Idiom-based vision-language matching (Subtask B) on AdMIRe multilingual (15 languages) 2.0 (test)
32Top-1 AccBest submission
Evaluation Results
| Method | Links | |
|---|---|---|
| Best submissionSystem architecture=SigLIP2, Idiom replacement=true, Zero-shot classification=true2026.02 | 32 | |
| Ensemble (Qwen3+GPT-4o)System architecture=Qwen3 + GPT-4o ensemble2026.02 | 29 | |
| GPT-4o2026.02 | 28 | |
| Qwen3-32B2026.02 | 26.5 | |
| Mistral-large2026.02 | 25.5 | |
| Llama3.1-70B2026.02 | 25 | |
| DeepSeek-v32026.02 | 23.5 | |
| CLIP baselineSystem architecture=XLM-RoBERTa-ViT-H-142026.02 | 6.7 |