Long-context reasoning on RULER 16K context length
94.62NIAHUltraLLaDA+BA-Att
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| UltraLLaDA+BA-AttBase Model=UltraLLaDA, Attention Method=BA-Att2026.05 | 94.62 | 44.89 | 53 | 99.4 | 80.93 | |
| UltraLLaDABase Model=UltraLLaDA, Attention Method=Full Attention2026.05 | 93 | 43.12 | 39.5 | 98.4 | 77.51 | |
| UltraLLaDA+XAttBase Model=UltraLLaDA, Attention Method=XAtt2026.05 | 75.09 | 35.37 | 40 | 7.4 | 58.38 | |
| LLaDA1.5+BA-AttBase Model=LLaDA1.5, Attention Method=BA-Att2026.05 | 16.5 | 18.06 | 74 | 30 | 26.62 | |
| LLaDA1.5+XAttBase Model=LLaDA1.5, Attention Method=XAtt2026.05 | 15.16 | 15.89 | 37 | 0.6 | 17.51 | |
| LLaDA1.5Base Model=LLaDA1.5, Attention Method=Full Attention2026.05 | 8.56 | 11.19 | 46.5 | 20.6 | 15.73 |