Needle-in-a-haystack Context Scaling on NIAH
100NIAH Success Rate (32K Context)Qwen3-4B
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Qwen3-4BScaling=YaRN2026.02 | 100 | 100 | 88.33 | 41.67 | 23.33 | 16.67 | 3.33 | 1.7 | |
| Qwen3-8BScaling=YaRN2026.02 | 100 | 100 | 88.33 | 41.67 | 23.33 | 16.67 | 3.33 | 1.7 | |
| Qwen3-14BScaling=YaRN2026.02 | 100 | 100 | 88.33 | 41.67 | 23.33 | 16.67 | 3.33 | 1.7 | |
| StateLM-14BSearch mechanism=w/o search2026.02 | 100 | 100 | 99.44 | 97.78 | 89.45 | 94.44 | 95 | 83.89 | |
| StateLM-8BSearch mechanism=w/o search2026.02 | 99.44 | 98.33 | 98.33 | 99.44 | 95.55 | 88.89 | 88.89 | 67.22 | |
| StateLM-4BSearch mechanism=w/o search2026.02 | 95 | 95.56 | 95.56 | 88.33 | 76.67 | 62.78 | 53.89 | 32.22 |