EYT-Bench (PersonaMem-v2) - Empathy: leaderboard
EYT-Bench (Enjoy Your Talk, 2026) evaluates LLMs as multi-turn conversational partners with a persona-grounded user simulator and an independent LLM judge; each turn the target model predicts the user state before replying. Judge-rated empathy of the replies on a 0 to 100 scale on the PersonaMem-v2 pool of free-text personas distilled from real conversations; higher is better.
Metric: Empathy (recognition, attribution, resonance, response, support) (0-100): judge rubric of five sub-indicators scored 0, 0.5 or 1 per turn, summed to 0-5, turn weighted with the first two turns at 0.10 each and rescaled to 0-100; multi-turn dialogues of up to 10 turns with a persona-grounded Gemini 3.1 Pro user simulator and a Gemini 3.1 Pro judge (reasoning effort high, temperature 0); PersonaMem-v2 pool (100 dialogues; personas are free-text paragraphs distilled from long real user-assistant interactions). Source: arxiv.org. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B | 73.1 | #246 |
| 2 | Qwen 3.5 27B | 71.3 | #359 |
| 3 | Qwen 3.5 35B A3B | 70.1 | #482 |
| 4 | Seed 2.0 Pro | 69.4 | #211 |
| 5 | DeepSeek V4 Pro | 67.8 | #188 |
| 6 | Seed 2.0 Lite | 67.8 | #289 |
| 7 | Gemini 3.1 Pro (Preview) (Thinking) | 67.8 | #94 (Gemini 3.1 Pro (Preview)) |
| 8 | DeepSeek V4 Flash | 67.6 | #258 |
| 9 | Claude Opus 4.7 | 65.3 | #79 |
| 10 | Claude Sonnet 4.6 | 62.9 | #161 |
| 11 | Seed 2.0 Mini | 57.6 | #508 |
| 12 | GPT-5.5 | 55.9 | #49 |
Interactive version: theaggregate.ai/benchmark?slug=eyt-bench-personamem-v2-empathy · How It Works · Data refreshed daily, snapshot 2026-11-13.