iPrefBench Personalization
Agents · 2026-10-26
iPrefBench's personalization score: the fraction of each task's personalization criteria satisfied, each a user-dependent decision whose value must match the user's intent and is either inferable from the activity history or obtainable only by asking (about 2.9 per task), macro-averaged over the 130 tasks, in percent. Same tasks, runs and Claude Sonnet 5 judge as task success.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 69.4 |
| Sonnet 5 | 64.8 |
| Kimi K3 | 63.9 |
| Gemini 3.1 Pro Preview | 60.9 |
| Qwen3.8-Flash-Next | 58.6 |
| Gemini 3.7 Flash | 55.1 |
| Qwen3.8 27B | 54.4 |