Needle Retrieval on NIAH (Context Length Sweep)
100Success Rate (8K, Single)Qwen3 w/ MATCH
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3 w/ MATCHBackbone=Qwen3, Sparsification Stage=Post-sparsified, k=42026.06 | 100 | 92.4 | 99.7 | 81.5 | 93 | 59.7 | 84.7 | 29 | 80.7 | 21 | 69.8 | 85.2 | |
| Qwen3 w/ MATCHBackbone=Qwen3, Sparsification Stage=Post-sparsified, k=82026.06 | 99 | 93 | 99 | 82.3 | 91.3 | 55.4 | 77.3 | 26.6 | 71 | 19.9 | 67.5 | 84.2 | |
| Qwen3Backbone=Qwen3, Sparsification Stage=Post-sparsified2026.06 | 87 | 91.9 | 80 | 84.1 | 72.7 | 58 | 90 | 24.4 | 47 | 13.6 | 60.3 | 78.7 | |
| Phi-3 w/ MATCHBackbone=Phi-3, Sparsification Stage=Pre-sparsified, k=82026.06 | 80 | 47.5 | 78.7 | 45 | 77 | 36.7 | — | — | — | — | 56.4 | 56.4 | |
| Phi-3 w/ MATCHBackbone=Phi-3, Sparsification Stage=Pre-sparsified, k=42026.06 | 78 | 43.6 | 75.3 | 40.5 | 75.3 | 36.9 | — | — | — | — | 53.8 | 53.8 | |
| Phi-3Backbone=Phi-3, Sparsification Stage=Pre-sparsified2026.06 | 24.3 | 18.2 | 14.3 | 8.2 | 3.7 | 4.2 | — | — | — | — | 11.7 | 11.7 |