Log-density Estimation on 2-component diagonal-covariance GMMs d=100, n=2048 context (test)
20.8Log-density MSEDiScoFormer
Evaluation results
| Method | Links | |
|---|---|---|
| DiScoFormerd_model=256, heads=8, layers=6, training steps=150k2025.11 | 20.8 | |
| Oracle h KDEbandwidth strategy=best fixed bandwidth found by grid search2025.11 | 781 | |
| Scott KDEbandwidth strategy=Scott's rule2025.11 | 967 |