Multi-turn Agentic Tool-use on TauBench Retail
50.3Success Rate (SR)HERO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| HEROModel=Qwen3-30B-A3B-Instruct2026.06 | 50.3 | 10.7 | |
| GRPOModel=Qwen3-30B-A3B-Instruct2026.06 | 47.8 | 10.8 | |
| HEROModel=Qwen3-4B-Instruct2026.06 | 34.7 | 9.6 | |
| Environment Feedback OnlyModel=Qwen3-30B-A3B-Instruct2026.06 | 34.5 | 12.2 | |
| BaseModel=Qwen3-30B-A3B-Instruct2026.06 | 33.6 | 11.9 | |
| GRPOModel=Qwen3-4B-Instruct2026.06 | 33.3 | 13.4 | |
| Environment Feedback OnlyModel=Qwen3-4B-Instruct2026.06 | 30.4 | 11.5 | |
| BaseModel=Qwen3-4B-Instruct2026.06 | 27.2 | 10.5 | |
| Full-Demo Privileged TeacherModel=Qwen3-30B-A3B-Instruct2026.06 | 26.1 | 12.6 | |
| Full-Demo Privileged TeacherModel=Qwen3-4B-Instruct2026.06 | 22.5 | 12.4 |