ParaIntent - Response Quality: leaderboard

Metric: Response quality score (%; held-out DeepSeek-V4-Pro judge of content, tone appropriateness and naturalness; over the ParaIntent synthetic test set (14 Chinese spoken-emotional-dialogue intent categories, balanced explicit and implicit samples); single-turn spoken response). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.

Top models

#ModelScore
1Qwen3-Omni58.65
2MiMo-Audio52.91
3Step-Audio-2-mini45.87
4GLM-4-Voice45.18
5Qwen2.5-Omni35.68

Interactive version: theaggregate.ai/benchmark?slug=paraintent-response-quality · How It Works · Data refreshed daily, snapshot 2026-09-26.