Text Generation on LM1B (test)
2.46EntropyUDLM
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| UDLMSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=42025.12 | 2.46 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=42025.12 | 2.48 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=82025.12 | 2.49 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=82025.12 | 2.49 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=162025.12 | 2.5 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=322025.12 | 2.5 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=642025.12 | 2.5 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=2562025.12 | 2.5 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=10242025.12 | 2.5 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=162025.12 | 2.51 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=1282025.12 | 2.51 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=5122025.12 | 2.51 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=322025.12 | 2.52 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=642025.12 | 2.52 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=2562025.12 | 2.52 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=10242025.12 | 2.52 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=1282025.12 | 2.53 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=16, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=5122025.12 | 2.53 | — | — | — | — | — | — | — | — | — | |
| ILMEvaluator Model=Llama 3.2 3B2026.06 | 2.8 | — | — | — | — | 4.67 | 21 | — | — | — | |
| DILM-SEvaluator Model=Llama 3.2 3B2026.06 | 2.87 | — | — | — | — | 4.65 | 30 | — | — | — | |
| DILM-MEvaluator Model=Llama 3.2 3B2026.06 | 2.99 | — | — | — | — | 4.76 | 48 | — | — | — | |
| UDLMSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=42025.12 | 3.05 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=42025.12 | 3.06 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=82025.12 | 3.09 | — | — | — | — | — | — | — | — | — | |
| CFMCategory=Baseline, one-step=true2026.05 | 3.1 | 269.72 | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=82025.12 | 3.12 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=162025.12 | 3.12 | — | — | — | — | — | — | — | — | — | |
| ARMEvaluator Model=Llama 3.2 3B2026.06 | 3.12 | — | — | — | — | 3.94 | 30 | — | — | — | |
| PAIRFLOWSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=162025.12 | 3.13 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=322025.12 | 3.13 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=642025.12 | 3.13 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=1282025.12 | 3.13 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=2562025.12 | 3.13 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=5122025.12 | 3.13 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=10242025.12 | 3.13 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=322025.12 | 3.14 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=642025.12 | 3.15 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=1282025.12 | 3.15 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=2562025.12 | 3.15 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=5122025.12 | 3.15 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=32, Training Iterations (Iter.)=200k, Number of Function Evaluations (NFE)=10242025.12 | 3.16 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=42025.12 | 3.57 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=42025.12 | 3.57 | — | — | — | — | — | — | — | — | — | |
| Coupling ModelsCategory=Ours, Temperature=0.50, top-p=0.95, latent scale zscale=0.8, one-step=true2026.05 | 3.6088 | 45.82 | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=82025.12 | 3.63 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=82025.12 | 3.65 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=162025.12 | 3.67 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=322025.12 | 3.68 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=162025.12 | 3.69 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=642025.12 | 3.69 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=5122025.12 | 3.69 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=322025.12 | 3.7 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=1282025.12 | 3.7 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=2562025.12 | 3.7 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=10242025.12 | 3.7 | — | — | — | — | — | — | — | — | — | |
| MDMEvaluator Model=Llama 3.2 3B2026.06 | 3.7 | — | — | — | — | 4.81 | 85 | — | — | — | |
| PAIRFLOWSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=642025.12 | 3.71 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=1282025.12 | 3.71 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=2562025.12 | 3.71 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=10242025.12 | 3.71 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=64, Training Iterations (Iter.)=400k, Number of Function Evaluations (NFE)=5122025.12 | 3.72 | — | — | — | — | — | — | — | — | — | |
| Duo + Di4CCategory=Baseline, one-step=true2026.05 | 3.79 | 292.94 | — | — | — | — | — | — | — | — | |
| DFM (ESD)Category=Baseline, one-step=true2026.05 | 3.79 | 68.11 | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=42025.12 | 3.98 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=42025.12 | 4 | — | — | — | — | — | — | — | — | — | |
| DFM (PSD)Category=Baseline, one-step=true2026.05 | 4.06 | 94.08 | — | — | — | — | — | — | — | — | |
| Coupling ModelsCategory=Ours, Temperature=0.60, top-p=0.95, latent scale zscale=0.8, one-step=true2026.05 | 4.0841 | 76.19 | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=82025.12 | 4.09 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=82025.12 | 4.11 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=162025.12 | 4.14 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=162025.12 | 4.16 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=322025.12 | 4.16 | — | — | — | — | — | — | — | — | — | |
| FMLMCategory=Baseline, one-step=true2026.05 | 4.16 | 119.34 | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=642025.12 | 4.17 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=1282025.12 | 4.17 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=2562025.12 | 4.17 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=322025.12 | 4.18 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=5122025.12 | 4.18 | — | — | — | — | — | — | — | — | — | |
| UDLMSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=10242025.12 | 4.18 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=642025.12 | 4.19 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=5122025.12 | 4.19 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=1282025.12 | 4.2 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=2562025.12 | 4.2 | — | — | — | — | — | — | — | — | — | |
| PAIRFLOWSequence Length (N)=128, Training Iterations (Iter.)=600k, Number of Function Evaluations (NFE)=10242025.12 | 4.2 | — | — | — | — | — | — | — | — | — | |
| MDLM + SDTTCategory=Baseline, one-step=true2026.05 | 4.31 | 1,429.48 | — | — | — | — | — | — | — | — | |
| LM1B test setCategory=Reference2026.05 | 4.31 | 47.47 | — | — | — | — | — | — | — | — | |
| Coupling ModelsCategory=Ours, Temperature=0.65, top-p=0.95, latent scale zscale=0.8, one-step=true2026.05 | 4.3192 | 98.96 | — | — | — | — | — | — | — | — | |
| Duo + DCDCategory=Baseline, one-step=true2026.05 | 4.33 | 1,224.52 | — | — | — | — | — | — | — | — | |
| MDLM + Di4CCategory=Baseline, one-step=true2026.05 | 4.38 | 1,217.1 | — | — | — | — | — | — | — | — | |
| Coupling ModelsCategory=Ours, Temperature=0.70, top-p=0.95, latent scale zscale=0.8, one-step=true2026.05 | 4.5716 | 134.06 | — | — | — | — | — | — | — | — | |
| ARBatch Size=42026.06 | — | 104.8 | — | — | — | — | — | — | 0.54 | 1 | |
| ARBatch Size=162026.06 | — | 104.8 | — | — | — | — | — | — | 2.03 | 1 | |
| ARBatch Size=1282026.06 | — | 104.8 | — | — | — | — | — | — | 15.4 | 1 | |
| BERT-MouthPretrained=true2022.11 | — | 142.89 | 0.2867 | 0.124 | — | — | — | — | — | — | |
| BERT-MouthSteps=64, Inference Time (secs)=2.182022.11 | — | 142.89 | — | — | — | — | — | — | — | — | |
| BERT-MouthSteps=512, Inference Time (secs)=14.392022.11 | — | 86.78 | — | — | — | — | — | — | — | — | |
| D3PMPretrained=false, Schedule=(T - t + 1)^-1, Time Step=LTE2022.11 | — | 82.34 | 0.3897 | 0.2347 | — | — | — | — | — | — | |
| D3PMPretrained=false, Schedule=(T - t + 1)^-1, Time Step=TAD2022.11 | — | 125.15 | 0.339 | 0.272 | — | — | — | — | — | — | |
| D3PMPretrained=false, Schedule=Spindle, Time Step=LTE2022.11 | — | 77.5 | 0.4241 | 0.2288 | — | — | — | — | — | — | |
| D3PM absorbinginference steps=1000, context window length=128, layers=12, vocabulary size=8192, number of seeds=22021.07 | — | 76.9 | — | — | 1.9 | — | — | — | — | — |