Agentic Task Evaluation on τ2-Bench Retail (avg@4, pass@4)
69.7Avg@4Snapshot
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SnapshotBackbone=Qwen3-30B-A3B2026.05 | 69.7 | 92.1 | |
| PPO-EWMABackbone=Qwen3-30B-A3B2026.05 | 67.82 | 92.1 | |
| SnapshotBackbone=Qwen3-4B2026.05 | 66.23 | 89.47 | |
| Linear_proxBackbone=Qwen3-30B-A3B2026.05 | 65.8 | 87.7 | |
| PPO-EWMABackbone=Qwen3-4B2026.05 | 65.72 | 90.35 | |
| Decoupled PPOBackbone=Qwen3-30B-A3B2026.05 | 65.43 | 89.47 | |
| Linear_proxBackbone=Qwen3-4B2026.05 | 64.4 | 86.84 | |
| Decoupled PPOBackbone=Qwen3-4B2026.05 | 63.96 | 88.6 |