PersTurnBench: leaderboard

PersTurnBench (2026) replays turns of real users conversations: the candidate model writes the next assistant reply and a verified personalized evaluator with user memory predicts how satisfied that user would be, calibrated per user. The score is user-macro satisfaction on a 1 to 5 scale, averaged within and then across users; higher is better.

Metric: User-macro satisfaction (1-5): calibrated predicted satisfaction averaged within each user and then across users, PersTurnBench replay of 300 turns from 90 users (180 user-task blocks): the candidate fills only the current assistant turn and the frozen Qwen3-8B memory evaluator predicts the user's turn satisfaction, reference-CDF calibrated per user; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Kimi K2.64.81#197
2GLM-5.14.8#200
3DeepSeek V4 Pro4.79#201
4GPT-5.54.69#56
5Claude Opus 4.74.64#86
6Gemini 3.1 Pro (Preview)4.63#102
7Qwen 3 8B4.47#1890
8MiniMax-M2.74.43#539
9DeepSeek V4 Flash4.42#273
10Gemini 3.1 Flash Lite4.34#377

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=persturnbench · How It Works · Data refreshed daily, snapshot 2026-11-22.