Visual Question Answering on TVQA, InfoVQA, ChartQA, DocVQA, and VizWiz
82.92TVQA AccuracyBaseline
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| BaselineBackbone=QWEN2.5-VL-7B-INSTRUCT, Token (budget ratio)=1.00×, Channel (budget ratio)=1.00×, KV Cache (budget ratio)=1.00×2026.05 | 82.92 | 80.12 | 83.04 | 94.4 | 70.58 | 82.21 | |
| FastV + RotateKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token (budget ratio)=0.40×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.25×2026.05 | 82.38 | 65.79 | 77.4 | 87.28 | 70.32 | 76.63 | |
| FastV + RotateKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token (budget ratio)=0.30×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.19×2026.05 | 81.8 | 56.82 | 71.32 | 80.13 | 69.86 | 71.99 | |
| FastVtoken onlyBackbone=QWEN2.5-VL-7B-INSTRUCT, Token (budget ratio)=0.25×, Channel (budget ratio)=1.00×, KV Cache (budget ratio)=0.25×2026.05 | 81.49 | 52.76 | 67.44 | 75.04 | 69.13 | 69.17 | |
| VisionZip + RotateKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token (budget ratio)=0.45×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.28×2026.05 | 80.84 | 70.51 | 79.56 | 91.29 | 70.13 | 78.47 | |
| FastVtoken onlyBackbone=QWEN2.5-VL-7B-INSTRUCT, Token (budget ratio)=0.20×, Channel (budget ratio)=1.00×, KV Cache (budget ratio)=0.20×2026.05 | 80.44 | 46.73 | 63.08 | 66.06 | 68.34 | 64.93 | |
| VisionZip + RotateKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token (budget ratio)=0.35×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.22×2026.05 | 78.48 | 63.86 | 75.84 | 87.91 | 69.38 | 75.09 | |
| FastV + ThinKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token (budget ratio)=0.40×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.25×2026.05 | 77.81 | 63.01 | 77.12 | 81.74 | 69.25 | 73.79 | |
| FastV + ThinKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token (budget ratio)=0.30×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.19×2026.05 | 77.76 | 54.79 | 71.4 | 75.58 | 68.82 | 69.67 | |
| VisionZiptoken onlyBackbone=QWEN2.5-VL-7B-INSTRUCT, Token (budget ratio)=0.28×, Channel (budget ratio)=1.00×, KV Cache (budget ratio)=0.28×2026.05 | 76.84 | 58.46 | 70.32 | 84.71 | 68.97 | 71.86 | |
| VisionZip + ThinKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token (budget ratio)=0.45×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.28×2026.05 | 75.72 | 67.13 | 79.64 | 85.83 | 68.93 | 75.45 | |
| VisionZiptoken onlyBackbone=QWEN2.5-VL-7B-INSTRUCT, Token (budget ratio)=0.22×, Channel (budget ratio)=1.00×, KV Cache (budget ratio)=0.22×2026.05 | 74.19 | 50.58 | 64.4 | 78.01 | 68.84 | 67.2 | |
| VisionZip + ThinKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token (budget ratio)=0.35×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.22×2026.05 | 73.83 | 60.95 | 76.44 | 82.78 | 68.35 | 72.47 | |
| BaselineBackbone=LLAMA3-LLAVA-NEXT-8B, Token (budget ratio)=1.00×, Channel (budget ratio)=1.00×, KV Cache (budget ratio)=1.00×2026.05 | 65.4 | 32.27 | 69.08 | 72.44 | 57.5 | 59.34 | |
| VisionZiptoken onlyBackbone=LLAMA3-LLAVA-NEXT-8B, Token (budget ratio)=0.28×, Channel (budget ratio)=1.00×, KV Cache (budget ratio)=0.28×2026.05 | 63.33 | 27.7 | 57.32 | 64.08 | 58.97 | 54.28 | |
| VisionZip + RotateKBackbone=LLAMA3-LLAVA-NEXT-8B, Token (budget ratio)=0.35×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.22×2026.05 | 62.73 | 27.7 | 60.16 | 64.73 | 59.33 | 54.93 | |
| FastVtoken onlyBackbone=LLAMA3-LLAVA-NEXT-8B, Token (budget ratio)=0.25×, Channel (budget ratio)=1.00×, KV Cache (budget ratio)=0.25×2026.05 | 62.69 | 25.91 | 59.64 | 56.39 | 58.56 | 52.64 | |
| FastV + RotateKBackbone=LLAMA3-LLAVA-NEXT-8B, Token (budget ratio)=0.40×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.25×2026.05 | 62.67 | 28.37 | 65.25 | 62.46 | 58.34 | 55.42 | |
| VisionZip + RotateKBackbone=LLAMA3-LLAVA-NEXT-8B, Token (budget ratio)=0.45×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.28×2026.05 | 62.62 | 28.97 | 62.88 | 66.99 | 58.94 | 56.08 | |
| VisionZiptoken onlyBackbone=LLAMA3-LLAVA-NEXT-8B, Token (budget ratio)=0.22×, Channel (budget ratio)=1.00×, KV Cache (budget ratio)=0.22×2026.05 | 62.07 | 26.11 | 55.16 | 59.18 | 58.95 | 52.29 | |
| FastV + RotateKBackbone=LLAMA3-LLAVA-NEXT-8B, Token (budget ratio)=0.30×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.19×2026.05 | 61.53 | 25.4 | 62.36 | 57.59 | 58.53 | 53.08 | |
| FastVtoken onlyBackbone=LLAMA3-LLAVA-NEXT-8B, Token (budget ratio)=0.20×, Channel (budget ratio)=1.00×, KV Cache (budget ratio)=0.20×2026.05 | 61.03 | 24.25 | 54.28 | 51.23 | 58.16 | 49.79 | |
| VisionZip + ThinKBackbone=LLAMA3-LLAVA-NEXT-8B, Token (budget ratio)=0.35×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.22×2026.05 | 54.24 | 27.81 | 55.56 | 50.76 | 58.2 | 49.31 | |
| VisionZip + ThinKBackbone=LLAMA3-LLAVA-NEXT-8B, Token (budget ratio)=0.45×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.28×2026.05 | 54.22 | 27.81 | 58.52 | 53.77 | 57.81 | 50.43 | |
| FastV + ThinKBackbone=LLAMA3-LLAVA-NEXT-8B, Token (budget ratio)=0.40×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.25×2026.05 | 54.07 | 25.51 | 60.56 | 51.72 | 57.7 | 49.91 | |
| FastV + ThinKBackbone=LLAMA3-LLAVA-NEXT-8B, Token (budget ratio)=0.30×, Channel (budget ratio)=0.25×, KV Cache (budget ratio)=0.19×2026.05 | 54.07 | 23.84 | 58.76 | 47.33 | 57.41 | 48.28 |