Agent Performance on AgentNoiseBench (Noisy setting tau2 1.0, test)
43.2Retail Avg@4 ScoreNoisyAgent
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| NoisyAgentBackbone Model=Qwen3-32B, Training Method=NoisyAgent2026.05 | 43.2 | 65.79 | 46 | 70 | 43.42 | 70.18 | |
| GRPOBackbone Model=Qwen3-32B, Training Method=GRPO2026.05 | 38.16 | 61.4 | 37 | 62 | 36.84 | 62.28 | |
| GSPOBackbone Model=Qwen3-32B, Training Method=GSPO2026.05 | 37.72 | 60.53 | 39 | 64 | 39.25 | 65.79 | |
| NoisyAgentBackbone Model=Qwen3-8B, Training Method=NoisyAgent2026.05 | 36.4 | 61.4 | 38 | 56 | 38.38 | 64.91 | |
| DAPOBackbone Model=Qwen3-32B, Training Method=DAPO2026.05 | 36.18 | 57.89 | 39.5 | 66 | 38.16 | 66.67 | |
| GSPOBackbone Model=Qwen3-8B, Training Method=GSPO2026.05 | 31.8 | 54.39 | 32.5 | 52 | 34.43 | 56.14 | |
| Qwen3-32BBackbone Model=Qwen3-32B, Training Method=Base2026.05 | 31.14 | 52.63 | 31.5 | 56 | 26.54 | 45.61 | |
| GRPOBackbone Model=Qwen3-8B, Training Method=GRPO2026.05 | 30.48 | 50.88 | 33.5 | 54 | 31.58 | 53.51 | |
| DAPOBackbone Model=Qwen3-8B, Training Method=DAPO2026.05 | 29.39 | 53.51 | 31 | 50 | 34.21 | 57.89 | |
| Qwen3-8BBackbone Model=Qwen3-8B, Training Method=Base2026.05 | 24.12 | 44.74 | 23 | 42 | 21.05 | 41.23 |