Financial Question Answering on FinVQA text-only (test)
77.8Accuracy (English)Qwen3-VL-32B-Instruct
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Qwen3-VL-32B-InstructEvaluation Protocol=Zero shot inference2026.05 | 77.8 | 70.9 | 66.1 | 60.6 | 69.4 | 57.3 | |
| Qwen3-VL-32B-InstructEvaluation Protocol=Supervised finetuning2026.05 | 77.2 | 68.9 | 68.5 | 69.4 | 68 | 62.3 | |
| Qwen3-VL-32B-InstructEvaluation Protocol=Constrained decoding2026.05 | 76.9 | 68.9 | 52.1 | 33.6 | 67.2 | 61.2 | |
| Qwen3-VL-8B-InstructEvaluation Protocol=Zero shot inference2026.05 | 74.1 | 58.5 | 58.4 | 49.5 | 56.5 | 46.4 | |
| Qwen2.5-VL-32B-InstructEvaluation Protocol=Supervised finetuning2026.05 | 74.1 | 54.3 | 53.1 | 54.8 | 46.3 | 50.4 | |
| Qwen2.5-VL-32B-InstructEvaluation Protocol=Zero shot inference2026.05 | 73.8 | 40.6 | 27.9 | 22.6 | 34.4 | 33.9 | |
| Gemma-3-27b-itEvaluation Protocol=Zero shot inference2026.05 | 73.7 | 70.4 | 67.4 | 69.7 | 68.9 | 65.6 | |
| Gemma-3-12b-itEvaluation Protocol=Zero shot inference2026.05 | 71.7 | 62.7 | 66.6 | 63.4 | 64.3 | 57.8 | |
| Qwen2.5-VL-32B-InstructEvaluation Protocol=Constrained decoding2026.05 | 70.5 | 48.9 | 35.1 | 35.4 | 45 | 50.4 | |
| Qwen3-VL-8B-InstructEvaluation Protocol=Supervised finetuning2026.05 | 69.8 | 53.7 | 55.8 | 56.1 | 54 | 51.7 | |
| Qwen3-VL-4B-InstructEvaluation Protocol=Zero shot inference2026.05 | 69.6 | 51.2 | 48.5 | 26.2 | 39.6 | 17.5 | |
| Gemma-3-27b-itEvaluation Protocol=Constrained decoding2026.05 | 69.2 | 64 | 61.3 | 63.4 | 64.3 | 61.2 | |
| Qwen3-VL-8B-InstructEvaluation Protocol=Constrained decoding2026.05 | 69 | 53.4 | 45 | 28 | 53.6 | 52.1 | |
| Qwen3-VL-4B-InstructEvaluation Protocol=Supervised finetuning2026.05 | 67.7 | 50.7 | 51.3 | 47.8 | 49.5 | 47.6 | |
| Gemma-3-27b-itEvaluation Protocol=Supervised finetuning2026.05 | 67.7 | 63.9 | 59.7 | 63.6 | 64.7 | 59.4 | |
| Qwen3-VL-4B-InstructEvaluation Protocol=Constrained decoding2026.05 | 67.2 | 49.6 | 40.8 | 14.3 | 49.1 | 46.7 | |
| Qwen2.5-VL-3B-InstructEvaluation Protocol=Constrained decoding2026.05 | 63.1 | 36.9 | 20.9 | 29.4 | 32.4 | 20.2 | |
| Qwen2.5-VL-3B-InstructEvaluation Protocol=Supervised finetuning2026.05 | 62.6 | 35.7 | 35.8 | 30 | 33.2 | 36.4 | |
| Gemma-3-12b-itEvaluation Protocol=Constrained decoding2026.05 | 62.3 | 58.2 | 58.5 | 57.3 | 59.3 | 55.3 | |
| Gemma-3-12b-itEvaluation Protocol=Supervised finetuning2026.05 | 62.3 | 55.7 | 53.5 | 55.8 | 56.5 | 54.4 | |
| Gemma-3-4b-itEvaluation Protocol=Zero shot inference2026.05 | 61.5 | 50.3 | 52.2 | 50.4 | 41.3 | 45.2 | |
| Qwen2.5-VL-7B-InstructEvaluation Protocol=Constrained decoding2026.05 | 61.2 | 39.8 | 34.2 | 18.3 | 34.1 | 40.1 | |
| Qwen2.5-VL-7B-InstructEvaluation Protocol=Supervised finetuning2026.05 | 61.1 | 34.1 | 41.5 | 34.5 | 30.6 | 41.1 | |
| Qwen2.5-VL-3B-InstructEvaluation Protocol=Zero shot inference2026.05 | 57.5 | 26.7 | 27.3 | 14.3 | 25.4 | 21.2 | |
| Gemma-3-4b-itEvaluation Protocol=Constrained decoding2026.05 | 56.3 | 46.3 | 49.1 | 48.4 | 49.6 | 40.9 | |
| Gemma-3-4b-itEvaluation Protocol=Supervised finetuning2026.05 | 56.3 | 47.8 | 48.9 | 47 | 50.3 | 44.7 | |
| Qwen2.5-VL-7B-InstructEvaluation Protocol=Zero shot inference2026.05 | 55.6 | 36.6 | 37.4 | 20.5 | 31.3 | 27.6 |