Long-context Recall on SQuAD, FDA, and SWDE Subset
72.3SQuAD Recall AccuracyQwen3
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Qwen3Backbone=8B, LS. Tokens=-, Zero-shot=true, HypeNet Setting=true2026.06 | 72.3 | 78.2 | 90.8 | 80.4 | |
| Qwen3Backbone=30B-A3B, LS. Tokens=-, Zero-shot=true, HypeNet Setting=true2026.06 | 58.7 | 81 | 90.8 | 76.8 | |
| HALOModel Scale=1.7B, Attention Type=GDN, LS. Tokens=234M2026.06 | 54.5 | 68.2 | 80.9 | 67.9 | |
| FlashMorphModel Scale=1.7B, Attention Type=GDN, LS. Tokens=20M2026.06 | 54.3 | 74.1 | 82.2 | 70.2 | |
| PostNAS*Model Scale=1.7B, Attention Type=GLA, LS. Tokens=50B2026.06 | 54.2 | 69.4 | 81.9 | 68.5 | |
| KL-LSModel Scale=1.7B, Attention Type=GDN, LS. Tokens=20B2026.06 | 54.1 | 73.5 | 81.2 | 69.6 | |
| PostNAS*Model Scale=1.7B, Attention Type=GDN, LS. Tokens=50B2026.06 | 52.8 | 67.9 | 82.4 | 67.7 | |
| FlashMorphBackbone=8B, LS. Tokens=20M, Zero-shot=true, HypeNet Setting=true2026.06 | 52.7 | 73.5 | 87.3 | 71.2 | |
| PostNAS*Model Scale=1.7B, Attention Type=Lightning, LS. Tokens=50B2026.06 | 51.3 | 64.7 | 80.9 | 65.7 | |
| HALOModel Scale=1.7B, Attention Type=Lightning, LS. Tokens=234M2026.06 | 51.3 | 70.5 | 82.5 | 68.1 | |
| FlashMorphModel Scale=1.7B, Attention Type=Lightning, LS. Tokens=20M2026.06 | 51.1 | 71.6 | 81.6 | 68.1 | |
| UniformBackbone=8B, LS. Tokens=N/A, Zero-shot=true, HypeNet Setting=true2026.06 | 49.7 | 63.3 | 84.1 | 65.7 | |
| UniformModel Scale=1.7B, Attention Type=GDN, LS. Tokens=N/A2026.06 | 48.2 | 59.2 | 78.9 | 62.1 | |
| FlashMorphModel Scale=1.7B, Attention Type=GLA, LS. Tokens=20M2026.06 | 47.3 | 73.7 | 82.3 | 67.7 | |
| Qwen3Model Scale=0.6B, Attention Type=Full2026.06 | 44.1 | 82.1 | 80.5 | 68.9 | |
| KL-LSModel Scale=1.7B, Attention Type=GLA, LS. Tokens=20B2026.06 | 44.1 | 53.5 | 78.6 | 58.7 | |
| Qwen3Backbone=0.6B, LS. Tokens=-, Zero-shot=true, HypeNet Setting=true2026.06 | 44.1 | 82.1 | 80.5 | 68.9 | |
| HALOModel Scale=1.7B, Attention Type=GLA, LS. Tokens=234M2026.06 | 43.7 | 41.4 | 74.2 | 53.1 | |
| UniformModel Scale=1.7B, Attention Type=GLA, LS. Tokens=N/A2026.06 | 43.6 | 57.3 | 81.2 | 60.7 | |
| PostNAS*Backbone=1.7B, LS. Tokens=50B, Zero-shot=true, HypeNet Setting=true2026.06 | 43.5 | 63.5 | 81.5 | 62.8 | |
| HALOBackbone=8B, LS. Tokens=234M, Zero-shot=true, HypeNet Setting=true2026.06 | 41.9 | 59.7 | 80.9 | 60.8 | |
| FlashMorphModel Scale=0.6B, Attention Type=Lightning, LS. Tokens=20M2026.06 | 41.7 | 62.4 | 76.2 | 60.1 | |
| UniformBackbone=1.7B, LS. Tokens=N/A, Zero-shot=true, HypeNet Setting=true2026.06 | 40.2 | 64.5 | 80.4 | 61.7 | |
| Qwen3Model Scale=1.7B, Attention Type=Full2026.06 | 39.8 | 79 | 85.1 | 67.9 | |
| Qwen3Backbone=1.7B, LS. Tokens=-, Zero-shot=true, HypeNet Setting=true2026.06 | 39.8 | 79 | 85.1 | 67.9 | |
| FlashMorphBackbone=1.7B, LS. Tokens=20M, Zero-shot=true, HypeNet Setting=true2026.06 | 39.8 | 70.2 | 81.4 | 63.8 | |
| KL-LSModel Scale=1.7B, Attention Type=Lightning, LS. Tokens=20B2026.06 | 39.5 | 49.4 | 74.8 | 54.5 | |
| UniformModel Scale=1.7B, Attention Type=Lightning, LS. Tokens=N/A2026.06 | 39 | 56.6 | 78.6 | 58.1 | |
| HALOBackbone=1.7B, LS. Tokens=234M, Zero-shot=true, HypeNet Setting=true2026.06 | 38.5 | 67.8 | 80.7 | 62.3 | |
| FlashMorphModel Scale=0.6B, Attention Type=GDN, LS. Tokens=20M2026.06 | 38.4 | 71.3 | 76.7 | 62.1 | |
| HALOModel Scale=0.6B, Attention Type=GLA, LS. Tokens=234M2026.06 | 36.4 | 68.5 | 75.4 | 60.1 | |
| FlashMorphModel Scale=0.6B, Attention Type=GLA, LS. Tokens=20M2026.06 | 35.4 | 70.7 | 76 | 60.7 | |
| FlashMorphBackbone=0.6B, LS. Tokens=20M, Zero-shot=true, HypeNet Setting=true2026.06 | 35.3 | 58.2 | 76.8 | 56.7 | |
| HALOModel Scale=0.6B, Attention Type=Lightning, LS. Tokens=234M2026.06 | 34.7 | 60.4 | 72.6 | 55.9 | |
| KL-LSModel Scale=0.6B, Attention Type=GDN, LS. Tokens=20B2026.06 | 33.5 | 72.8 | 75.6 | 60.6 | |
| KL-LSModel Scale=0.6B, Attention Type=GLA, LS. Tokens=20B2026.06 | 32.7 | 61.4 | 75.2 | 56.4 | |
| UniformModel Scale=0.6B, Attention Type=GLA, LS. Tokens=N/A2026.06 | 31.1 | 55.6 | 75 | 53.9 | |
| UniformModel Scale=0.6B, Attention Type=Lightning, LS. Tokens=N/A2026.06 | 30.3 | 51.3 | 71.8 | 51.1 | |
| UniformModel Scale=0.6B, Attention Type=GDN, LS. Tokens=N/A2026.06 | 30.3 | 55.5 | 72.2 | 52.7 | |
| KL-LSModel Scale=0.6B, Attention Type=Lightning, LS. Tokens=20B2026.06 | 29.3 | 58.5 | 70.5 | 52.8 | |
| KL-LSBackbone=1.7B, LS. Tokens=20B, Zero-shot=true, HypeNet Setting=true2026.06 | 26.7 | 48.8 | 73.9 | 49.8 | |
| HALOModel Scale=0.6B, Attention Type=GDN, LS. Tokens=234M2026.06 | 26.5 | 62 | 71.2 | 53.2 | |
| KL-LSBackbone=0.6B, LS. Tokens=20B, Zero-shot=true, HypeNet Setting=true2026.06 | 25.7 | 58.4 | 70.8 | 51.6 | |
| UniformBackbone=30B-A3B, LS. Tokens=N/A, Zero-shot=true, HypeNet Setting=true2026.06 | 23 | 35.9 | 76.8 | 45.2 | |
| HALOBackbone=0.6B, LS. Tokens=234M, Zero-shot=true, HypeNet Setting=true2026.06 | 22.8 | 62.8 | 73.5 | 53 | |
| FlashMorphBackbone=30B-A3B, LS. Tokens=20M, Zero-shot=true, HypeNet Setting=true2026.06 | 21.4 | 53.4 | 81.5 | 52.1 | |
| HALOBackbone=30B-A3B, LS. Tokens=234M, Zero-shot=true, HypeNet Setting=true2026.06 | 20.8 | 41 | 77 | 46.3 | |
| UniformBackbone=0.6B, LS. Tokens=N/A, Zero-shot=true, HypeNet Setting=true2026.06 | 20.7 | 52.6 | 70.6 | 48 |