBR) decoding. In the reported setup, MBR generates multiple hypotheses, scores each hypothesis with a utility function such as ERRANT $S$6 or GLEU, and returns the single best full sentence. The authors emphasize that if every sampled hypothesis contains some incorrect extra change, MBR must still select one of them. Edit-level majority voting avoids that limitation because it evaluates edits separately rather than committing to one complete candidate. The paper further notes that MBR can be disadvantaged when the utility is an n-gram metric such as GLEU, since brevity penalty and sentence-level scoring can distort selection.

A common misconception is that “majority voting” here means choosing the most common corrected sentence. It does not. The operative unit of aggregation is the extracted edit, and the winning set of edits may correspond to no sampled sentence at all.

4. Empirical behavior across datasets and languages

The paper evaluates the method on nine benchmarks spanning English, Czech, German, Ukrainian, Korean, Hindi, and Romanian. The English datasets are BEA-2019, CWEB-G, and JFLEG. The other-language datasets are AKCES-GEC, Falko-Merlin, UNLP-2023, Kor-learner, Hi-GEC, and RONACC. Across these nine benchmarks, the proposed method outperforms both greedy and MBR decoding in most cases (Goto et al., 13 May 2026).

The strongest gains occur on datasets with relatively few required edits per sentence. The paper identifies CWEB-G as showing the largest improvements and reports that, on that benchmark, Llama-3.1-8B-Instruct improves by 14.0 $S$7 points relative to greedy decoding. Improvements are also reported on BEA-2019, UNLP-2023, and Hi-GEC. By contrast, performance is less consistent, and sometimes worse, on JFLEG, Falko-Merlin, and RONACC, which the paper characterizes as requiring more substantial fluency editing rather than minimal correction.

This dataset-dependent pattern is central to the paper’s interpretation of the method. When the task primarily requires sparse, precise edits, high-consensus filtering suppresses over-correction and improves precision. When the task permits or rewards broader rewriting, strict edit consensus can filter out too much and produce under-correction. The authors connect this to average source-reference edit distance rather than to language identity itself: lower edit distance datasets tend to benefit more from the method.

The paper also states that the method can approach or surpass some fine-tuned non-LLM systems in certain settings. It does not present this as a universal result, but as part of a broader empirical picture in which inference-time edit aggregation can materially change the operating regime of general-purpose LLMs on GEC.

5. Thresholding, prompt stability, and validity of frequency signals

The threshold $S$8 is the main hyperparameter controlling the strictness of voting. Larger $S$9 yields stricter filtering, so fewer edits survive; smaller $H_i$0 yields more aggressive correction, at the cost of retaining more low-consensus edits. The paper tunes $H_i$1 on development data to maximize the target metric, with $H_i$2 fixed in the experiments. It explicitly describes $H_i$3 as a control knob for the precision-recall trade-off: higher $H_i$4 produces more conservative correction and less over-correction, whereas lower $H_i$5 allows more edits but increases the risk of noise (Goto et al., 13 May 2026).

The paper provides evidence that the threshold behaves as intended. On very low-error-density data, higher thresholds can be beneficial; on fluency-oriented data, smaller thresholds may be needed to avoid under-correction. A case study described in the paper shows greedy decoding introducing unnecessary edits such as dialect shifts and extraneous text, with moderate $H_i$6 removing much of the noise and high $H_i$7 retaining only the most clearly supported correction.

A second empirical result concerns prompt robustness. The authors evaluate 10 different instruction templates and report that the proposed method achieves higher average scores and lower variance than greedy decoding across prompts. This suggests that the method is less sensitive to prompt wording, because prompt-induced boilerplate and unstable phrasing tend not to recur consistently at the edit level and are therefore filtered out.

The paper also directly tests its central hypothesis that edit frequency correlates with correction validity. Using eight sampled candidates, it groups edits by frequency and measures precision by frequency group. The reported pattern is that precision generally increases as edit frequency increases. That finding is not merely anecdotal: it supplies the empirical justification for treating repeated edits as high-confidence candidates for acceptance.

For implementation and reproducibility, the authors release two repositories: GEC-DATASETS, for dataset loading and access, and GEC-LLM, an LLM inference wrapper built around vLLM with support for few-shot sampling and ensemble voting. The paper’s code example uses LLMGECFewShot, EnsembleVoting(min_vote=7), and evaluation with GLEU or ERRANT-based metrics.

6. Position within majority-voting research

Edit-level majority voting belongs to a broader family of aggregation procedures that replace reliance on a single model output with some notion of consensus, but its aggregation unit is unusually fine-grained. In multi-agent debate, for example, majority-style aggregation can operate at the level of agent agreement thresholds. One recent study defines majority consensus as discussion termination once more than $H_i$8 of agents agree, with the final answer extracted from the last agent; that is an agent-level protocol rather than an edit-level one (Kaesberg et al., 26 Feb 2025). Edit-level majority voting is therefore not a debate protocol and does not aggregate agent opinions or dialogue acts.

It also differs from semantic best-of-$H_i$9 selection methods such as Radial Consensus Score (RCS), which map entire candidate answers into embedding space, compute a weighted Fréchet mean as a semantic center, and select the candidate closest to that center (Nguyen et al., 14 Apr 2026). RCS extends majority voting from discrete answer counts to geometric consensus over answer embeddings; edit-level majority voting instead decomposes each candidate into atomic corrections and aggregates those edits directly. Both approaches move beyond exact-string majority counts, but they do so at different representational levels: full-answer semantics in one case and source-conditioned edit structure in the other.

This suggests a broader taxonomy of majority-based inference. Some methods choose among whole answers, some aggregate over semantic embeddings, some terminate multi-agent discussions when agreement thresholds are met, and edit-level majority voting aggregates over local transformations applied to a shared source sentence. In the specific domain of GEC, the paper’s evidence indicates that this edit-centric formulation is particularly well matched to the precision-sensitive, minimal-edit character of the task (Goto et al., 13 May 2026).