PersonaMem v1 - 128K - Track Full Preference Evolution: leaderboard
PersonaMem v1 measures evolving user preferences across multi-session conversation histories using 128K-token context and discriminative multiple-choice response selection. This series measures Track Full Preference Evolution accuracy. Printed figure values are rounded to 0.01 and converted to percentages; not the 1M, RAG, or generative setting. Source: arXiv 2504.14225v1 Figure 3 and authors' results_qa_types.png.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O4 Mini | 73 | #185 |
| 2 | Gemini 2.0 Flash | 70 | #351 |
| 3 | GPT-4o | 68 | #357 |
| 4 | DeepSeek R1 | 68 | #262 |
| 5 | GPT-4.5 | 68 | #278 |
| 6 | Gemini 2.0 Flash Lite | 68 | #455 |
| 7 | GPT-4.1 | 67 | #258 |
| 8 | O1 | 67 | #190 |
| 9 | Llama 4 Maverick | 66 | #486 |
| 10 | Gemini 1.5 Flash | 65 | #620 |
| 11 | GPT-4o Mini | 60 | #598 |
| 12 | Claude 3.5 Haiku | 55 | #571 |
| 13 | O3 Mini | 54 | #282 |
| 14 | Claude 3.7 Sonnet | 45 | #259 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=personamem-v1-128k-track-full-preference-evolution · How It Works · Data refreshed daily, snapshot 2026-11-05.