ScriptMoE tops VLM baseline on 229-language OCR benchmark
TL;DR
- ScriptMoE handles 229 languages across 10 scripts in one recognizer, using a sparse Mixture-of-Experts decoder that routes each image to top-2 script-aligned experts.
- On the authors' TextMuSS-Bench (10 scripts, 10,899 images), the model reaches 82.06% accuracy, 1.31% above the strongest scene-text-recognition baseline.
- Dropping ScriptMoE into PP-OCRv5 lifts CC-OCR end-to-end F1 from 65.71% to 80.89%, slightly ahead of the best VLM at 80.73%.
A single recognizer spanning 229 languages across 10 scripts reaches 82.06% accuracy on the authors' TextMuSS-Bench, edging the strongest scene-text baseline by 1.31%, according to a paper posted to arXiv by Xingsong Ye, Zhineng Chen and colleagues.
The design keeps a single visual encoder and replaces the dense decoder with a sparse Mixture-of-Experts block. An image-level router "dispatches each image to the top-2 script-aligned experts" while a shared expert "absorbs cross-script knowledge," per the abstract. Training runs on TextMuSS-10M, a new synthetic dataset the authors built to give "balanced and sufficient supervision where real data is unavailable."
The end-to-end number is the more striking one. Swapping ScriptMoE for the recognizer inside PP-OCRv5 lifts CC-OCR multilingual F1 from 65.71% to 80.89%, "slightly surpassing the best VLM (80.73%) at a fraction of the parameter count." No parameter figure appears in the abstract, so the size gap is a claim rather than a measured number.
Originally reported by paper
Read the original article →Original headline: ScriptMoE Reads 229 Languages in One Model, Edges Best VLM on Multilingual OCR at a Fraction of Parameters