Web Image Translation on Multilingual Web Image Translation ZH-EN
65.9BLEUVaaWIT on Qwen3 (8B)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| VaaWIT on Qwen3 (8B)Backbone=Qwen3 (8B), Tuning Strategy=DSAM + VAA, Trainable Parameters=50M2026.05 | 65.9 | 94.8 | |
| Full Fine-TuningBackbone=Qwen3-VL (8B), Tuning Strategy=Full Fine-Tuning, Trainable Parameters=8B2026.05 | 61.8 | 89.6 | |
| VaaWIT on LLaMA3.1 (8B)Backbone=LLaMA3.1 (8B), Tuning Strategy=DSAM + VAA, Trainable Parameters=50M2026.05 | 59.3 | 88.7 | |
| GPT4.1†Tuning Strategy=Zero-Shot, Evaluation Date=October 20252026.05 | 46.1 | 94.6 | |
| Gemini2.5 Pro†Tuning Strategy=Zero-Shot, Evaluation Date=October 20252026.05 | 40.1 | 90.1 | |
| Qwen3-VL (32B)Tuning Strategy=Zero-Shot, Parameters=32B2026.05 | 39.2 | 84.4 | |
| LoRABackbone=Qwen3-VL (8B), Tuning Strategy=LoRA (r=8), Trainable Parameters=30M2026.05 | 35 | 83.6 | |
| Chain-of-ThoughtBackbone=Qwen3-VL (8B), Tuning Strategy=Prompting2026.05 | 32.7 | 81.7 | |
| Qwen3-VL (8B)Tuning Strategy=Zero-Shot, Parameters=8B2026.05 | 32.3 | 81.3 | |
| EasyOCR_Google APIType=Cascaded Model2026.05 | 9.9 | 63.5 | |
| PP-OCR_Microsoft APIType=Cascaded Model2026.05 | 9.6 | 63.8 | |
| LLaMA3.2 (90B)Tuning Strategy=Zero-Shot, Parameters=90B2026.05 | 7.9 | 48.9 | |
| LLaMA3.2 (11B)Tuning Strategy=Zero-Shot, Parameters=11B2026.05 | 2.8 | 45.7 | |
| LLaVA-OV (7B)Tuning Strategy=Zero-Shot, Parameters=7B2026.05 | 1.7 | 42.6 |