HIPE-OCRepair is an ICDAR 2026 competition dedicated to OCR post-correction for multilingual historical documents, with a particular focus on LLM-assisted approaches. The goal is to evaluate whether modern large language models can effectively address the “OCR debt” accumulated in vast digitized collections, and to provide a unified, high-quality benchmark for this long-standing challenge.

Despite major progress in OCR engines, historical documents remain difficult to digitize accurately: degraded scans, complex typography, multilingual variation, and layout noise often produce errors that propagate downstream into search, NLP, and digital humanities research. Because large-scale re-digitization is impractical for most institutions, post-correction remains the most realistic solution.

HIPE-OCRepair introduces a new benchmark, standardized evaluation metrics, curated multilingual datasets, and a leaderboard to systematically assess LLMs and alternative approaches for OCR post-correction.

Background: Why a New OCR Post-Correction Benchmark?

Previous OCR post-correction competitions (ICDAR2017 Competition on Post-OCR Text Correction, ICDAR 2019 Competition on Post-OCR Text Correction) advanced the field, but limitations remain:

  • inconsistent transcription policies
  • heterogeneous segmentation
  • variable annotation quality
  • lack of large-scale multilingual data
  • no unified, LLM-ready benchmark

Meanwhile, the recent surge of LLM-based OCR post-correction studies shows both promise and contradictory results (please check ➡️ References), partly because of inconsistent datasets, evaluation methods, or ground truth quality. HIPE-OCRepair addresses this gap by providing a standardized, reproducible, community-wide benchmark.

Benchmark

The HIPE-OCRepair benchmark provides a unified, multilingual, and LLM-ready evaluation suite for OCR post-correction across multiple historical collections. It is designed to capture the diversity and complexity of real-world archival material while offering consistent segmentation, metadata, and ground truth.

The benchmark includes:

  • Curated multilingual datasets (English, French, German) covering newspapers, monographs, and other historical sources
  • Harmonized ground truth with consistent transcription and segmentation standards
  • Noisy OCR inputs aligned at the line level, the minimal and most meaningful correction unit
  • Rich metadata (date, collection, document type, OCR engine quality indicators) to support context-aware correction

For details on datasets, task setup, and evaluation, see the ➡️ Tasks & Data page.

Evaluation Overview

More about the evaluation methodology can be found on the ➡️ Evaluation page.

Registration & Participation

The full schedule, including data releases, leaderboard openings, evaluation phase, and deadlines, is available on the ➡️ Timeline page.

After registration, for questions, participants may contact the organizers via the HIPE-OCRepair 2026 mailing list.

About the Organizers

HIPE-OCRepair is organized by members of the Impresso project (EPFL and University of Zurich), with extensive experience in historical document processing and shared tasks, including previous coordination of CLEF-HIPE 2020/2022/2026.

See the full organizer biographies on the ➡️ Organizers page.