StanceBench - Compassion and Empathy: leaderboard
StanceBench tests whether audio-capable models can read interpersonal stance from conversational speech: for each stance dimension, speakers were role-prompted toward a positive or a negative pole, and the judge must tell them apart from the audio. This board covers compassionate and supportive versus callous delivery. The score is the area under the ROC curve of the signed stance score the judge gives, against the role-derived poles, so 0.5 is chance.
Metric: AUROC (0-1 scale) of the judge's conversation-level signed stance score in separating the 127 role-prompted positive-pole and negative-pole single-speaker instances of the Compassion and Empathy dimension (S1), 25 percent sample of the Seamless Interaction Improvised subset; binary pole choice asked in both pole orders, malformed outputs dropped. Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash (Thinking) | 0.96 | #453 (Gemini 2.5 Flash) |
| 2 | Qwen2.5-Omni-7B | 0.92 | #1789 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=stancebench-compassion-and-empathy · How It Works · Data refreshed daily, snapshot 2026-11-13.