List your company
ResearchFormerly WizWand

Commonsense Reasoning on OBQA

89.2AccuracyHydraLoRA
39.38452.31765.2578.183Oct 17, 2022May 21, 2023Dec 24, 2023Jul 27, 2024Mar 1, 2025Oct 3, 2025May 8, 2026
Updated 5mo ago

Evaluation results

MethodLinks
HydraLoRABackbone=Qwen2.5-3B2026.0289.2——
MoSLoRABackbone=Qwen2.5-3B2026.0289——
UQ4CTBackbone=Llama3.1-8B2024.1088.43.341—
MixLoRABackbone=Llama3.1-8B2024.1088.276.58—
LoRABackbone=Llama3.1-8B2024.1088.0017.3—
HypLoRABackbone=Qwen2.5-3B2026.0287.8——
BLoB(Mean)Backbone=Llama3.1-8B, Variant=Mean2024.1087.6016.83—
BLoB(N=10)Backbone=Llama3.1-8B, N=102024.1087.133.841—
MC DropBackbone=Llama3.1-8B2024.1087.1217.24—
LoRABackbone=Qwen2.5-3B2026.0287——
EnsembleBackbone=Llama3.1-8B2024.1086.478.63—
LABackbone=Llama3.1-8B2024.1086.00611.975—
BoHAModel=Llama-3.2-3B, #Params=48.63M2025.0985.73——
FFTModel=Llama-3.2-3B, #Params=3.21B2025.0985——
ABBAModel=Llama-3.2-3B, #Params=48.63M2025.0984.27——
GraLoRAModel=Llama-3.2-3B, #Params=48.63M2025.0983.73——
HiRAModel=Llama-3.2-3B, #Params=48.63M2025.0983.32——
BLoBBackbone=Llama2-7B, Method=LoRA, Steps=5,000, N=02024.0682.738.360.56
LAPBackbone=Llama2-7B, Method=LoRA, Steps=5,0002024.0682.128.70.52
BBBBackbone=Llama2-7B, Method=LoRA, Steps=5,0002024.0682.0611.380.66
BLoBBackbone=Llama2-7B, Method=LoRA, Steps=5,000, N=52024.0681.793.40.53
MCDBackbone=Llama2-7B, Method=LoRA, Steps=5,0002024.0681.7213.10.77
rsLoRAModel=Llama-3.2-3B, #Params=48.63M2025.0981.72——
MLEBackbone=Llama2-7B, Method=LoRA, Steps=5,0002024.0681.5212.550.73
BLoBBackbone=Llama2-7B, Method=LoRA, Steps=5,000, N=102024.0681.523.770.5
MAPBackbone=Llama2-7B, Method=LoRA, Steps=5,0002024.0681.3813.260.75
ENSBackbone=Llama2-7B, Method=LoRA, Steps=5,0002024.0681.3815.341.06
PiSSAModel=Llama-3.2-3B, #Params=48.63M2025.0980.8——
L-LoRA-SParameters=10M, Phase=MAP2026.0580.816.41.12
L-LoRA-XSRank=100, Phase=MAP2026.0580.815.61.12
L-LoRA-SParameters=10M, Phase=Bayesian2026.0580.66.50.6
MoREparams/task=4.5M, Backbone=Llama2-7B2026.0380.5——
L-LoRA-XSRank=100, Phase=Bayesian2026.0580.210.70.63
DoRAModel=Llama-3.2-3B, #Params=49.40M2025.0979.8——
LoRAModel=Llama-3.2-3B, #Params=48.63M2025.0979.6——
Laplace (LA)Phase=Bayesian2026.0578.96.40.65
L-LoRA-SParameters=4M, Phase=MAP2026.0578.816.51.09
L-LoRA-SParameters=4M, Phase=Bayesian2026.0578.810.10.77
Laplace (LA)Phase=MAP2026.0578.716.10.99
B-LoRA-XSRank=64, Phase=Bayesian2026.0578.610.70.74
EPTparams/task=3.3M, Backbone=Llama2-7B2026.0378.4——
B-LoRA-XSRank=64, Phase=MAP2026.0578.214.40.91
FullBackbone=Deepseek-R1-Distill-Qwen-7B2025.1278——
L-LoRA-XSRank=64, Phase=Bayesian2026.05789.20.66
L-LoRA-XSRank=64, Phase=MAP2026.0576.8181.16
B-LoRA-XSRank=32, Phase=MAP2026.0576.815.70.96
B-LoRA-XSRank=32, Phase=Bayesian2026.0576.410.20.71
SnapKVBackbone=Deepseek-R1-Distill-Qwen-7B, Budget=5122025.1276——
L-LoRA-XSRank=32, Phase=MAP2026.057618.51.05
L-LoRA-XSRank=32, Phase=Bayesian2026.0575.460.67
SnapKVBackbone=Deepseek-R1-Distill-Qwen-7B, Budget=2562025.1275——
LoRAparams/task=2.1M, Backbone=Llama2-7B2026.0374——
FFTModel=Llama-3.2-1B, #Params=1.24B2025.0974——
BoHAModel=Llama-3.2-1B, #Params=22.54M2025.0973.33——
ABBAModel=Llama-3.2-1B, #Params=22.54M2025.0972.93——
Base ModelBackbone=Llama3.1-8B2024.1072.811.39—
DoRAModel=Llama-3.2-1B, #Params=22.92M2025.0972.2——
PiSSAModel=Llama-3.2-1B, #Params=22.54M2025.0972.2——
DRAGON2022.1072——
LoRAModel=Llama-3.2-1B, #Params=22.54M2025.0971.4——
GraLoRAModel=Llama-3.2-1B, #Params=22.54M2025.0971.4——
rsLoRAModel=Llama-3.2-1B, #Params=22.54M2025.0971.11——
SnapKVBackbone=Deepseek-R1-Distill-Qwen-7B, Budget=1282025.1271——
HiRAModel=Llama-3.2-1B, #Params=22.54M2025.0970.2——
H2OBackbone=Deepseek-R1-Distill-Qwen-7B, Budget=3842025.1269——
MultiLoRAparams/task=10M, Backbone=Llama2-7B2026.0368.2——
SnapKVBackbone=Deepseek-R1-Distill-Qwen-7B, Budget=3842025.1268——
QAGNN2022.1067.8——
MOELoRAparams/task=4.5M, Backbone=Llama2-7B2026.0367.8——
H2OBackbone=Deepseek-R1-Distill-Qwen-7B, Budget=5122025.1267——
GreaseLM2022.1066.9——
RoBERTa2022.1064.9——
H2OBackbone=Deepseek-R1-Distill-Qwen-7B, Budget=2562025.1264——
CKT-baseBackbone=T5-base, Method=CKT2023.0661.58——
CALMBackbone=T5-base2023.0660.9——
LLaMAParameters=65B, Zero-shot=true, Likelihood Normalization=Answer-conditional2023.0260.2——
ERNIE-baseBackbone=ERNIE-base2023.0658.9——
LLaMAParameters=33B, Zero-shot=true, Likelihood Normalization=Answer-conditional2023.0258.6——
T5-base + SSMBackbone=T5-base, Augmentation=SSM2023.0658.6——
KnowBERTBackbone=KnowBERT2023.0658.5——
T5-base + TIBackbone=T5-base, Augmentation=TI2023.0658.43——
T5-baseBackbone=T5-base2023.0658.2——
T5-base + CSKG (Rule)Backbone=T5-base, Augmentation=CSKG (Rule)2023.0657.97——
GPT-3Parameters=175B, Zero-shot=true, Likelihood Normalization=Answer-conditional2023.0257.6——
BERT-baseBackbone=BERT-base2023.0657.6——
LLaMAParameters=7B, Zero-shot=true, Likelihood Normalization=Answer-conditional2023.0257.2——
CKT w/ GPT-2Backbone=GPT-2, Method=CKT2023.0656.95——
T5-base + KDBackbone=T5-base, Augmentation=KD2023.0656.54——
LLaMAParameters=13B, Zero-shot=true, Likelihood Normalization=Answer-conditional2023.0256.4——
T5-base + CSKG (TI)Backbone=T5-base, Augmentation=CSKG (TI)2023.0656.17——
PaLMParameters=540B, Zero-shot=true, Likelihood Normalization=Answer-conditional2023.0253.4——
COMETBackbone=BART2023.0651.2——
PaLMParameters=62B, Zero-shot=true, Likelihood Normalization=Answer-conditional2023.0250.4——
PASERQuantization=GPTQ 4 bits, Backbone=LLaMA2-13B2025.0244.2——
NuggetsQuantization=GPTQ 4 bits, Backbone=LLaMA2-13B2025.0242.7——
H2OBackbone=Deepseek-R1-Distill-Qwen-7B, Budget=1282025.1242——
w/o trainingQuantization=w/o Quant, Backbone=LLaMA2-13B2025.0242——
PASERQuantization=RTN 4 bits, Backbone=LLaMA2-13B2025.0241.7——
IFDQuantization=GPTQ 4 bits, Backbone=LLaMA2-13B2025.0241.4——
Instruction MiningQuantization=GPTQ 4 bits, Backbone=LLaMA2-13B2025.0241.3——