Multimodal Financial Question Answering on FinVQA multimodal
69.4Accuracy (en)Qwen3-VL-32B-Instruct
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Qwen3-VL-32B-InstructInference Mode=Supervised finetuning, Scale=27-32B2026.05 | 69.4 | 67.7 | 66.1 | 59.9 | 60.1 | 53.4 | |
| Gemma-3-4b-itInference Mode=Zero shot inference, Scale=27-32B2026.05 | 69.3 | 62.4 | 59.9 | 55.6 | 61.9 | 63.5 | |
| Qwen3-VL-32B-InstructInference Mode=Zero shot inference, Scale=27-32B2026.05 | 68.3 | 51.8 | 49.1 | 31.3 | 49.1 | 45.9 | |
| Qwen2.5-VL-32B-InstructInference Mode=Zero shot inference, Scale=27-32B2026.05 | 67.2 | 36 | 22.5 | 9 | 32.8 | 30.1 | |
| Gemma-3-4b-itInference Mode=Zero shot inference, Scale=7-12B2026.05 | 66.7 | 55 | 60.4 | 58.7 | 60.3 | 58.2 | |
| Qwen3-VL-8B-InstructInference Mode=Zero shot inference, Scale=7-12B2026.05 | 64.6 | 49.2 | 47.6 | 27.5 | 44.4 | 41.8 | |
| Qwen2.5-VL-32B-InstructInference Mode=Supervised finetuning, Scale=27-32B2026.05 | 64.1 | 57.1 | 54.1 | 59.1 | 52.3 | 59.3 | |
| Qwen3-VL-32B-InstructInference Mode=Constrained decoding, Scale=27-32B2026.05 | 62.4 | 63 | 65.2 | 58.7 | 61.9 | 53.4 | |
| Qwen3-VL-4B-InstructInference Mode=Zero shot inference, Scale=3-4B2026.05 | 61.9 | 46 | 47.1 | 19 | 37 | 12.2 | |
| Qwen2.5-VL-7B-InstructInference Mode=Zero shot inference, Scale=7-12B2026.05 | 61.4 | 46.6 | 49.2 | 26.5 | 42.9 | 45.5 | |
| Gemma-3-4b-itInference Mode=Supervised finetuning, Scale=27-32B2026.05 | 61.1 | 55.5 | 57.3 | 54.1 | 58.1 | 50.1 | |
| Gemma-3-4b-itInference Mode=Constrained decoding, Scale=7-12B2026.05 | 59.3 | 52.4 | 53.5 | 52.4 | 54.5 | 54.5 | |
| Gemma-3-4b-itInference Mode=Constrained decoding, Scale=27-32B2026.05 | 59.3 | 52.9 | 58.8 | 52.9 | 59.8 | 50.8 | |
| Qwen2.5-VL-32B-InstructInference Mode=Constrained decoding, Scale=27-32B2026.05 | 58.7 | 55.6 | 54.5 | 54.5 | 48.1 | 55 | |
| Qwen3-VL-8B-InstructInference Mode=Constrained decoding, Scale=7-12B2026.05 | 58.2 | 55 | 57.8 | 52.9 | 56.6 | 51.3 | |
| Qwen3-VL-8B-InstructInference Mode=Supervised finetuning, Scale=7-12B2026.05 | 58.2 | 55.6 | 57.8 | 53.4 | 57.7 | 51.3 | |
| Gemma-3-4b-itInference Mode=Supervised finetuning, Scale=7-12B2026.05 | 57.1 | 52.4 | 55.1 | 55 | 55 | 55.6 | |
| Qwen2.5-VL-3B-InstructInference Mode=Supervised finetuning, Scale=3-4B2026.05 | 55.6 | 40.7 | 43.3 | 24.9 | 36.5 | 42.9 | |
| Qwen3-VL-4B-InstructInference Mode=Constrained decoding, Scale=3-4B2026.05 | 55.6 | 55 | 62 | 50.8 | 60.8 | 46 | |
| Qwen2.5-VL-3B-InstructInference Mode=Constrained decoding, Scale=3-4B2026.05 | 54.5 | 41.8 | 46.5 | 24.3 | 35.4 | 40.2 | |
| Qwen3-VL-4B-InstructInference Mode=Supervised finetuning, Scale=3-4B2026.05 | 54.5 | 55 | 64.7 | 51.3 | 59.3 | 48.1 | |
| Qwen2.5-VL-7B-InstructInference Mode=Supervised finetuning, Scale=7-12B2026.05 | 54.5 | 40.7 | 50.8 | 45.5 | 42.9 | 49.7 | |
| Gemma-3-4b-itInference Mode=Constrained decoding, Scale=3-4B2026.05 | 53.4 | 52.9 | 54.5 | 44.4 | 54 | 29.6 | |
| Qwen2.5-VL-7B-InstructInference Mode=Constrained decoding, Scale=7-12B2026.05 | 53.4 | 41.3 | 52.4 | 47.6 | 47.6 | 49.2 | |
| Qwen2.5-VL-3B-InstructInference Mode=Zero shot inference, Scale=3-4B2026.05 | 52.4 | 33.9 | 31 | 13.2 | 25.9 | 22.2 | |
| Gemma-3-4b-itInference Mode=Supervised finetuning, Scale=3-4B2026.05 | 51.9 | 50.8 | 51.3 | 46 | 56.1 | 31.7 | |
| Gemma-3-4b-itInference Mode=Zero shot inference, Scale=3-4B2026.05 | 37.6 | 44.4 | 42.8 | 41.8 | 43.4 | 30.2 |