Long-context language understanding on LongBench (MFQA, Qasp, HotQA, TrQA, PRe, Avg)
30.94MFQA ScoreTidalDecode
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| TidalDecodeBudget (K)=40962025.08 | 30.94 | 13.85 | 13.71 | 86.3 | 78 | 44.56 | |
| LessIsMoreBudget (K)=40962025.08 | 30.9 | 14.34 | 12.58 | 87.06 | 79 | 44.78 | |
| Full AttentionBudget (K)=Full2025.08 | 30.76 | 14.56 | 11.5 | 86.56 | 77 | 44.08 | |
| LessIsMoreBudget (K)=10242025.08 | 29.87 | 14.2 | 12.04 | 87.42 | 75.58 | 43.82 | |
| QuestBudget (K)=40962025.08 | 28.92 | 13.63 | 12.15 | 85.91 | 72.5 | 42.62 | |
| TidalDecodeBudget (K)=10242025.08 | 28.57 | 11.11 | 9.82 | 79.78 | 75.17 | 40.89 | |
| QuestBudget (K)=10242025.08 | 26.21 | 12.19 | 10.75 | 83.47 | 63.84 | 39.29 |