Multihop Question Answering on HotpotQA In-Distribution (Context Window Scaling)
59.4Accuracy (8K Context)MemReread
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| MemRereadScale=4B2026.05 | 59.4 | 57.8 | 51.6 | 46.9 | 49.2 | 53.1 | 51.6 | 54.7 | 53 | |
| ReMemR1Scale=4B2026.05 | 58.6 | 54.7 | 57 | 50 | 51.6 | 45.3 | 50 | 52.3 | 52.4 | |
| MemAgentScale=4B2026.05 | 53.9 | 55.5 | 46.9 | 50.8 | 43.8 | 43 | 43.8 | 42.2 | 47.5 | |
| MemAgentScale=1.7B2026.05 | 44.5 | 39.1 | 29.7 | 23.4 | 25 | 20.3 | 21.1 | 19.5 | 27.8 | |
| MemRereadScale=1.7B2026.05 | 43 | 43 | 32 | 23.4 | 22.7 | 22.7 | 23.4 | 17.2 | 28.4 | |
| ReMemR1Scale=1.7B2026.05 | 27.3 | 28.1 | 31.2 | 19.5 | 25.8 | 25.8 | 21.9 | 17.2 | 24.6 |