List your company

Language Modeling on OpenWebText (OWT) (Gen. PPL, Ent.)

15.06Gen. PPLReal data
-36.2896310.3202656.931,003.5398Feb 13, 2026Mar 2, 2026Mar 20, 2026Apr 6, 2026Apr 24, 2026May 11, 2026May 29, 2026

Evaluation results

MethodLinks
Real data2026.0515.065.44
Bitstream DiffusionNFE=256, Sampling Strategy=Stochastic2026.0527.065.26
AR Transformer2026.0535.95.58
LangFlowNFE=10242026.0536.535.25
CoFReBudget=768, Train=≈123h + 30k, VRAM=93.44 GiB/GPU2026.0537.8465.142
Bitstream DiffusionNFE=256, Sampling Strategy=Deterministic2026.0546.325.13
MDLM + SDTTBudget=768, Train=≈139h + SDTT, VRAM=112.4 GiB/GPU2026.0547.045.45
Duo + DCDSteps=12026.0247.132.8
CoFReBudget=384, Train=≈123h + 30k, VRAM=93.44 GiB/GPU2026.0548.7555.283
SCDDSampling Steps=1024, pu=0.2, Context Length=5122026.0255.7—
SCDDSampling Steps=512, pu=0.2, Context Length=5122026.0258.2—
SCDDSampling Steps=256, pu=0.2, Context Length=5122026.0259.6—
SCDDSampling Steps=128, pu=0.2, Context Length=5122026.0260.7—
SCDDSampling Steps=1024, pu=0.1, Context Length=5122026.0261.3—
FLMSampling steps=10242026.0262.235.33
FLM2026.0562.235.33
MDLM + SDTTBudget=384, Train=≈139h + SDTT, VRAM=112.4 GiB/GPU2026.0562.295.49
SCDDSampling Steps=512, pu=0.1, Context Length=5122026.0263.6—
GIDD+Sampling Steps=1024, pu=0.1, Context Length=5122026.0263.8—
GIDD+Sampling Steps=512, pu=0.1, Context Length=5122026.0264.8—
GIDD+Sampling Steps=256, pu=0.1, Context Length=5122026.0265—
CoFReBudget=192, Train=≈123h + 30k, VRAM=93.44 GiB/GPU2026.0565.1825.38
SCDDSampling Steps=256, pu=0.1, Context Length=5122026.0266—
GIDD+Sampling Steps=128, pu=0.1, Context Length=5122026.0266.7—
SCDDSampling Steps=64, pu=0.2, Context Length=5122026.0267.1—
SCDDSampling Steps=128, pu=0.1, Context Length=5122026.0267.6—
ReMDM-CapSampling Steps=1024, cap=0.01, Context Length=5122026.0268.3—
GIDD+Sampling Steps=1024, pu=0.2, Context Length=5122026.0271.2—
GIDD+Sampling Steps=64, pu=0.1, Context Length=5122026.0271.4—
SCDDSampling Steps=64, pu=0.1, Context Length=5122026.0271.8—
GIDD+Sampling Steps=512, pu=0.2, Context Length=5122026.0272—
GIDD+Sampling Steps=256, pu=0.2, Context Length=5122026.0273.2—
ReMDM-CapSampling Steps=512, cap=0.01, Context Length=5122026.0273.9—
SCDDSampling Steps=32, pu=0.2, Context Length=5122026.0274.5—
GIDD+Sampling Steps=128, pu=0.2, Context Length=5122026.0275.1—
FMLMSteps=42026.0276.375.05
DuoSampling steps=10242026.0277.695.55
Duo2026.0577.695.55
SCDDSampling Steps=32, pu=0.1, Context Length=5122026.0278.6—
GIDD+Sampling Steps=64, pu=0.2, Context Length=5122026.0279—
ReMDM-ConfidenceSampling Steps=1024, Context Length=5122026.0280.5—
ReMDM-CapSampling Steps=256, cap=0.01, Context Length=5122026.0281.7—
GIDD+Sampling Steps=32, pu=0.1, Context Length=5122026.0282.1—
ReMDM-ConfidenceSampling Steps=512, Context Length=5122026.0283.9—
ReMDM-ConfidenceSampling Steps=256, Context Length=5122026.0287.9—
MDLMSampling Steps=1024, Context Length=5122026.0288.5—
MDLM + SDTTBudget=192, Train=≈139h + SDTT, VRAM=112.4 GiB/GPU2026.0589.175.53
GIDD+Sampling Steps=32, pu=0.2, Context Length=5122026.0290.5—
MDLMSampling Steps=512, Context Length=5122026.0291.9—
MDLMSampling Steps=256, Context Length=5122026.0294.8—
ReMDM-CapSampling Steps=128, cap=0.01, Context Length=5122026.0295.9—
Duo + DCDSteps=22026.0296.593.77
Duo + Di4CSteps=12026.0297.773.36
ReMDM-ConfidenceSampling Steps=128, Context Length=5122026.0298.1—
CoFReBudget=96, Train=≈123h + 30k, VRAM=93.44 GiB/GPU2026.05101.7915.434
MDLMSampling Steps=128, Context Length=5122026.02104.7—
MDLM2026.05105.155.63
Duo + DCDSteps=42026.02108.214.82
ReMDM-ConfidenceSampling Steps=64, Context Length=5122026.02118.3—
ReMDM-CapSampling Steps=64, cap=0.01, Context Length=5122026.02120.9—
MDLMSampling steps=10242026.02121.095.65
MDLMSampling Steps=64, Context Length=5122026.02123.6—
FMLMSteps=12026.02129.324.53
FMLMSteps=22026.02134.265.07
CANDISampling steps=10242026.02143.135.71
CANDI2026.05143.135.71
Duo + Di4CSteps=42026.02150.674.81
Duo + Di4CSteps=22026.02165.814.65
ReMDM-CapSampling Steps=32, cap=0.01, Context Length=5122026.02166.3—
ReMDM-ConfidenceSampling Steps=32, Context Length=5122026.02167.6—
MDLMSampling Steps=32, Context Length=5122026.02169.9—
MDLM + SDTTBudget=96, Train=≈139h + SDTT, VRAM=112.4 GiB/GPU2026.05193.055.58
MDLM + Di4CSteps=42026.02239.275.4
MDLM + SDTTSteps=42026.02339.735.38
MDLM + Di4CSteps=22026.02758.235.35
MDLM + SDTTSteps=22026.02877.225.34
MDLM + SDTTSteps=12026.021,260.865.26
MDLM + Di4CSteps=12026.021,298.85.29