Scientific Question Answering on SciQA
99.1AccuracyDDP
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| DDPbackbone=Gemini-3-Pro2026.04 | 99.1 | — | — | — | — | — | — | — | — | |
| Gemini-3-pro2026.04 | 98.1 | — | — | — | — | — | — | — | — | |
| Gemini-1.5-pro2026.04 | 94.8 | — | — | — | — | — | — | — | — | |
| GPT-4o2026.04 | 94.3 | — | — | — | — | — | — | — | — | |
| Qwen3-VL-4B-Instruct2026.02 | 92.7 | — | — | — | — | — | — | — | — | |
| SwiftInference Mode=Thinking Mode2025.06 | 92.48 | 99.1 | 7.61 | 8.84 | 7.73 | 6.76 | 7.11 | 8.45 | — | |
| RefitInference Mode=Thinking Mode2025.06 | 91.59 | 96 | 7.64 | 8.87 | 7.76 | 6.81 | 7.13 | 8.43 | — | |
| Qwen2.5-VL-32BToken Budget=100%, Backbone Model=Qwen2.5-VL-32B2026.04 | 91.4 | — | — | — | — | — | — | — | — | |
| MolmoAct2026.02 | 91.1 | — | — | — | — | — | — | — | — | |
| STaR-GATE-DInference Mode=Thinking Mode2025.06 | 90.27 | 21.7 | 6.58 | 8.19 | 6.72 | 5.78 | 5.73 | 7.24 | — | |
| IWPToken Budget=35.3%, Backbone Model=Qwen2.5-VL-32B2026.04 | 90 | — | — | — | — | — | — | — | — | |
| DivPruneToken Budget=35.3%, Backbone Model=Qwen2.5-VL-32B2026.04 | 88.9 | — | — | — | — | — | — | — | — | |
| VisionZipToken Budget=35.3%, Backbone Model=Qwen2.5-VL-32B2026.04 | 88.7 | — | — | — | — | — | — | — | — | |
| IWPToken Budget=22.2%, Backbone Model=Qwen2.5-VL-32B2026.04 | 87.6 | — | — | — | — | — | — | — | — | |
| VisionZipToken Budget=22.2%, Backbone Model=Qwen2.5-VL-32B2026.04 | 87.6 | — | — | — | — | — | — | — | — | |
| DivPruneToken Budget=22.2%, Backbone Model=Qwen2.5-VL-32B2026.04 | 86.8 | — | — | — | — | — | — | — | — | |
| CDPrunerToken Budget=35.3%, Backbone Model=Qwen2.5-VL-32B2026.04 | 85.2 | — | — | — | — | — | — | — | — | |
| FastVToken Budget=35.3%, Backbone Model=Qwen2.5-VL-32B2026.04 | 85 | — | — | — | — | — | — | — | — | |
| VisionZipToken Budget=11.1%, Backbone Model=Qwen2.5-VL-32B2026.04 | 84.5 | — | — | — | — | — | — | — | — | |
| IWPToken Budget=11.1%, Backbone Model=Qwen2.5-VL-32B2026.04 | 83.5 | — | — | — | — | — | — | — | — | |
| FastVToken Budget=22.2%, Backbone Model=Qwen2.5-VL-32B2026.04 | 83.2 | — | — | — | — | — | — | — | — | |
| CDPrunerToken Budget=22.2%, Backbone Model=Qwen2.5-VL-32B2026.04 | 83.1 | — | — | — | — | — | — | — | — | |
| DivPruneToken Budget=11.1%, Backbone Model=Qwen2.5-VL-32B2026.04 | 82.5 | — | — | — | — | — | — | — | — | |
| STaR-GATEInference Mode=Thinking Mode2025.06 | 81.86 | 90.3 | 7.08 | 8.17 | 7.27 | 6.49 | 6.69 | 7.69 | — | |
| StepDPOInference Mode=Thinking Mode2025.06 | 81.86 | 7.5 | 6.29 | 7.87 | 6.36 | 5.57 | 5.42 | 6.89 | — | |
| FastVToken Budget=11.1%, Backbone Model=Qwen2.5-VL-32B2026.04 | 80.2 | — | — | — | — | — | — | — | — | |
| Xiaomi-Robotics-0pre-training=with VL data2026.02 | 79.4 | — | — | — | — | — | — | — | — | |
| CDPrunerToken Budget=11.1%, Backbone Model=Qwen2.5-VL-32B2026.04 | 79.3 | — | — | — | — | — | — | — | — | |
| DPOInference Mode=Thinking Mode2025.06 | 79.2 | 5.8 | 5.96 | 7.61 | 6.08 | 5.29 | 5.14 | 6.5 | — | |
| Qwen-VL2026.04 | 67.1 | — | — | — | — | — | — | — | — | |
| LLaVA-v1.52026.04 | 66.8 | — | — | — | — | — | — | — | — | |
| BaseInference Mode=Thinking Mode2025.06 | 49.56 | 73.5 | 7 | 8.12 | 7.03 | 6.11 | 6.84 | 7.78 | — | |
| pi_0.52026.02 | 28 | — | — | — | — | — | — | — | — | |
| pi_02026.02 | 0 | — | — | — | — | — | — | — | — | |
| Xiaomi-Robotics-0pre-training=without VL data2026.02 | 0 | — | — | — | — | — | — | — | — | |
| ATSBackbone=Dynamic-LLaVA-7B, Token reduction=25%2026.04 | — | — | — | — | — | — | — | — | 63 | |
| Bunny-2BLLM=Qwen1.5-1.8B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 64.6 | |
| DASHBackbone=Dynamic-LLaVA-7B, Token reduction=25%2026.04 | — | — | — | — | — | — | — | — | 64.9 | |
| Dynamic-LLaVA-7B_IBackbone=Dynamic-LLaVA-7B, Token reduction=25%2026.04 | — | — | — | — | — | — | — | — | 69.1 | |
| EViTBackbone=Dynamic-LLaVA-7B, Token reduction=25%2026.04 | — | — | — | — | — | — | — | — | 64.1 | |
| FastVLLM=Vicuna-7B, Retain Ratio (%)=∼ 10%2026.05 | — | — | — | — | — | — | — | — | 51.1 | |
| Imp-2BLLM=Qwen1.5-1.8B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 66.1 | |
| InternVL2-2BLLM=InternLM2-1.8B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 94.1 | |
| LLaVA-1.5LLM=Vicuna-7B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 66.8 | |
| LLaVA-FastVBackbone=Dynamic-LLaVA-7B, Token reduction=25%2026.04 | — | — | — | — | — | — | — | — | 68.7 | |
| LLaVA-HiREDBackbone=Dynamic-LLaVA-7B, Token reduction=25%2026.04 | — | — | — | — | — | — | — | — | 66.4 | |
| LLaVA-MiniLLM=Vicuna-7B, Retain Ratio (%)=∼ 10%2026.05 | — | — | — | — | — | — | — | — | 69.7 | |
| LLaVA-MODLLM=Qwen1.5-1.8B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 68 | |
| LLaVA-MODLLM=Qwen2-1.5B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 69.2 | |
| LLaVA-PruMerge+Backbone=Dynamic-LLaVA-7B, Token reduction=25%2026.04 | — | — | — | — | — | — | — | — | 68.3 | |
| Mini-Gemini-2BLLM=Gemma-2B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 63.1 | |
| MiniCPM-V2LLM=MiniCPM-2.4B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 80.4 | |
| MoE-LLaVA-2BLLM=Qwen-1.5-1.8B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 63.1 | |
| PruMergeLLM=Vicuna-7B, Retain Ratio (%)=∼ 5%2026.05 | — | — | — | — | — | — | — | — | 68.5 | |
| PruMerge+LLM=Vicuna-7B, Retain Ratio (%)=∼ 25%2026.05 | — | — | — | — | — | — | — | — | 68.3 | |
| PyramidDropLLM=Vicuna-7B, Retain Ratio (%)=∼ 10%2026.05 | — | — | — | — | — | — | — | — | 69 | |
| Qwen-VL-ChatLLM=Qwen-7B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 68.2 | |
| Qwen2-VL-2BLLM=Qwen2-1.5B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 78.1 | |
| SparseVLMLLM=Vicuna-7B, Retain Ratio (%)=∼ 10%2026.05 | — | — | — | — | — | — | — | — | 62.2 | |
| TinyLLaVA†LLM=Qwen1.5-0.5B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 60.9 | |
| TinyLLaVA†LLM=Qwen2.5-0.5B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 59.1 | |
| TinyLLaVA†LLM=Qwen1.5-1.8B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 65.3 | |
| TinyLLaVA†LLM=Qwen2.5-1.5B, Retain Ratio (%)=100%2026.05 | — | — | — | — | — | — | — | — | 72 | |
| ToMeBackbone=Dynamic-LLaVA-7B, Token reduction=25%2026.04 | — | — | — | — | — | — | — | — | 62.7 | |
| VanillaBackbone=Dynamic-LLaVA-7B, Token reduction=0%2026.04 | — | — | — | — | — | — | — | — | 66.8 | |
| VEN-VLLLM=Qwen3-0.6B, Retain Ratio (%)=∼ 7.5%2026.05 | — | — | — | — | — | — | — | — | 84.2 | |
| VEN-VLLLM=Qwen2.5-1.5B, Retain Ratio (%)=∼ 10%2026.05 | — | — | — | — | — | — | — | — | 90.5 | |
| VEN-VLLLM=Qwen3-1.7B, Retain Ratio (%)=∼ 7.5%2026.05 | — | — | — | — | — | — | — | — | 94.5 | |
| VisionZipLLM=Vicuna-7B, Retain Ratio (%)=∼ 10%2026.05 | — | — | — | — | — | — | — | — | 69 |