EYT-Bench (PersonaMem-v2) - Latent Intent: leaderboard
Metric: Per-turn latent-intent accuracy (%; the target model first predicts the user latent-intent label from 8 classes, scored against the simulator gold label, before generating its reply; 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. Saturation forecast: Around December 2026. 17 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | DeepSeek V4 Pro | 79.9 |
| 2 | DeepSeek V4 Flash | 70.1 |
| 3 | GPT-5.5 | 47.9 |
| 4 | Gemini 3.1 Pro (Preview) (Thinking) | 47.3 |
| 5 | Claude Opus 4.7 | 41.5 |
| 6 | Claude Sonnet 4.6 | 24.9 |
| 7 | Qwen 3.5 397B A17B | 15.3 |
| 8 | Qwen 3.5 35B A3B | 14.2 |
| 9 | Qwen 3.5 27B | 13.1 |
| 10 | Seed 2.0 Pro | 12 |
| 11 | Seed 2.0 Mini | 11.7 |
| 12 | Seed 2.0 Lite | 7.7 |
Interactive version: theaggregate.ai/benchmark?slug=eyt-bench-personamem-v2-latent-intent · How It Works · Data refreshed daily, snapshot 2026-09-29.