Web Agent Task Completion on XBench DeepSearch
79AccuracySlimSearcher (SFT+RL)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SlimSearcher (SFT+RL)Backbone Policy=Tongyi-DeepResearch, Optimization Strategy=SFT+RL2026.06 | 79 | 5.92 | 4,888 | |
| WebLeaperType=Open-weights Agent2026.06 | 78 | 14.06 | 4,466.5 | |
| SlimSearcher (SFT+RL)Backbone Policy=Qwen3-30B-A3B-Instruct-2507, Optimization Strategy=SFT+RL2026.06 | 77 | 10.87 | 5,092 | |
| SlimSearcher (SFT)Backbone Policy=Tongyi-DeepResearch, Optimization Strategy=SFT2026.06 | 77 | 10.22 | 5,155 | |
| SlimSearcher (SFT)Backbone Policy=Qwen3-30B-A3B-Instruct-2507, Optimization Strategy=SFT2026.06 | 74 | 13.44 | 5,515 | |
| Tongyi-DeepResearch2026.06 | 71.3 | 14.26 | 6,918 | |
| Prompt ControlBackbone Policy=Tongyi-DeepResearch, Optimization Strategy=Prompt Control2026.06 | 67.6 | 12.5 | 6,321 | |
| OpenAI o3Type=Proprietary System2026.06 | 67 | — | — | |
| Claude-4-SonnetType=Proprietary System2026.06 | 64.6 | — | — | |
| Qwen3-30B-A3B-Instruct-2507Type=Open-weights Agent, Unified Environment=true2026.06 | 55 | 15.04 | 2,395 | |
| Kimi-K2-Instruct-0905Type=Open-weights Agent2026.06 | 54 | 5.98 | 1,316 | |
| DeepSeek-V3Type=Open-weights Agent2026.06 | 53 | 2.23 | 422 | |
| WebExplorerType=Open-weights Agent2026.06 | 51.7 | 9.05 | 2,246 | |
| Qwen3-235B-A22B-InstructType=Open-weights Agent2026.06 | 49 | 8.84 | 938 |