Turn-taking decision on RolePlayConv zero-shot (test)
79PrecisionModeratorLM-Think
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| ModeratorLM-ThinkRole conditioning=true, Reasoning (Thinking)=true, Chunking strategy=dynamic, Decoding strategy=Sampling for reasoning2026.06 | 79 | 82 | 79 | 91 | 3 | 3 | |
| ModeratorLMRole conditioning=true, Reasoning (Thinking)=false, Chunking strategy=dynamic, Decoding strategy=Greedy2026.06 | 71 | 57 | 61 | 76 | 5 | 14 | |
| MP-BaselineRole conditioning=false, Chunking strategy=dynamic2026.06 | 40 | 48 | 42 | 67 | 14 | — | |
| MoshiRole conditioning=false, Chunking strategy=dynamic, Variant=Moshika, Tolerance window=±0.5 s2026.06 | 15 | 34 | 21 | 50 | 47 | — |