Agentic Tool Use on Multi-step Tool-Calling Tasks (Model Aggregate Test)
99Success RateGold CMTF
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Gold CMTFTool Filtering Strategy=Gold-contract CMTF2026.06 | 99 | 0.01 | 0 | 1 | 2,532 | |
| Learned CMTFTool Filtering Strategy=Learned contract CMTF (Opus 4.8 Hybrid)2026.06 | 98 | 0.02 | 0 | 1 | 2,528 | |
| All toolsTool Filtering Strategy=No filtering (full exposure)2026.06 | 77.5 | 1.461 | 2.7 | 100 | 26,172 | |
| Keyword top-5Tool Filtering Strategy=Keyword-based2026.06 | 62 | 2.265 | 3.2 | 5 | 4,482 | |
| State-awareTool Filtering Strategy=State-aware filtering2026.06 | 62 | 2.196 | 0 | 5.733 | 4,648 |