Intent Coverage and Efficiency on MultiWOZ 2.2 (held-out)
85.6Coverage AUCBN-Guided
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| BN-GuidedThreshold (τ)=0.5, top-k=5, Ground-truth replay=true, Backbone=MiniMax-M2.5, Description=coverage augmented by BN top-5 predictions2026.04 | 85.6 | 2.73 | 1.04 | 53.8 | 18.5 | |
| BaselineThreshold (τ)=0.5, top-k=5, Ground-truth replay=true, Description=cumulative coverage using only actual user intents2026.04 | 74.2 | 3.95 | — | — | — |