Reverse on Reverse OOD lengths: 1.5x, 2x, 4x, 8x
100Exact Match AccuracyTop-K (K=32)
Evaluation Results
| Method | Links | |
|---|---|---|
| Top-K (K=32)Layers (L)=6, OOD Sequence Length Scale=1.5x2025.06 | 100 | |
| ASEntmaxLayers (L)=6, OOD Sequence Length Scale=1.5x2025.06 | 100 | |
| ASEntmaxLayers (L)=6, OOD Sequence Length Scale=2x2025.06 | 99.8 | |
| EntmaxLayers (L)=6, OOD Sequence Length Scale=1.5x2025.06 | 99 | |
| Top-K (K=32)Layers (L)=6, OOD Sequence Length Scale=2x2025.06 | 98.7 | |
| ASEntmaxLayers (L)=6, OOD Sequence Length Scale=4x2025.06 | 96.4 | |
| EntmaxLayers (L)=6, OOD Sequence Length Scale=2x2025.06 | 86 | |
| Top-K (K=32)Layers (L)=6, OOD Sequence Length Scale=4x2025.06 | 57 | |
| ASEntmaxLayers (L)=6, OOD Sequence Length Scale=8x2025.06 | 56.7 | |
| SSMaxLayers (L)=6, OOD Sequence Length Scale=1.5x2025.06 | 54.6 | |
| SoftmaxLayers (L)=6, OOD Sequence Length Scale=1.5x2025.06 | 36 | |
| EntmaxLayers (L)=6, OOD Sequence Length Scale=4x2025.06 | 28.5 | |
| EntmaxLayers (L)=6, OOD Sequence Length Scale=8x2025.06 | 0.2 | |
| SoftmaxLayers (L)=6, OOD Sequence Length Scale=2x2025.06 | 0 | |
| SoftmaxLayers (L)=6, OOD Sequence Length Scale=4x2025.06 | 0 | |
| SoftmaxLayers (L)=6, OOD Sequence Length Scale=8x2025.06 | 0 | |
| Top-K (K=32)Layers (L)=6, OOD Sequence Length Scale=8x2025.06 | 0 | |
| SSMaxLayers (L)=6, OOD Sequence Length Scale=2x2025.06 | 0 | |
| SSMaxLayers (L)=6, OOD Sequence Length Scale=4x2025.06 | 0 | |
| SSMaxLayers (L)=6, OOD Sequence Length Scale=8x2025.06 | 0 |