List your company

Language Modeling on Penn Treebank (PTB) (test)

14.72Perplexityfull-precision
12.164429.414746.66563.9153Sep 18, 2018Dec 24, 2019Mar 30, 2021Jul 5, 2022Oct 10, 2023Jan 14, 2025Apr 21, 2026

Evaluation results

MethodLinks
full-precisionBits (W-E-A)=full-prec., Size (MB)=474.92022.0314.72——————
QuantGPTBits (W-E-A)=8-8-8, Size (MB)=121.42022.0314.9——————
QuantGPTBits (W-E-A)=4-4-8, Size (MB)=62.42022.0314.95——————
LSQBits (W-E-A)=8-8-8, Size (MB)=121.42022.0315.43——————
LAQBits (W-E-A)=8-8-8, Size (MB)=121.42022.0315.87——————
PACTBits (W-E-A)=8-8-8, Size (MB)=121.42022.0316.11——————
QuantGPTBits (W-E-A)=2-2-8, Size (MB)=33.02022.0316.12——————
LAQBits (W-E-A)=4-4-8, Size (MB)=62.42022.0316.55——————
LAQBits (W-E-A)=2-2-8, Size (MB)=33.02022.0318.25——————
GPT2 + GFSA#Params=117M, Fine-tuning=true2023.1219.45——————
GPT2#Params=117M, Fine-tuning=true2023.1219.513——————
PACTBits (W-E-A)=4-4-8, Size (MB)=62.42022.0320.17——————
HSNK (Symbols)=100, L (Walk steps)=52022.0720.25—————9.3
HSNK (Symbols)=50, L (Walk steps)=52022.0722.47—————9.5
FP16Bits=-, Backbone=LLAMA-2-7B2026.0422.51——————
FP16Bits=–, Backbone=LLAMA-2-7B2026.0422.51——————
Optimus Alambda=0.05, Optimized for=PPL2022.0723.58—————3.78
GPT2Fine-tuned=single epoch2022.0724.23——————
HSNK (Symbols)=100, L (Walk steps)=202022.0725.48—————23.77
FBI-LLMBits=W1A16, Backbone=LLAMA-2-7B2026.0429.6——————
HSNK (Symbols)=50, L (Walk steps)=202022.0730.38—————26.06
Optimus Blambda=1, Optimized for=MI2022.0735.53—————8.18
LBLLMBits=W(1+1)A4, Backbone=LLAMA-2-7B2026.0436.79——————
LBLLMBits=W(1+1)A4, Backbone=LLAMA-2-7B2026.0436.79——————
AWD-LSTM + MoS + Partial Shuffled + Adversarial TrainingParams=22M, dynamic evaluation=true2019.0646.01——————
AWD-LSTM + MoS + Adversarial TrainingParams=22M, dynamic evaluation=true2019.0646.52——————
AWD-LSTM-MoS + dynamic evaluation with FRAGEParas=24M2018.0946.54——————
AWD-LSTM + MoS + Partial ShuffledParams=22M, dynamic evaluation=true2019.0647.49——————
AWD-LSTM + MoSParams=22M, dynamic evaluation=true2019.0647.69——————
AWD-LSTM-MoS + dynamic evaluationParas=24M2018.0947.69——————
AWD-LSTM + Adversarial TrainingParams=24M, dynamic evaluation=true2019.0648.72——————
ADV-AWD-LSTM + MaxUpParams=24M, Dynamic evaluation=true2020.0250.29——————
AWD-LSTMParams=24M, dynamic evaluation=true2019.0651.1——————
ADV-AWD-LSTMParams=24M, Dynamic evaluation=true2020.0251.1——————
AWD-LSTM + continuous cache pointer with FRAGEParas=24M2018.0951.8——————
AWD-LSTM + MoS + Partial Shuffled + Adversarial TrainingParams=22M, dynamic evaluation=false2019.0652.2——————
AWD-LSTM-DOC#Params=23M2022.0352.4——————
AWD-LSTM + continuous cache pointerParas=24M2018.0952.8——————
AWD-LSTM + MoS + Adversarial TrainingParams=22M, dynamic evaluation=false2019.0652.87——————
AWD-LSTM-MoS with FRAGEParas=24M2018.0953.31——————
iVAEMIEncoder/Decoder=one-layer LSTMs2022.0753.44—————12.5
AWD-LSTM + MoS + Partial ShuffledParams=22M, dynamic evaluation=false2019.0653.92——————
ADV-AWD-LSTM + MaxUpParams=24M, Dynamic evaluation=false2020.0254.27——————
LSTM + MoSSource=[49], Search Method=manual2019.0254.422—————
AWD-LSTM-MoS#Params=22M2022.0354.4——————
AWD-LSTM + MoSParams=22M, dynamic evaluation=false2019.0654.44——————
AWD-LSTM + MoSParams=22M, Dynamic evaluation=false2020.0254.44——————
AWD-LSTM-MoSParas=24M2018.0954.44——————
AWD-FWM (ours)Number of parameters=24M, Seeds=32020.1154.48——————
AWD-TXLNumber of parameters=24M, Seeds=3, AWD-style regularisation=true, Model averaging=true, Softmax temperature tuning=true2020.1154.52——————
AWD-LSTM + Adversarial TrainingParams=24M, dynamic evaluation=false2019.0655.01——————
ADV-AWD-LSTMParams=24M, Dynamic evaluation=false2020.0255.01——————
AWD-LSTM-MoS w/o finetune with FRAGEParas=24M2018.0955.23——————
Random search WSSource=Ours, Comparable Search Space=Y, Search Method=random2019.0255.5230.2511.25——
DARTS (second order)Source=[34], Comparable Search Space=Y, Search Method=gradient-based2019.0255.723112——
DARTS (2nd)Params (M)=23, Search Cost (GPU days)=0.252019.1055.7——————
DARTS (second order)Source=Ours (reproduced), Comparable Search Space=Y, Search Method=gradient-based2019.0255.923112——
AWD-LSTM-MoS w/o finetuneParas=24M2018.0955.97——————
LSTM + SEParams (M)=222019.1056——————
DARTSParams=23M2019.0656.1——————
AWD-LSTM + FRAGEParams=24M, Dynamic evaluation=false2020.0256.1——————
AWD-LSTM with FRAGEParas=24M2018.0956.1——————
ON-LSTM#Params=25M2022.0356.2——————
DM-LSTM#Params=24M2022.0356.2——————
ASHA + LSTM + DropConnectSource=[30], Comparable Search Space=N, Search Method=HP-tuned, Search Cost Unit=CPU-days2019.0256.324——13——
ENAS*Source=[41], Comparable Search Space=Y, Search Method=RL2019.0256.3240.5————
ASHA baselineSource=Ours, Comparable Search Space=Y, Search Method=random2019.0256.423——2——
AWD-LSTMParams=24M, Dynamic evaluation=false2020.0256.5——————
AWD-TXL (ours)Number of parameters=24M, Seeds=3, AWD-style regularisation=true, Model averaging=true, Softmax temperature tuning=true2020.1156.5——————
Fraternal dropout#Params=24M2022.0356.8——————
LSTM + DropConnectSource=[38], Search Method=manual2019.0257.324—————
AWD-LSTMParams=24M, dynamic evaluation=false2019.0657.3——————
AWD-LSTMParas=24M2018.0957.3——————
AWD-LSTM#Params=24M2022.0357.3——————
AWD-LSTM2022.0557.3——————
AWD-LSTMNumber of parameters=24M, Seeds=32020.1157.3——————
GDASParams (M)=23, Search Cost (GPU days)=0.42019.1057.5——————
DARTS (first order)Source=[34], Comparable Search Space=Y, Search Method=gradient-based2019.0257.6230.511.5——
DARTS (1st)Params (M)=23, Search Cost (GPU days)=0.132019.1057.6——————
AWD-LSTM w/o finetune with FRAGEParas=24M2018.0958——————
LSTM + SCParams (M)=242019.1058.3——————
ENASSource=[34], Comparable Search Space=Y, Search Method=random2019.0258.6240.5————
ENASParams (M)=24, Search Cost (GPU days)=0.52019.1058.6——————
LSTMParams (M)=242019.1058.8——————
AWD-LSTM w/o finetuneParas=24M2018.0958.8——————
Random search baselineSource=[34], Comparable Search Space=Y, Search Method=random2019.0259.423——2——
PRPN2022.0362——————
NAS-RNNParams=54M2019.0662.4——————
NAS-RNNParams=54M, Dynamic evaluation=false2020.0262.4——————
NASSource=[51], Comparable Search Space=N, Search Method=RL2019.026425——10,000——
NASParams (M)=25, Search Cost (GPU days)=10^42019.1064——————
V-RHNParams (M)=232019.1065.4——————
Variational RHN#Params=23M2022.0365.4——————
Variational LSTM + weight tyingParams=51M2019.0668.5——————
BWABits=W(1+1)A4, Backbone=LLAMA-2-7B2026.0469.46——————
Pointer Sentinel-LSTM#Params=21M2022.0370.9——————
Variational LSTMParams=19M2019.0673.4——————
CBS-1-TModel Configuration=L22018.1277.39————65—
CBS-1Model Configuration=L22018.1277.56————69—
CBS-5Model Configuration=L22018.1278.61————73—