The HIPE-OCRepair-2026 competition, part of ICDAR 2026, evaluated the effectiveness of large language models in correcting OCR errors in historical documents. The contest focused on improving search and access to digitized collections by addressing legacy OCR issues without re-digitization.
- Participants corrected noisy transcripts from English, French, and German newspapers and printed works from the 17th to 20th centuries.
- The evaluation used a retrieval-oriented scoring approach rather than diplomatic scoring to reflect practical use cases.
- Four teams submitted systems ranging from zero-shot prompting to continued pre-training and fine-tuning.
- Results showed that modern LLM-assisted systems significantly improve OCR quality, though performance varies across languages and noise levels.
- Over-correction on low-noise inputs was identified as a recurring challenge, highlighting the need for evaluation beyond character error reduction.
The dataset, scorer, and evaluation pipeline are publicly released to support future research in this area.