Skill Decomposition and Retrieval on COMPSKILLBENCH 50 queries 1.0
92DASKILLWEAVER (qwen-max + SAD)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SKILLWEAVER (qwen-max + SAD)Backbone=qwen-max, Method Variation=+ SAD2026.06 | 92 | 98 | 39.4 | |
| LLM-Direct (100 skills shown)Backbone=qwen-max, Evaluation Protocol=LLM-Direct, Skill Prompting Constraints=100 skills shown2026.06 | 90 | 96 | 21.1 | |
| SKILLWEAVER (qwen-max Vanilla)Backbone=qwen-max, Method Variation=Vanilla2026.06 | 66 | 82 | 35.9 | |
| ReAct-style (iterative)Backbone=qwen-max, Evaluation Protocol=ReAct-style (iterative)2026.06 | 0 | 4 | 15.4 |