Visual Question Answering on VQA, GQA, SQA, and VQAText IMG v2
80.9VQAv2 AccuracyCogVLM-chat-1.1-17B
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| CogVLM-chat-1.1-17BBackbone=CogVLM-chat-1.1-17B, Token Retention=all 1225 tokens (100%)2026.02 | 80.9 | 58.2 | 68.4 | 69.2 | 69.2 | 100 | |
| InternVL-Chat-13BBackbone=InternVL-Chat-13B, Token Retention=all 576 tokens (100%)2026.02 | 79.3 | 62.9 | 66.3 | 57 | 66.4 | 100 | |
| Qwen-VL-7BBackbone=Qwen-VL-7B, Token Retention=all 256 tokens (100%)2026.02 | 78.8 | 59.3 | 67.1 | 63.8 | 67.3 | 100 | |
| ConsensusDropBackbone=Qwen-VL-7B, Token Retention=128 tokens (50%)2026.02 | 78.1 | 58.9 | 66.4 | 61.1 | 66.1 | 98.2 | |
| VisPrunerBackbone=Qwen-VL-7B, Token Retention=128 tokens (50%)2026.02 | 77.4 | 57.8 | 65.9 | 59.6 | 65.2 | 96.9 | |
| ConsensusDropBackbone=InternVL-Chat-13B, Token Retention=144 tokens (75%)2026.02 | 76.9 | 60.5 | 67.9 | 56.1 | 65.4 | 98.5 | |
| VisPrunerBackbone=InternVL-Chat-13B, Token Retention=144 tokens (75%)2026.02 | 76.7 | 60.2 | 67.3 | 55.3 | 64.9 | 97.7 | |
| FastVBackbone=Qwen-VL-7B, Token Retention=128 tokens (50%)2026.02 | 76.5 | 56.9 | 65.3 | 58.2 | 64.2 | 95.4 | |
| ConsensusDropBackbone=CogVLM-chat-1.1-17B, Token Retention=123 tokens (90%)2026.02 | 75.1 | 49.3 | 69.6 | 59.5 | 63.4 | 91.6 | |
| VisPrunerBackbone=CogVLM-chat-1.1-17B, Token Retention=123 tokens (90%)2026.02 | 74.6 | 48.4 | 68.2 | 59.6 | 62.7 | 90.6 | |
| FastVBackbone=CogVLM-chat-1.1-17B, Token Retention=123 tokens (90%)2026.02 | 74.2 | 40.3 | 63.5 | 41.9 | 55 | 79.5 | |
| FastVBackbone=InternVL-Chat-13B, Token Retention=144 tokens (75%)2026.02 | 74.1 | 58.2 | 66.6 | 55.6 | 63.6 | 95.8 |