List your company
ResearchFormerly WizWand

Multi-modal Long-context Benchmarking on MileBench

57.23Task T ScoreFlashCache
37.646842.730947.81552.8991Nov 20, 2025
Updated 8mo ago

Evaluation results

MethodLinks
FlashCacheBackbone=Qwen2.5-VL-32B, KV Cache retention ratio (rho)=0.12025.1157.2359.0816.120.17
SnapKVBackbone=Qwen2.5-VL-32B, KV Cache retention ratio (rho)=0.12025.1157.0658.7116.0618.83
StreamingLLMBackbone=Qwen2.5-VL-32B, KV Cache retention ratio (rho)=0.12025.115756.152.520
H2OBackbone=Qwen2.5-VL-32B, KV Cache retention ratio (rho)=0.12025.1156.9458.332.521
Full CacheBackbone=Qwen2.5-VL-32B, KV Cache retention ratio (rho)=1.02025.1156.5165.2127.1918.17
FlashCacheBackbone=Qwen2.5-VL-32B, KV Cache retention ratio (rho)=0.052025.1156.2553.118.2926.33
SnapKVBackbone=Qwen2.5-VL-32B, KV Cache retention ratio (rho)=0.052025.1155.6949.076.2519.83
FlashCacheBackbone=Qwen2.5-VL-7B, KV Cache retention ratio (rho)=0.12025.1155.6560.9318.4514.5
Full CacheBackbone=Qwen2.5-VL-7B, KV Cache retention ratio (rho)=1.02025.1155.5969.1727.3514.17
SnapKVBackbone=Qwen2.5-VL-7B, KV Cache retention ratio (rho)=0.12025.1155.5861.37.1915.5
LOOK-MBackbone=Qwen2.5-VL-7B, KV Cache retention ratio (rho)=0.12025.1155.0858.545.7813.67
StreamingLLMBackbone=Qwen2.5-VL-7B, KV Cache retention ratio (rho)=0.12025.1154.7159.236.5615.83
H2OBackbone=Qwen2.5-VL-7B, KV Cache retention ratio (rho)=0.12025.1154.5858.656.2514.83
FlashCacheBackbone=Qwen2.5-VL-7B, KV Cache retention ratio (rho)=0.052025.1154.5556.1910.6314.83
MEDABackbone=Qwen2.5-VL-7B, KV Cache retention ratio (rho)=0.12025.1154.2457.055.4713.67
H2OBackbone=Qwen2.5-VL-32B, KV Cache retention ratio (rho)=0.052025.1152.3148.441.2523.67
StreamingLLMBackbone=Qwen2.5-VL-32B, KV Cache retention ratio (rho)=0.052025.1151.9442.341.2526.83
LOOK-MBackbone=Qwen2.5-VL-32B, KV Cache retention ratio (rho)=0.12025.1151.9158.992.8124.67
MEDABackbone=Qwen2.5-VL-32B, KV Cache retention ratio (rho)=0.052025.1151.8650.012.6526.33
MEDABackbone=Qwen2.5-VL-7B, KV Cache retention ratio (rho)=0.052025.1151.7452.955.1615.33
LOOK-MBackbone=Qwen2.5-VL-7B, KV Cache retention ratio (rho)=0.052025.1151.6953.675.3116.33
SnapKVBackbone=Qwen2.5-VL-7B, KV Cache retention ratio (rho)=0.052025.1151.5256.336.2515.33
MEDABackbone=Qwen2.5-VL-32B, KV Cache retention ratio (rho)=0.12025.1151.3148.452.9728.67
H2OBackbone=Qwen2.5-VL-7B, KV Cache retention ratio (rho)=0.052025.1150.4752.775.3111.33
LOOK-MBackbone=Qwen2.5-VL-32B, KV Cache retention ratio (rho)=0.052025.1149.9745.772.527.83
StreamingLLMBackbone=Qwen2.5-VL-7B, KV Cache retention ratio (rho)=0.052025.1149.4653.645.6315
LOOK-MBackbone=LLaVA-OneVision-1.5-8B, KV Cache retention ratio (rho)=0.052025.1148.0156.9912.511.5
FlashCacheBackbone=LLaVA-OneVision-1.5-8B, KV Cache retention ratio (rho)=0.12025.1147.0958.562012.83
H2OBackbone=LLaVA-OneVision-1.5-8B, KV Cache retention ratio (rho)=0.052025.1147.0156.9912.511.5
Full CacheBackbone=LLaVA-OneVision-1.5-8B, KV Cache retention ratio (rho)=1.02025.1146.9765.3327.0311.67
SnapKVBackbone=LLaVA-OneVision-1.5-8B, KV Cache retention ratio (rho)=0.12025.1146.9757.6210.4712.67
H2OBackbone=LLaVA-OneVision-1.5-8B, KV Cache retention ratio (rho)=0.12025.1146.8458.5115.9411.83
LOOK-MBackbone=LLaVA-OneVision-1.5-8B, KV Cache retention ratio (rho)=0.12025.1146.8458.5115.9411.83
StreamingLLMBackbone=LLaVA-OneVision-1.5-8B, KV Cache retention ratio (rho)=0.12025.1146.8158.46.7212.5
StreamingLLMBackbone=LLaVA-OneVision-1.5-8B, KV Cache retention ratio (rho)=0.052025.1146.5257.055.1612
SnapKVBackbone=LLaVA-OneVision-1.5-8B, KV Cache retention ratio (rho)=0.052025.1145.9757.827.9713
FlashCacheBackbone=LLaVA-OneVision-1.5-8B, KV Cache retention ratio (rho)=0.052025.1145.5151.1215.5613.33
MEDABackbone=LLaVA-OneVision-1.5-8B, KV Cache retention ratio (rho)=0.12025.1144.2352.7614.0612.17
MEDABackbone=LLaVA-OneVision-1.5-8B, KV Cache retention ratio (rho)=0.052025.1138.448.2510.3113