OCR models keep writing the right word in the wrong alphabet — and frontier models do it too
VikParuchuri · x · 2026-10-16
Datalab founder Vik Paruchuri investigates "script confusions": a rare OCR failure where the model outputs the correct word in the wrong writing system. On an old hymnal page, "vain" came out as the Chinese character 罪 (= sin) — the model didn't misread, it effectively translated into the wrong script. Other cases include Finnish words ending in Arabic or Greek letters.
- Testing showed the bug affects every frontier model tried, including Gemini 3.8 Flash and Claude Opus 5.5
- Errors are rare (a fraction of a percent of pages) but sneaky: edit-distance metrics barely move and searching for the right word finds nothing
- Best explanation: inside the model, same-meaning words across scripts share nearly identical embeddings ("cruel" sits close to 残酷)
- Fix: a novel training technique shipped in Chandra 2.5, cutting script confusions 7x vs Chandra OCR 2.4 on a hard test set, beating both frontier models
Related event: Datalab founder explains why VLMs confuse scripts during OCR(5 posts)→
More from Models
- NEURA launches Neuron AI, a multimodal PLA model giving robots touch, force feedback and memory — CyberRobooo · 2026-10-17
- Sentience claims per-user models: train, continually update and serve one model per person — soleio · 2026-10-17
- Four Gemini 4 Argon agents beat Minecraft in 27 hours, writing 2,783 programs — GoogleDeepMind · 2026-10-17
- Meta open-sources Muse Spark Open to mixed reception: 'trash' license, nerfed benchmark scores — gabriberton · 2026-10-17
- If new diffusion models are open-weight only, what's left for commercial use? — blackpanda8484 · 2026-10-17
- Reflection Beam OSS architecture: sliding window attention, gated attention, QK Norm — A_K_Nain · 2026-10-17