EditVeto Gates LLM OCR Rewrites to Curb Unwanted Edits

Summary:EditVeto places a veto gate after a full-context LLM rewrite for OCR restoration. The rewritten text is character-aligned with the OCR input, and a 0.6B-parameter classifier keeps or reverts each changed block. Across eight HIPE-OCRepair-2026 test sets, the gate improves all five evaluated rewriters, including two unseen during training: weighted cMER falls by 0.0026–0.0044 and preference scores rise by 0.15–0.25. The results are not uniformly positive. On published competition outputs, gating

---

Core idea: rewrite globally, then veto edits selectively

EditVeto addresses a central failure mode in OCR restoration for historical documents. LLMs can repair noise, malformed characters, and recognition omissions, but they may also rewrite text that was already correct. Rather than asking the model to generate only around suspected error locations, the method retains full-context rewriting and adds a reversible decision layer at the output stage.

The pipeline has three steps:

1. Character-align the rewritten text with the original OCR transcription;
2. Split the differences into contiguous edit blocks;
3. Use a 0.6B-parameter gate to decide whether each block should be kept or reverted.

The final output is not regenerated. It is assembled from the full rewrite by accepting or reverting individual blocks. This converts correction into an auditable binary decision rather than an implicit probability decision buried inside the generator.

Why not generate correction spans directly?\n

The paper's oracle analysis motivates the design: when true error locations are known, generating each span independently performs worse than applying no correction at all. In contrast, rewriting the full document and then vetoing selected edits can help.

The explanation is contextual. OCR errors in historical text may involve character confusion, broken lines, abbreviations, archaic spelling, and references to information elsewhere in the document. Independent span generation removes that context and may introduce locally plausible but globally inconsistent text. A full rewrite uses cross-position information; the gate then limits unnecessary modifications.

The oracle setting is nevertheless optimistic because it assumes that true error locations are known. Production systems do not have character-level ground truth, and the gate itself can produce false acceptances and false rejections. Its deployment value therefore depends on gate calibration under realistic error distributions.

How the gate operates

Instead of training another large OCR-restoration model, EditVeto uses a 0.6B-parameter classifier to arbitrate edits. It is trained on rewrites produced by two Qwen3 models, then frozen during evaluation across five rewriters, including two not observed during training.

Separating gate training from upstream generation has two engineering benefits:

• The gate can be inserted into an existing OCR-repair pipeline without retraining the main rewriter;
• It exhibits some transfer across rewriters instead of being locked to one generator's output distribution.

The trade-off is sensitivity to alignment and block construction. Character alignment errors, document segmentation artifacts, Unicode normalization, and systematic drift can concentrate failures in a small number of dense edit blocks. A production implementation should therefore expose per-block confidence and route ambiguous edits to review rather than forcing every decision into a hard binary action.

Evaluation: consistently helpful on average, not universally beneficial

Across eight HIPE-OCRepair-2026 test sets, all five evaluated rewriters improved. Two were unseen during gate training:

• Weighted cMER decreased by 0.0026–0.0044;
• Preference scores increased by 0.15–0.25;
• Paired bootstrap tests yielded approximately p ~ 0.0002 for each reported improvement.

These results support the claim that a frozen gate can transfer between rewriters. Improvements on unseen rewriters reduce the likelihood that the mechanism is merely memorizing a particular generator's editing patterns.

Competition outputs expose an important boundary. Gating improved cMER for three of four teams but increased the winning run's error from 0.0050 to 0.0065. EditVeto is not a free polishing layer. When an upstream rewriter is already highly accurate, learned editing preferences may instead preserve or introduce more errors than they remove.

Attribution in the ensemble results

In a second disclosed test, a retrained gate continued to improve LoRA-tuned rewriters, reducing cMER by 0.0011–0.0021. An ensemble of their gated edits reached cMER 0.0069 and preference 0.928. The fully public system remained behind the winner at 0.0050 and approximately tied with the second-ranked run at 0.0071.

The paper notes, however, that most of the ensemble gain came from fine-tuning and pooling rather than gating. The 0.0069 result should therefore not be attributed to EditVeto alone. A stronger ablation would separately report the original rewriter, LoRA-tuned rewriter, pooling alone, pooling plus gating, and the interaction between gating and each rewriter.

A third, secondary evaluation is also more modest. The unchanged gate reduces zero-shot Qwen3.8-27B cMER from 0.0114 to 0.0108, but it improves the preference score only for the LoRA-tuned version. The transfer gain is small and should not be generalized to every larger model or document domain.

Engineering trade-offs

EditVeto adds character alignment, edit-block extraction, and 0.6B gate inference after the expensive rewrite. Relative to rerunning a large repair model, this is comparatively controllable and can be parallelized by block. However, a complete implementation must retain the source text, rewritten text, alignment data, and per-block decisions for reproducibility and rollback.

It is most appropriate for:

• Batch post-processing in digital archives, libraries, and historical-document databases;
• Review pipelines where unwanted modification is more costly than a missed correction;
• Systems that already operate several LLM rewriters and need a common output-review layer;
• Human-in-the-loop workflows requiring an auditable record of accepted and reverted edits.

For tables, multi-column layouts, complex scans, and documents with extensive formatting changes, character alignment may be unstable. A production system should merge purely formatting or whitespace edits, assign confidence scores to semantic blocks, and preserve the source for high-impact uncertain edits.

Deployment recommendations

Do not enable EditVeto as a universal post-processor without domain-specific validation. Build evaluation sets that cover OCR noise types, languages and historical periods, layout complexity, OCR engines, rewriter versions, and the proportion of text that is already correct.

A practical rollout is:

1. Align rewritten and source text;
2. Normalize and segment edit blocks, removing whitespace-only and formatting-only differences;
3. Produce a keep/revert decision and confidence score for every block;
4. Preserve the source for high-impact blocks and route low-confidence decisions to review;
5. Track cMER, preference scores, and correct-text corruption rates separately.

Overall acceptance should not depend on aggregate cMER alone. For archival restoration, the more informative objective is the two-dimensional trade-off between recovery and over-correction, including worst-group performance across model sizes, document domains, and OCR-noise levels.

Conclusion

EditVeto turns unrestricted LLM OCR rewriting into an auditable sequence of block-level accept-and-revert decisions. Its gains across several rewriters, including unseen ones, show that a compact gate can transfer beyond the generators observed during training.

It is not a universal performance upgrade: the winning competition run degrades after gating, and much of the ensemble advantage comes from LoRA tuning and pooling. The transferable research conclusion is that global rewriting followed by output-level arbitration is more promising than isolated span generation. Whether EditVeto is worthwhile in production depends on the actual ratio of the cost of corrupting correct text to the cost of leaving an OCR error unresolved.

---

Source: ArXiv NLP/LLM (cs.CL) (2026-11-05)

— END —

Tags: #OCR Restoration Follow #EditVeto Follow #Qwen3 Follow #LLM Gating Follow #Document Intelligence Follow