Long-context reasoning on RULER 8K context length
98.5NIAH ScoreUltraLLaDA+BA-Att
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| UltraLLaDA+BA-AttBase Model=UltraLLaDA, Attention Method=BA-Att2026.05 | 98.5 | 57.65 | 68.5 | 100 | 87.71 | |
| UltraLLaDABase Model=UltraLLaDA, Attention Method=Full Attention2026.05 | 98.28 | 55.29 | 62 | 100 | 86.22 | |
| UltraLLaDA+XAttBase Model=UltraLLaDA, Attention Method=XAtt2026.05 | 86.97 | 36.3 | 62 | 24.2 | 70.5 | |
| LLaDA1.5Base Model=LLaDA1.5, Attention Method=Full Attention2026.05 | 51.56 | 41.6 | 48.02 | 61.6 | 50.25 | |
| LLaDA1.5+BA-AttBase Model=LLaDA1.5, Attention Method=BA-Att2026.05 | 50.28 | 58.23 | 48.1 | 59.4 | 51.85 | |
| LLaDA1.5+XAttBase Model=LLaDA1.5, Attention Method=XAtt2026.05 | 39.53 | 33.21 | 35.5 | 0.4 | 34.93 |