Multimodal Benchmarking on MMBench (Accuracy, Avg)
67.7AccuracyLLaVA-1.5
Evaluation results
| Method | Links | ||
|---|---|---|---|
| LLaVA-1.5LLM=LLaMA-13B [59], Model Type=Dense Model2025.07 | 67.7 | 64.2 | |
| Our MethodLLM=Phi2-2.7B [62], Model Type=Sparse Model, Expert Selection (Top-k)=Top42025.07 | 66.8 | 62.3 | |
| MoE-LLaVA-4Top2LLM=Phi2-2.7B [62], Model Type=Sparse Model, Expert Selection (Top-k)=4Top22025.07 | 65.2 | 61.1 | |
| LLaVA-1.5LLM=Vicuna-7B [60], Model Type=Dense Model2025.07 | 64.3 | 61.2 | |
| Our MethodLLM=StableLM-1.6B [65], Model Type=Sparse Model, Expert Selection (Top-k)=Top42025.07 | 60.6 | 58.8 | |
| MoE-LLaVA-4Top2LLM=StableLM-1.6B [65], Model Type=Sparse Model, Expert Selection (Top-k)=4Top22025.07 | 60.2 | 57.6 | |
| LLaVA-PhiLLM=Phi2-2.7B [62], Model Type=Dense Model2025.07 | 59.8 | — | |
| MobileVLMLLM=MobileLLaMA-2.7B [63], Model Type=Dense Model2025.07 | 59.6 | — | |
| IDEFICSLLM=IDEFICS-65B [58], Model Type=Dense Model2025.07 | 54.5 | — | |
| Qwen-VLLLM=Qwen-7B [4], Model Type=Dense Model2025.07 | 38.2 | — |