Fudan's ScriptMoE recognizes text in 229 languages, beating VLMs with far fewer parameters
Fudan-University · hf · 2026-09-23
Fudan University released an all-in-one multilingual scene text recognition (STR) system built from two components:
- TextMuSS-10M: a large-scale synthetic scene text dataset spanning 10 scripts and 229 languages, providing supervision where real data is scarce.
- ScriptMoE: a script-aware Mixture-of-Experts architecture that shares one visual encoder and replaces the dense decoder with a sparse MoE block — an image-level router dispatches each image to the top-2 script-aligned experts, plus a shared expert for cross-script knowledge.
Results:
- 82.06% accuracy on their TextMuSS-Bench (10 scripts, 10,899 images), beating the strongest STR baseline by 1.31 points.
- Swapping only the recognizer in PP-OCRv5 lifts the CC-OCR end-to-end multilingual F1 from 65.71% to 80.89%, narrowly beating the best VLM (80.73%) at a fraction of the parameter count.
Positioned as simpler than per-language experts and lighter yet more accurate than VLMs.
More from Research
- TensorSharp's logit-reading approach beats LocalJev at structured decisions, 3.3x faster — fuzhongkai · 2026-09-23
- Naproche: a proof assistant that reads math proofs written in controlled natural language — zetalyrae · 2026-09-23
- Nanjing University used AI to design proteins 4x stronger than any natural protein — MikePFrank · 2026-09-23
- Open-source GUI agents top out at 8% task success on composite cross-device tasks — maier_ak · 2026-09-23
- JarvisGUI benchmark tests GUI agents across Android, Windows and Ubuntu in one workflow — maier_ak · 2026-09-23
- Cross-DURIAN: A New Benchmark for Multi-Device GUI Agents — maier_ak · 2026-09-23