Tool Retrieval on ToolBench Multi-Domain
86.07NDCG@1 (I1)ToolOmni
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| ToolOmniBackbone=Qwen3-4B-Instruct, Maximum retrieval turns=42026.04 | 86.07 | 84.09 | 84.8 | 80.63 | 72.96 | 73.91 | 78.9 | 71.44 | 71.82 | 78.29 | |
| ToolRetrieverMechanism=Dense retrieval2026.04 | 81.03 | 80.88 | 84.52 | 75.91 | 69.53 | 75.21 | 77.52 | 69.15 | 74.24 | 76.44 | |
| IterFeedbackMaximum retrieval turns=42026.04 | 71.77 | 70.49 | 74.82 | 63.18 | 55.41 | 61.32 | 67.43 | 59.23 | 63.14 | 65.2 | |
| ToolGenMechanism=Generative retrieval2026.04 | 68.84 | 72 | 78.76 | 46.77 | 53.55 | 62.43 | 75.68 | 75.26 | 84.48 | 68.64 | |
| EmbSim*Backbone=text-embedding-3-large, Source=Cited from ToolGen2026.04 | 54 | 50.82 | 55.86 | 40.84 | 36.67 | 39.55 | 18 | 17.77 | 20.7 | 37.13 | |
| BM25*Source=Cited from ToolGen2026.04 | 22.77 | 22.64 | 25.61 | 18.29 | 20.74 | 22.18 | 10 | 10.08 | 12.33 | 18.29 |