Document Visual Question Answering on DocVQA (ANLS)
96.4ANLSFull Precision
Evaluation Results
| Method | Links | |
|---|---|---|
| Full PrecisionModel=Qwen2.5-VL-72B-Instruct, Weight bits=16 bit, Memory=136.74 GB, Evaluation protocol=zero-shot2026.07 | 96.4 | |
| Full PrecisionModel=Qwen2.5-VL-7B-Instruct, Weight bits=16 bit, Memory=15.45 GB, Evaluation protocol=zero-shot2026.07 | 95.17 | |
| Full PrecisionModel=Qwen2.5-VL-32B-Instruct, Weight bits=16 bit, Memory=62.31 GB, Evaluation protocol=zero-shot2026.07 | 94.8 | |
| GPTQModel=Qwen2.5-VL-72B-Instruct, Weight bits=3 bit, Memory=28.61 GB, Evaluation protocol=zero-shot2026.07 | 94.25 | |
| GPTQModel=Qwen2.5-VL-7B-Instruct, Weight bits=3 bit, Memory=2.66 GB, Evaluation protocol=zero-shot2026.07 | 93.36 | |
| SPOT-EBase Model=Gemini-2.5-Flash, Source Type=Closed-Source, Test-time Adaptation=SPOT-E2026.06 | 93 | |
| GPTQModel=Qwen2.5-VL-32B-Instruct, Weight bits=3 bit, Memory=12.71 GB, Evaluation protocol=zero-shot2026.07 | 92.65 | |
| SAB-LVLMModel=Qwen2.5-VL-72B-Instruct, Weight bits=1.07 bit, Memory=23.23 GB, Evaluation protocol=zero-shot2026.07 | 92.56 | |
| SPOT-EBase Model=GPT-4o, Source Type=Closed-Source, Test-time Adaptation=SPOT-E2026.06 | 92.3 | |
| Gemini-2.5-FlashBase Model=Gemini-2.5-Flash, Source Type=Closed-Source, Test-time Adaptation=None2026.06 | 91.5 | |
| GPT-4oBase Model=GPT-4o, Source Type=Closed-Source, Test-time Adaptation=None2026.06 | 91.1 | |
| ARB-LLMModel=Qwen2.5-VL-72B-Instruct, Weight bits=1.19 bit, Memory=23.23 GB, Evaluation protocol=zero-shot2026.07 | 90.42 | |
| SAB-LVLMModel=Qwen2.5-VL-32B-Instruct, Weight bits=1.07 bit, Memory=10.28 GB, Evaluation protocol=zero-shot2026.07 | 90.36 | |
| BiLLMModel=Qwen2.5-VL-72B-Instruct, Weight bits=1.08 bit, Memory=24.23 GB, Evaluation protocol=zero-shot2026.07 | 90.1 | |
| ARB-LLMModel=Qwen2.5-VL-32B-Instruct, Weight bits=1.16 bit, Memory=10.28 GB, Evaluation protocol=zero-shot2026.07 | 90 | |
| SPOT-EBase Model=GPT-4o-mini, Source Type=Closed-Source, Test-time Adaptation=SPOT-E2026.06 | 88 | |
| SPOT-EBase Model=Qwen3-VL-8B, Source Type=Open-Source, Test-time Adaptation=SPOT-E2026.06 | 86.8 | |
| SPOT-EBase Model=Qwen2.5-VL-7B, Source Type=Open-Source, Test-time Adaptation=SPOT-E2026.06 | 86.5 | |
| Qwen3-VL-8BBase Model=Qwen3-VL-8B, Source Type=Open-Source, Test-time Adaptation=None2026.06 | 86.2 | |
| GPT-4o-miniBase Model=GPT-4o-mini, Source Type=Closed-Source, Test-time Adaptation=None2026.06 | 86 | |
| BaselineKeep ratio=100%, FLOPs Saved=—2026.05 | 85.91 | |
| AsymVLMKeep ratio=75%, FLOPs Saved=28%2026.05 | 85.75 | |
| Qwen2.5-VL-7BBase Model=Qwen2.5-VL-7B, Source Type=Open-Source, Test-time Adaptation=None2026.06 | 85.7 | |
| SAB-LVLMModel=Qwen2.5-VL-7B-Instruct, Weight bits=1.07 bit, Memory=2.14 GB, Evaluation protocol=zero-shot2026.07 | 85.34 | |
| PB-LLMModel=Qwen2.5-VL-72B-Instruct, Weight bits=1.70 bit, Memory=22.99 GB, Evaluation protocol=zero-shot2026.07 | 85.33 | |
| AsymVLMKeep ratio=65%, FLOPs Saved=39%2026.05 | 85.07 | |
| SparseVLMKeep ratio=75%, FLOPs Saved=18%2026.05 | 84.58 | |
| SparseVLMKeep ratio=65%, FLOPs Saved=35%2026.05 | 84.4 | |
| SPOT-EBase Model=InternVL3-8B, Source Type=Open-Source, Test-time Adaptation=SPOT-E2026.06 | 83.7 | |
| FastVKeep ratio=75%, FLOPs Saved=0%2026.05 | 83.53 | |
| SPOT-EBase Model=InternVL2.5-8B, Source Type=Open-Source, Test-time Adaptation=SPOT-E2026.06 | 83.5 | |
| SparseVLMKeep ratio=50%, FLOPs Saved=47%2026.05 | 83.38 | |
| ChatVLA2026.06 | 83.3 | |
| AsymVLMKeep ratio=50%, FLOPs Saved=54%2026.05 | 83.22 | |
| SPOT-EBase Model=LLaVA-OV-7B, Source Type=Open-Source, Test-time Adaptation=SPOT-E2026.06 | 82.8 | |
| InternVL3-8BBase Model=InternVL3-8B, Source Type=Open-Source, Test-time Adaptation=None2026.06 | 82.7 | |
| SPOT-EBase Model=LLaVA-NeXT-7B, Source Type=Open-Source, Test-time Adaptation=SPOT-E2026.06 | 82 | |
| InternVL2.5-8BBase Model=InternVL2.5-8B, Source Type=Open-Source, Test-time Adaptation=None2026.06 | 82 | |
| FastVKeep ratio=65%, FLOPs Saved=0%2026.05 | 81.74 | |
| LLaVA-OV-7BBase Model=LLaVA-OV-7B, Source Type=Open-Source, Test-time Adaptation=None2026.06 | 81 | |
| ARB-LLMModel=Qwen2.5-VL-7B-Instruct, Weight bits=1.07 bit, Memory=2.14 GB, Evaluation protocol=zero-shot2026.07 | 80.48 | |
| SyVLA2026.06 | 80.01 | |
| LLaVA-NeXT-7BBase Model=LLaVA-NeXT-7B, Source Type=Open-Source, Test-time Adaptation=None2026.06 | 80 | |
| FastVKeep ratio=50%, FLOPs Saved=0%2026.05 | 77.3 | |
| PB-LLMModel=Qwen2.5-VL-32B-Instruct, Weight bits=1.70 bit, Memory=10.24 GB, Evaluation protocol=zero-shot2026.07 | 66.1 | |
| Wall-Oss2026.06 | 63.62 | |
| BiLLMModel=Qwen2.5-VL-32B-Instruct, Weight bits=1.08 bit, Memory=10.73 GB, Evaluation protocol=zero-shot2026.07 | 60.85 | |
| BiLLMModel=Qwen2.5-VL-7B-Instruct, Weight bits=1.08 bit, Memory=2.24 GB, Evaluation protocol=zero-shot2026.07 | 33.08 | |
| PB-LLMModel=Qwen2.5-VL-7B-Instruct, Weight bits=1.17 bit, Memory=2.15 GB, Evaluation protocol=zero-shot2026.07 | 9.4 |