Spoken Dialogue System (SDS) Semantic Quality Evaluation on Eval2000 (test)
12.1ROUGE-LMulti turn CoT E2E
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Multi turn CoT E2Etype=E2E baseline2026.01 | 12.1 | 21.2 | 68.3 | 6.18 | 10.2 | — | |
| Multi turn CoT E2E + RLAIF (Single-Reward)Backbone=Multi turn CoT E2E, Post-training=RLAIF (Single-Reward)2026.01 | 11.9 | 19.9 | 56.5 | 6.33 | 7.1 | 55.4 | |
| Multi turn CoT E2E + RLAIF (Joint-Reward-v2)Backbone=Multi turn CoT E2E, Post-training=RLAIF (Joint-Reward-v2)2026.01 | 11.9 | 19.9 | 59.9 | 6.33 | 7.5 | 54.4 | |
| Multi turn CoT E2E + RLAIF (Joint-Reward-v1)Backbone=Multi turn CoT E2E, Post-training=RLAIF (Joint-Reward-v1)2026.01 | 11.8 | 19.6 | 61.3 | 6.29 | 8.5 | 52.6 | |
| Direct E2Etype=E2E baseline2026.01 | 8.4 | 302.2 | 51.5 | 5.5 | 24.2 | — | |
| Moshitype=Duplex SDS baseline2026.01 | 8.1 | 136.5 | 57.8 | 5.71 | 21 | — |