ICDAR 2026 HIPE-OCRepair Competition on LLM-Assisted OCR Post-Correction for Historical Documents

Imagine a massive, dusty library where millions of old newspapers and books have been scanned into computers. The problem? The machines that read the text (called OCR) are like tired, myopic librarians from the 1990s. They squint at faded ink, confused by old-fashioned fonts and torn paper, and they often type out gibberish instead of the real words. For decades, fixing this meant either re-scanning the whole library (impossible) or trying to manually fix every typo (too slow).

Enter the HIPE-OCRepair-2026 competition, a high-stakes challenge where teams of AI experts brought in their newest "super-librarians": Large Language Models (LLMs). These are the brainy, context-hungry AIs that can read a sentence and guess what comes next with incredible fluency. The goal? To see if these digital geniuses could clean up the messy, typo-ridden transcripts without making things worse.

The Big Test: Can the AI Fix the Mess Without Breaking It?

The organizers set up a tricky game. They gave the AI teams noisy, broken text from historical documents in English, French, and German, dating back as far as the 1600s. The catch? The AI had to fix the text without seeing the original picture of the page. It was like trying to fix a torn letter just by reading the blurry photocopy, with no original to compare it to.

The teams had to be incredibly careful. If an AI "hallucinated"—meaning it confidently invented a word that wasn't there or "modernized" an old spelling to make it sound cool today—it failed the test. The goal wasn't to rewrite history; it was to recover the exact words the author wrote, just without the scanner's mistakes.

The Results: One Team Dominates, But It's Not a Magic Wand

After the dust settled, the results were clear but nuanced.

The Champion: The team BnF-Mistral took the top spot, and they didn't just win; they dominated. Their secret sauce wasn't just using a smart AI; it was like giving that AI a massive, specialized training camp. They took a 24-billion-parameter model (a huge brain) and fed it billions of tokens of historical French documents from before 1900, then fine-tuned it specifically on the competition's data. They even built a "judge-and-retry" loop, where the AI checks its own work, spots if it's hallucinating, and tries again up to three times.

The numbers show just how effective this was. On the test sets, the BnF-Mistral team reduced the error rate (called cMER) significantly. For example, on a noisy German dataset, they dropped the error rate from 0.0546 down to 0.0082. On a French dataset, they went from 0.0184 to 0.0040. In plain English, they turned a text that was barely readable into something almost perfect.

The Runners-Up: Other teams tried different strategies. One team, BLOCR, used a "zero-shot" approach, meaning they just asked a pre-trained AI to fix the text without any special training, relying on clever instructions. They did pretty well, especially on English texts, but they couldn't quite catch the heavily trained BnF-Mistral team. Another team, L3i, tried fine-tuning a smaller model, but it didn't perform as well as the heavyweights.

The "No-Change" Baseline: There was also a "do nothing" team. They just handed back the messy text as-is. Surprisingly, on the cleanest, least noisy documents, doing nothing was actually a decent strategy. Why? Because the fancy AIs sometimes got too eager. When the text was already 99% correct, the AI would sometimes "over-correct," changing a correct old-fashioned word into a modern one, or deleting a word it thought was a mistake. This taught the researchers a vital lesson: More AI isn't always better if the AI gets too confident.

What the Paper Rules Out (and What It Doesn't)

The paper is very clear about what this competition did not prove.

• It's not a "fix-all" solution: The paper explicitly states that while LLMs are great, they aren't a magic wand that solves every problem instantly. Performance varied wildly depending on the language, the type of document, and how messy the original scan was.
• It's not about perfect historical preservation: The scoring system was designed for searchability, not for keeping every tiny historical detail (like weird old punctuation or line breaks). The paper argues that for finding documents in a library, getting the words right is more important than keeping the layout perfect. So, if an AI smoothed out a weird hyphen to make a word searchable, that was a win, even if it wasn't "diplomatically" faithful to the original page.
• It's not solved yet: The authors emphasize that "over-correction" on clean text is still a recurring challenge. They didn't find a way to perfectly stop AIs from being too creative when they shouldn't be.

The Takeaway for a Curious Teen

Think of this competition like a car repair contest. The "cars" were millions of old, broken-down text transcripts. The "mechanics" were AI models.

• The BnF-Mistral team was like a mechanic who spent years studying old engines and brought a full toolkit. They fixed the cars so well they were almost new.
• The Zero-Shot teams were like mechanics who just showed up with a generic manual and said, "I'll try to fix it." They did a good job, but they couldn't match the specialist.
• The Baseline was the mechanic who said, "It's not that broken, I'll just leave it alone." And on the least broken cars, they were right!

The main finding is that specialized, well-trained AI can massively improve our access to history, turning unreadable scans into searchable text. But the paper warns us: we have to be careful. If we let these AIs run wild, they might start "fixing" things that weren't broken, changing history in the process. The future of digital libraries depends on finding that sweet spot where the AI is smart enough to fix the typos, but humble enough to leave the history alone.