Agentic Task on τ2-Bench Telecom (avg@2, pass@2)
45Avg@2 ScorePPO-EWMA
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PPO-EWMABackbone=Qwen3-30B-A3B2026.05 | 45 | 57.5 | |
| SnapshotBackbone=Qwen3-30B-A3B2026.05 | 45 | 57.5 | |
| Decoupled PPOBackbone=Qwen3-30B-A3B2026.05 | 44.75 | 55 | |
| Linear_proxBackbone=Qwen3-30B-A3B2026.05 | 44 | 55 | |
| PPO-EWMABackbone=Qwen3-4B2026.05 | 42.5 | 52.5 | |
| SnapshotBackbone=Qwen3-4B2026.05 | 42.5 | 52.5 | |
| Decoupled PPOBackbone=Qwen3-4B2026.05 | 40 | 50 | |
| Linear_proxBackbone=Qwen3-4B2026.05 | 37.5 | 50 |