Long-context reasoning on RULER 4K context length
99.25NIAHLLaDA1.5
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| LLaDA1.5Base Model=LLaDA1.5, Attention Method=Full Attention2026.05 | 99.25 | 47.78 | 88.5 | 100 | 89.74 | |
| LLaDA1.5+BA-AttBase Model=LLaDA1.5, Attention Method=BA-Att2026.05 | 98.16 | 41.2 | 85.5 | 99.4 | 87.54 | |
| UltraLLaDA+BA-AttBase Model=UltraLLaDA, Attention Method=BA-Att2026.05 | 98.16 | 67.78 | 75 | 100 | 90.06 | |
| UltraLLaDABase Model=UltraLLaDA, Attention Method=Full Attention2026.05 | 97.31 | 66.69 | 68.5 | 100 | 88.37 | |
| UltraLLaDA+XAttBase Model=UltraLLaDA, Attention Method=XAtt2026.05 | 95.12 | 49.61 | 70.5 | 57.4 | 81.43 | |
| LLaDA1.5+XAttBase Model=LLaDA1.5, Attention Method=XAtt2026.05 | 74.66 | 26.55 | 61 | 3 | 59.64 |