Long-context reasoning on RULER 32K context length
92.78NIAHUltraLLaDA
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| UltraLLaDABase Model=UltraLLaDA, Attention Method=Full Attention2026.05 | 92.78 | 29.29 | 29 | 98.4 | 73.63 | |
| UltraLLaDA+BA-AttBase Model=UltraLLaDA, Attention Method=BA-Att2026.05 | 90.97 | 27.85 | 32 | 100 | 72.88 | |
| UltraLLaDA+XAttBase Model=UltraLLaDA, Attention Method=XAtt2026.05 | 74.94 | 30.19 | 15.5 | 1.8 | 53.28 |