Single-Objective Task Evaluation on 2WikiMultihopQA, HotpotQA, Bamboogle, Frames, and BrowseComp-Plus
77.5Accuracy (2WikiMultihopQA)Search-R1 (SFT + RL)
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Search-R1 (SFT + RL)Backbone=Qwen2.5-14B-Instruct, Training Protocol (SFT/RL)=SFT + RL2025.10 | 77.5 | 72.3 | 62.4 | 37.6 | 17.7 | 53.5 | 19.3 | |
| MemAct (SFT+RL)Backbone=Qwen2.5-14B-Instruct, Training Protocol (SFT/RL)=SFT + RL, Memory Management Strategy=MemAct2025.10 | 76.7 | 71 | 61.8 | 38.5 | 20.7 | 53.7 | 8.2 | |
| MemAct (SFT)Backbone=Qwen2.5-14B-Instruct, Training Protocol (SFT/RL)=SFT, Memory Management Strategy=MemAct2025.10 | 76.4 | 70.5 | 61.6 | 35.9 | 16 | 52.1 | 9.4 | |
| Qwen3-235BBackbone=Qwen3-235B2025.10 | 60.3 | 69 | 50.5 | 43 | 26.8 | 50 | 16.7 | |
| Qwen2.5-14BBackbone=Qwen2.5-14B-Instruct2025.10 | 58 | 65.5 | 48.8 | 27.5 | 11.1 | 42.1 | 6.5 | |
| MEM1 (RL)Backbone=Qwen2.5-14B-Instruct, Training Protocol (SFT/RL)=RL2025.10 | 56.5 | 66 | 47 | 28.5 | 14.1 | 42.4 | 5 | |
| SummarizationBackbone=Qwen2.5-14B-Instruct, Memory Management Strategy=Summarization2025.10 | 54 | 69.2 | 44.2 | 33.5 | 12 | 42.6 | 4.1 | |
| Sliding WindowBackbone=Qwen2.5-14B-Instruct, Memory Management Strategy=Sliding Window2025.10 | 53.5 | 56 | 47.2 | 21.5 | 8.5 | 37.3 | 3.8 | |
| A-MEMBackbone=Qwen2.5-14B-Instruct, Memory Management Strategy=A-MEM2025.10 | 52.8 | 59.1 | 45.3 | 25 | 9.4 | 38.3 | 3.9 |