List your company
ResearchFormerly WizWand

Commonsense Question Answering on CSQA

82.72AccuracySAC Single-task
23.606438.953254.369.6468Nov 4, 2024Feb 5, 2025May 9, 2025Aug 11, 2025Nov 12, 2025Feb 13, 2026May 18, 2026
Updated 5mo ago

Evaluation results

MethodLinks
SAC Single-taskBackbone=Qwen2-72B-Instruct2024.1182.72—
SAC Multi-taskBackbone=Qwen2-72B-Instruct2024.1182.56—
No ControlBackbone=Qwen2-72B-Instruct2024.1182.39—
InfLLMv2Model Size=8B, Context=short-context2026.0580.6—
NSAModel Size=8B, Context=short-context2026.0580.4—
DashAttentionModel Size=8B, Context=short-context2026.0580.3—
FullAttnModel Size=8B, Context=short-context2026.0580—
No ControlBackbone=Qwen2-7B-Instruct2024.1178.62—
SAC Single-taskBackbone=Qwen2-7B-Instruct2024.1177.07—
SAC Multi-taskBackbone=Qwen2-7B-Instruct2024.1176.9—
Top-K ReducedSparsity Level=Mid Sparsity, K=4, Avg. K=4.002026.0572.48—
MoE-DynamicSparsity Level=High Sparsity, phi=0.1, Avg. K=3.902026.0571.09—
BEAMSparsity Level=High Sparsity, beta=0.1, Avg. K=1.082026.0570.93—
MoE-DynamicSparsity Level=Mid Sparsity, phi=0.3, Avg. K=4.312026.0570.6—
AdaMoESparsity Level=High Sparsity, Null=128, Avg. K=2.112026.0570.27—
AdaMoESparsity Level=Mid Sparsity, Null=64, Avg. K=3.252026.0569.86—
16-bit BaselineModel Variant=LLaMA-3-8B2024.1169.2—
DeepSeekV2-LiteK=6, Avg. K=6.002026.0568.14—
BEAMSparsity Level=Extreme Sparsity, beta=1.0, Avg. K=0.482026.0567.57—
Top-K PruningSparsity Level=Mid Sparsity, K=4, Avg. K=4.002026.0567.4—
16-bit BaselineModel Variant=LLaMA-2-13B2024.1167.3—
Top-K ReducedSparsity Level=High Sparsity, K=2, Avg. K=2.002026.0566.83—
ANVFP4Group Size (GS)=16, Model Variant=LLaMA-2-13B2024.1166.2—
NVFP4Group Size (GS)=16, Model Variant=LLaMA-2-13B2024.1165.3—
BEAMSparsity Level=Mid Sparsity, beta=0.01, Avg. K=2.612026.0565.11—
MXFP4Group Size (GS)=32, Model Variant=LLaMA-2-13B2024.1165.1—
NVFP4Group Size (GS)=32, Model Variant=LLaMA-2-13B2024.1165—
16-bit BaselineModel Variant=LLaMA-2-7B2024.1164.9—
AMXFP4Group Size (GS)=32, Model Variant=LLaMA-2-13B2024.1164.9—
ANVFP4Group Size (GS)=16, Model Variant=LLaMA-3-8B2024.1164.9—
ANVFP4Group Size (GS)=32, Model Variant=LLaMA-2-13B2024.1164.7—
NVFP4Group Size (GS)=16, Model Variant=LLaMA-3-8B2024.1163.4—
ANVFP4Group Size (GS)=32, Model Variant=LLaMA-3-8B2024.1162.9—
NVFP4Group Size (GS)=16, Model Variant=LLaMA-2-7B2024.1162.6—
MXFP4-POTGroup Size (GS)=32, Model Variant=LLaMA-2-13B2024.1162.2—
AMXFP4Group Size (GS)=32, Model Variant=LLaMA-3-8B2024.1162.2—
ANVFP4Group Size (GS)=32, Model Variant=LLaMA-2-7B2024.1162.2—
ANVFP4Group Size (GS)=16, Model Variant=LLaMA-2-7B2024.1162.2—
MXFP4Group Size (GS)=32, Model Variant=LLaMA-3-8B2024.1162—
AMXFP4Group Size (GS)=32, Model Variant=LLaMA-2-7B2024.1162—
NVFP4Group Size (GS)=32, Model Variant=LLaMA-3-8B2024.1161.9—
MXFP4Group Size (GS)=32, Model Variant=LLaMA-2-7B2024.1161.6—
NVFP4Group Size (GS)=32, Model Variant=LLaMA-2-7B2024.1161.4—
MXFP4-POTGroup Size (GS)=32, Model Variant=LLaMA-2-7B2024.1159.4—
Top-K PruningSparsity Level=High Sparsity, K=2, Avg. K=2.002026.0558.64—
MXFP4-POTGroup Size (GS)=32, Model Variant=LLaMA-3-8B2024.1158.6—
Top-K ReducedSparsity Level=Extreme Sparsity, K=1, Avg. K=1.002026.0552.91—
MIDUS-HMLTraining Stage=SFT-1B2025.1250.04—
OpT-DeUSTraining Stage=SFT-1B2025.1249.55—
LESATraining Stage=SFT-1B2025.1247.91—
Avg-DeUSTraining Stage=SFT-1B2025.1247.01—
OpT-DeUSTraining Stage=CPT-1B2025.1246.44—
MIDUS-HMLTraining Stage=CPT-1B2025.1246.27—
LESATraining Stage=CPT-1B2025.1245.21—
Llama ProTraining Stage=SFT-1B2025.1242.59—
Avg-DeUSTraining Stage=CPT-1B2025.1241.93—
Llama ProTraining Stage=CPT-1B2025.1235.14—
SOLARTraining Stage=SFT-1B2025.1227.44—
BaseTraining Stage=SFT-1B2025.1226.37—
SOLARTraining Stage=CPT-1B2025.1226.29—
BaseTraining Stage=CPT-1B2025.1225.88—
Gemma 7BFramework=Adaptive HPF2024.06—5.5771
Llama-3 8BFramework=Adaptive HPF2024.06—5.9136
Mistral 7BFramework=Adaptive HPF2024.06—5.9073
Phi-3 3.8BFramework=Adaptive HPF2024.06—5.6793