AI 評測
MatrAIx Synthesizes 8.3 Billion Personas Using 1,290-Dimensional Attributes, Open-Sources Million-Record Core Dataset for AI Product Testing
MatrAIx combines correlated demographic attributes to create 8.3 billion simulated users and lets persona agents interact with surveys, chatbots, websites, and applications. The team has released a core dataset of roughly one million records and its evaluation code, but current validation shows only that the agents follow assigned personas—not that they can replace testing with real people.
The newly released [MatrAIx](https://arxiv.org/abs/2608.04205) attempts to expand AI evaluation beyond fixed question-and-answer tests to examine “how different users interact with products.” Its Persona 8B describes demographic, device, occupation, preference, and behavioral contexts using 1,290 categorical attributes. Some records are derived from publicly available, human-authored profiles, while the rest are sampled using a graphical model that preserves dependencies between attributes. The stated 8.3 billion figure refers to the total number of persona records that can be generated—not 8.3 billion real people or records collected individually by humans.
The system also provides four interactive environments—Survey, AI Chatbot, Web, and App—along with 1,010 application tasks spanning more than 25 domains, including business, software, finance, and healthcare. The study used Claude Opus 4.8, GPT-5.5, and Claude Haiku 4.5 to run eight representative tasks, producing 18,189 evaluation trajectories. Across 400 controlled tests, persona agents correctly exhibited or suppressed specified behaviors in 366 cases over ten behavioral attributes, yielding a 91.5% adherence rate. The [open-source repository](https://github.com/MatrAIx-ai/MatrAIx-Persona-8B) includes evaluation infrastructure, task examples, and MIT-licensed code. The team has also published a quality-filtered core dataset of roughly one million records on [Hugging Face](https://huggingface.co/datasets/MatrAIx2026/MatrAIx2026), comprising 599,847 human-grounded records and 400,000 synthetic records.
The technical value lies in enabling large-scale, conditional regression testing of the same chatbot agent or frontend workflow to identify failures that emerge only for particular backgrounds or device configurations. Persona adherence, however, is not equivalent to accurate prediction of human behavior: the primary validation sample included only 400 tests, some quality assessments still relied on LLM judges, and the models may amplify stereotypes into supposed “group responses.” Engineering teams should treat MatrAIx as a coverage tool and calibrate high-risk conclusions through research involving real people.