Agentic task on τ2-Bench Airline
60Avg@4PPO-EWMA
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PPO-EWMABackbone=Qwen3-30B-A3B2026.05 | 60 | 82 | |
| SnapshotBackbone=Qwen3-30B-A3B2026.05 | 59 | 80 | |
| Decoupled PPOBackbone=Qwen3-30B-A3B2026.05 | 57 | 76 | |
| SnapshotBackbone=Qwen3-4B2026.05 | 56 | 76 | |
| Linear_proxBackbone=Qwen3-4B2026.05 | 54 | 72 | |
| PPO-EWMABackbone=Qwen3-4B2026.05 | 54 | 74 | |
| Decoupled PPOBackbone=Qwen3-4B2026.05 | 53.5 | 72 | |
| Linear_proxBackbone=Qwen3-30B-A3B2026.05 | 53.5 | 74 |