Web Image Translation on Multilingual Web Image Translation JA-EN
55.3BLEUVaaWIT on LLaMA3.1 (8B)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| VaaWIT on LLaMA3.1 (8B)Backbone=LLaMA3.1 (8B), Tuning Strategy=DSAM + VAA2026.05 | 55.3 | 94.6 | |
| VaaWIT on Qwen3 (8B)Backbone=Qwen3 (8B), Tuning Strategy=DSAM + VAA2026.05 | 47.2 | 86.6 | |
| Full Fine-TuningBackbone=Qwen3-VL (8B), Tuning Strategy=Full Fine-Tuning2026.05 | 45 | 84.3 | |
| LoRABackbone=Qwen3-VL (8B), Tuning Strategy=LoRA (r=8)2026.05 | 33.1 | 73.9 | |
| Qwen3-VL (32B)Tuning Strategy=Zero-Shot, Parameters=32B2026.05 | 26.9 | 65.8 | |
| Chain-of-ThoughtBackbone=Qwen3-VL (8B), Tuning Strategy=Prompting2026.05 | 26.2 | 67 | |
| Qwen3-VL (8B)Tuning Strategy=Zero-Shot, Parameters=8B2026.05 | 26.1 | 67.2 | |
| GPT4.1†Tuning Strategy=Zero-Shot2026.05 | 25.8 | 68.2 | |
| Gemini2.5 Pro†Tuning Strategy=Zero-Shot2026.05 | 20.9 | 66.1 | |
| LLaMA3.2 (90B)Tuning Strategy=Zero-Shot, Parameters=90B2026.05 | 14.8 | 49.3 | |
| LLaVA-OV (7B)Tuning Strategy=Zero-Shot, Parameters=7B2026.05 | 11.2 | 47.5 | |
| EasyOCR_Google APIType=Cascaded Model2026.05 | 9.1 | 56.8 | |
| PP-OCR_Microsoft APIType=Cascaded Model2026.05 | 5.2 | 46.6 | |
| LLaMA3.2 (11B)Tuning Strategy=Zero-Shot, Parameters=11B2026.05 | 3.1 | 43.1 |