List your company
ResearchFormerly WizWand

Science Question Answering on SQA (accuracy)

98Accuracy (SQA)Qwen3-VL-235B
54.00865.42976.8588.271Sep 16, 2025Oct 31, 2025Dec 16, 2025Jan 31, 2026Mar 17, 2026May 2, 2026Jun 17, 2026
Updated 4mo ago

Evaluation results

MethodLinks
Qwen3-VL-235BToken Retention Ratio=100%2026.0598
F3AToken Retention Ratio=60%2026.0594.66
VisionZipToken Retention Ratio=60%2026.0594.1
DivPruneToken Retention Ratio=60%2026.0594.02
CDPrunerToken Retention Ratio=60%2026.0593.74
FastVToken Retention Ratio=60%2026.0593.21
F3AToken Retention Ratio=40%2026.0592.71
DivPruneToken Retention Ratio=40%2026.0591.73
CDPrunerToken Retention Ratio=40%2026.0591.52
FastVToken Retention Ratio=40%2026.0590.96
VisionZipToken Retention Ratio=40%2026.0590.88
F3AToken Retention Ratio=20%2026.0589.53
Qwen2.5-VL-7BToken Retention=100%2026.0688.4
CDPrunerToken Retention Ratio=20%2026.0587.62
DivPruneToken Retention Ratio=20%2026.0587.11
VisionZipToken Retention Ratio=20%2026.0586.88
StepPruneBackbone=Qwen2.5-VL-7B, Token Pruning Rate=66.7%2026.0686.7
FastVToken Retention Ratio=20%2026.0585.9
StepPruneBackbone=Qwen2.5-VL-7B, Token Pruning Rate=77.8%2026.0685.4
SPAREToken Retention=20%2026.0684.9
VanillaBackbone=Qwen2.5-VL-7B, Token Pruning Rate=0%2026.0684.7
VanillaBackbone=Qwen2-VL-7B, Token Retention Ratio=100.0%2026.0684.6
CDPrunerToken Retention=20%2026.0684.3
StepPruneBackbone=Qwen2.5-VL-7B, Token Pruning Rate=88.9%2026.0683.2
SPAREToken Retention=10%2026.0682.5
CDPrunerToken Retention=10%2026.0682.1
HoloVBackbone=Qwen2.5-VL-7B, Token Pruning Rate=66.7%2026.0679.8
HoloVBackbone=Qwen2.5-VL-7B, Token Pruning Rate=77.8%2026.0679.8
HoloVBackbone=Qwen2.5-VL-7B, Token Pruning Rate=88.9%2026.0679.5
DCP-PruneBackbone=Qwen2-VL-7B, Token Retention Ratio=5%2026.0679.1
FastVBackbone=Qwen2.5-VL-7B, Token Pruning Rate=66.7%2026.0678.5
SCOPEBackbone=Qwen2-VL-7B, Token Retention Ratio=5%2026.0678.3
FlowCutBackbone=Qwen2-VL-7B, Token Retention Ratio=5%2026.0678.2
FastVBackbone=Qwen2.5-VL-7B, Token Pruning Rate=77.8%2026.0678
FastVBackbone=Qwen2.5-VL-7B, Token Pruning Rate=88.9%2026.0677.4
LLaVA Vicuna STSMLLM=LLaVA, LLM=Vicuna, Method=STS, Activated Params=0.27B, FLOPs=82.2%2025.0970.2
Upper BoundBase Model=LLaVA-NeXT-7B, Token Retention Budget=unpruned2026.0470.1
Mipha Phi-1.5 BaselineMLLM=Mipha, LLM=Phi-1.5, Method=Baseline, Activated Params=1.30B2025.0969.9
CLASPBase Model=LLaVA-NeXT-7B, Token Retention Budget=640 Tokens2026.0469.8
Mipha Phi-1.5 STSMLLM=Mipha, LLM=Phi-1.5, Method=STS, Activated Params=1.08B, FLOPs=83.5%2025.0968.9
LLaVA Vicuna BaselineMLLM=LLaVA, LLM=Vicuna, Method=Baseline, Activated Params=0.32B2025.0968.4
SparseVLMBase Model=LLaVA-NeXT-7B, Token Retention Budget=640 Tokens2026.0467.6
CLASPBase Model=LLaVA-NeXT-7B, Token Retention Budget=320 Tokens2026.0467
SparseVLMBase Model=LLaVA-NeXT-7B, Token Retention Budget=320 Tokens2026.0466.1
MoE-LLaVA StableLM STSMLLM=MoE-LLaVA, LLM=StableLM, Method=STS, Activated Params=1.33B, FLOPs=83.9%2025.0963.4
MoE-LLaVA Qwen BaselineMLLM=MoE-LLaVA, LLM=Qwen, Method=Baseline, Activated Params=1.80B2025.0963.1
MoE-LLaVA StableLM BaselineMLLM=MoE-LLaVA, LLM=StableLM, Method=Baseline, Activated Params=1.60B2025.0962.6
MoE-LLaVA Qwen STSMLLM=MoE-LLaVA, LLM=Qwen, Method=STS, Activated Params=1.49B, FLOPs=83.9%2025.0962.1
CLASPBase Model=LLaVA-NeXT-7B, Token Retention Budget=160 Tokens2026.0460.9
SparseVLMBase Model=LLaVA-NeXT-7B, Token Retention Budget=160 Tokens2026.0457.8
Mipha Phi-2 BaselineMLLM=Mipha, LLM=Phi-2, Method=Baseline, Activated Params=2.70B2025.0957.5
Mipha Phi-2 STSMLLM=Mipha, LLM=Phi-2, Method=STS, Activated Params=2.24B, FLOPs=83.8%2025.0955.7