# Qwen improves LLM-based OCR correction with veto gate

Digest AI · Research · published 2026-11-05T05:00:00Z

Canonical: https://digestai.news/story/qwen-improves-llm-based-ocr-correction-with-veto-gate

## Summary

Researchers at Qwen have developed a new method called EditVeto to improve the accuracy of language models (LLMs) in correcting Optical Character Recognition (OCR) errors found in historical documents. The technique, described in an arXiv preprint, involves placing a veto gate after a full-context LLM rewrite, which character-aligns the rewrite with the original OCR input and then splits any differences into blocks. A 0.6 billion-parameter gate is used to decide whether each block should be kept or reverted based on true error locations. The study found that this approach significantly reduces errors in five tested rewriters, including two not seen during training. When applied to published competition outputs, the veto gate lowered cMER for three of four teams but slightly increased it for the winning run. In a second look at the data, an ensemble of gated edits reached a cMER of 0.0069 and a preference score of 0.928, primarily through fine-tuning and pooling rather than gating. Despite being behind the top two runs in published competitions, EditVeto showed promise as a third-party solution that can be built using only public models and data.

## Key points

- Qwen developed EditVeto to improve LLM-based OCR correction
- Gate character-aligns rewrite with original OCR input
- 0.6 billion-parameter gate decides whether to keep or revert blocks

## Why it matters

This research could help reduce errors in automated document processing, potentially improving the accuracy of AI-driven text analysis for historical documents and other applications.

## Sources

1. [Rewrite, Then Veto: Rewriter-Agnostic Edit Gating for LLM-Based Post-OCR Correction](https://arxiv.org/abs/2611.05539) (arXiv cs.CL, 2026-11-05, primary source)

## Cite

Digest AI, "Qwen improves LLM-based OCR correction with veto gate", 5 November 2026, https://digestai.news/story/qwen-improves-llm-based-ocr-correction-with-veto-gate

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/qwen-improves-llm-based-ocr-correction-with-veto-gate.json