Empathetic Dialogue Response Generation on AvaMERG
76.3Win RateES4R
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| ES4RComparison=vs. w/o Dual-Attn, Aspect=Empathy Depth2026.01 | 76.3 | 18.8 | 4.9 | |
| ES4RComparison Baseline=ES4R w/o Emp. TTS, Evaluation Aspect=Empathy Expressiveness (DMOS-E)2026.01 | 76.2 | 18.4 | 5.4 | |
| ES4RComparison Baseline=Llama-Omni2, Evaluation Aspect=Empathy Expressiveness (DMOS-E)2026.01 | 75.6 | 18.2 | 6.2 | |
| ES4RComparison=vs. w/o Dual-Attn, Aspect=Response Specificity2026.01 | 74.9 | 19.1 | 6 | |
| ES4RComparison=vs. w/o Dual-Attn, Aspect=Topic Understanding2026.01 | 72.7 | 21.3 | 6 | |
| ES4RComparison=vs. w/o Dual-Attn, Aspect=Actionable Advice2026.01 | 71.8 | 22.1 | 6.1 | |
| ES4RComparison=vs. w/o Dual-Attn, Aspect=Emotion Recognition2026.01 | 70.5 | 23.2 | 6.3 | |
| ES4RComparison=vs. Llama-Omni2, Aspect=Response Specificity2026.01 | 70.3 | 23.2 | 6.5 | |
| ES4RComparison Baseline=Llama-Omni2, Evaluation Aspect=Emotional Consistency (DMOS-C)2026.01 | 69.7 | 23.1 | 7.2 | |
| ES4RComparison=vs. Llama-Omni2, Aspect=Empathy Depth2026.01 | 69.1 | 24.3 | 6.6 | |
| ES4RComparison=vs. Llama-Omni2, Aspect=Topic Understanding2026.01 | 68.5 | 24.8 | 6.7 | |
| ES4RComparison Baseline=OpenS2S, Evaluation Aspect=Empathy Expressiveness (DMOS-E)2026.01 | 68.4 | 23.7 | 7.9 | |
| ES4RComparison=vs. Llama-Omni2, Aspect=Actionable Advice2026.01 | 67.7 | 25.4 | 6.9 | |
| ES4RComparison Baseline=ES4R w/o Emp. TTS, Evaluation Aspect=Emotional Consistency (DMOS-C)2026.01 | 67.3 | 25.8 | 6.9 | |
| ES4RComparison=vs. Llama-Omni2, Aspect=Emotion Recognition2026.01 | 66.9 | 26.1 | 7 | |
| ES4RComparison=vs. w/o Cross-Attn, Aspect=Response Specificity2026.01 | 66.9 | 25.7 | 7.4 | |
| ES4RComparison=vs. w/o Cross-Attn, Aspect=Empathy Depth2026.01 | 65.1 | 27.5 | 7.4 | |
| ES4RComparison=vs. w/o Cross-Attn, Aspect=Topic Understanding2026.01 | 64.2 | 28.3 | 7.5 | |
| ES4RComparison=vs. w/o Cross-Attn, Aspect=Actionable Advice2026.01 | 63.6 | 28.9 | 7.5 | |
| ES4RComparison=vs. w/o Cross-Attn, Aspect=Emotion Recognition2026.01 | 62.5 | 29.8 | 7.7 | |
| ES4RComparison Baseline=OpenS2S, Evaluation Aspect=Emotional Consistency (DMOS-C)2026.01 | 61.2 | 29.5 | 9.3 | |
| ES4RComparison=vs. OpenS2S, Aspect=Response Specificity2026.01 | 60.8 | 30.2 | 9 | |
| ES4RComparison=vs. OpenS2S, Aspect=Empathy Depth2026.01 | 59.1 | 31.6 | 9.3 | |
| ES4RComparison Baseline=ES4R w/o Emp. TTS, Evaluation Aspect=Speech Quality (DMOS-Q)2026.01 | 58.9 | 32.4 | 8.7 | |
| ES4RComparison=vs. OpenS2S, Aspect=Topic Understanding2026.01 | 58.3 | 32.1 | 9.6 | |
| ES4RComparison=vs. OpenS2S, Aspect=Actionable Advice2026.01 | 57.4 | 33.2 | 9.4 | |
| ES4RComparison=vs. OpenS2S, Aspect=Emotion Recognition2026.01 | 56.2 | 34.5 | 9.3 | |
| ES4RComparison Baseline=Llama-Omni2, Evaluation Aspect=Speech Quality (DMOS-Q)2026.01 | 48.3 | 40.1 | 11.6 | |
| ES4RComparison Baseline=OpenS2S, Evaluation Aspect=Speech Quality (DMOS-Q)2026.01 | 47.8 | 35.4 | 16.8 |