Long-horizon agentic tasks on BrowseComp Our Settings
62.5Pass@1AgentSwing
Evaluation results
| Method | Links | |
|---|---|---|
| AgentSwingBase Model=DeepSeek-v3.22026.03 | 62.5 | |
| AgentSwingBase Model=Tongyi-DR-30B-A3B2026.03 | 60.5 | |
| AgentSwingBase Model=GPT-OSS-120B2026.03 | 60 | |
| DeepSeek-v3.2Context Management=Discard-All2026.03 | 58 | |
| Tongyi-DR-30B-A3BContext Management=Discard-All2026.03 | 58 | |
| Tongyi-DR-30B-A3BContext Management=Summary2026.03 | 55 | |
| Tongyi-DR-30B-A3BContext Management=Keep-Last-N2026.03 | 53 | |
| GPT-OSS-120BContext Management=Keep-Last-N2026.03 | 52.5 | |
| DeepSeek-v3.2Context Management=Keep-Last-N2026.03 | 52 | |
| GPT-OSS-120BContext Management=Discard-All2026.03 | 50.5 | |
| DeepSeek-v3.2Context Management=Summary2026.03 | 48.5 | |
| GPT-OSS-120BContext Management=Summary2026.03 | 48 | |
| Tongyi-DR-30B-A3BContext Management=Baseline (w/o CM)2026.03 | 48 | |
| DeepSeek-v3.2Context Management=Baseline (w/o CM)2026.03 | 43.5 | |
| GPT-OSS-120BContext Management=Baseline (w/o CM)2026.03 | 39.5 |