Long-Context Retrieval on RULER
80.1AccuracyFull attention
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Full attentionBackbone Model=LLaMA-3.1-70B2026.03 | 80.1 | 0 | |
| Full attentionBackbone Model=LLaMA-3.1-8B2026.03 | 79.8 | 0 | |
| QuestBackbone Model=LLaMA-3.1-70B, Budget=8K/4K/8K2026.03 | 79.8 | 93 | |
| QuestBackbone Model=LLaMA-3.1-70B, Budget=4K/3K/4K2026.03 | 79.8 | 97 | |
| QuestBackbone Model=LLaMA-3.1-70B, Budget=2K2026.03 | 79.5 | 98 | |
| MAC-AttentionBackbone Model=LLaMA-3.1-8B2026.03 | 78.8 | 95 | |
| QuestBackbone Model=LLaMA-3.1-70B, Budget=1K2026.03 | 78.3 | 99 | |
| MAC-AttentionBackbone Model=LLaMA-3.1-70B2026.03 | 78 | 99 | |
| QuestBackbone Model=LLaMA-3.1-8B, Budget=8K/4K/8K2026.03 | 77 | 93 | |
| QuestBackbone Model=LLaMA-3.1-8B, Budget=4K/3K/4K2026.03 | 75.5 | 97 | |
| Full attentionBackbone Model=Phi-4-Mini2026.03 | 74.4 | 0 | |
| QuestBackbone Model=LLaMA-3.1-8B, Budget=2K2026.03 | 73.4 | 98 | |
| MAC-AttentionBackbone Model=Phi-4-Mini2026.03 | 73.1 | 77 | |
| QuestBackbone Model=LLaMA-3.1-8B, Budget=1K2026.03 | 72.4 | 99 |