Multi-hop Question Answering on 2WikiMultiHopQA (Context Window Scaling)
73.4Performance (8K Context)Qwen3 + CDT
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3 + CDTScale=4B2026.05 | 73.4 | 61.7 | 44.5 | 39.8 | 37.5 | — | — | — | — | |
| Qwen3 + MemRereadScale=4B2026.05 | 70.3 | 71.1 | 59.4 | 64.1 | 54.7 | 55.5 | 46.9 | 45.3 | 58.4 | |
| MemRereadScale=4B2026.05 | 70.3 | 71.1 | 59.4 | 64.1 | 54.7 | 55.5 | 46.9 | 45.3 | 58.4 | |
| Qwen3Scale=4B2026.05 | 67.2 | 60.9 | 52.3 | 43.8 | 46.9 | — | — | — | — | |
| InfMemScale=4B2026.05 | 53.1 | 62.5 | 55.5 | 53.1 | 55.5 | 68 | 53.1 | 49.2 | 56.3 |