MatrAIx: Simulating the World with 8.3 Billion Persona Agents
Xiaomin Li, Yuexing Hao, Jianheng Hou, Jintao Huang, Qianfeng Wen, Shirley Huang, Yifan Liu, Xiaoyi Liu, Yilan Fan, Yijun Wang, Koutian Wu, Ruoqi Gao, Muhammad Ahmed Mohsin, Jing Tang, Brihi Joshi, Heming Liu, Zheyuan Deng, Zonglin Di, Sankalp Jajee, Jiuyao Lu, Zhiwei Zhang, Saksham Kapoor, Ishan Gupta, Yunhan Zhao, Chanwoo Park, Yucheng Lu, Bing Hu, Weihang Xiao, Aravind Mohan, Hanwen Xing, Runyu Zhang, Mihir Kulshreshtha, Yuanda Xu, Qianyu Zhu, Dianzhuo Wang, Yuxin Xiao, Bowen Jiang, Yongye Su, Wenhao Chai, Zuxin Liu, Lawrence Yunliang Chen, Xuandong Zhao, Ethan Ye, Shivam Patel, Jason Xie, Alex Martin Richmond, Weixiang Ding, Emre Okcular, Diya Mathew, Ziheng Wang, Rana M. Shahroz Khan, Zhejian Peng, Fang Wu, Fan Nie, Xinyang Han, Yubin Kim, Jiawei Zhang, Zhenting Qi, Huangyuan Su, Xu Pan, Abinitha Gourabathina, Hyewon Jeong, Hemanth Neelgund Ramesh, Kumail Alhamoud, Kimia Hamidieh, Zidi Xiong, Samuel Schmidgall, Pengrui Han, Yepeng Huang, Yongheng Wang, Bowen Yang, Alex Gu, Yuchu Wang, Akshay Paruchuri, Brenna Li, Hejie Cui, Jiayuan Ding, Chaosheng Dong, Jiahao Wang, Yixuan He, Chi Wang, Pamela Bhattacharya, Tianyi Peng, Paul Pu Liang, Mitchell Gordon, Yilun Du, Marinka Zitnik, James Zou, Prasanna Tambe, Philip Torr, Emily Fox, Asu Ozdaglar, Dawn Song
cs.AI
2026-08-05
MatrAIx builds 8.3B personas across 1,290 attributes and drives Opus 4.8, GPT 5.5, and Haiku 4.5 as simulated product users; personas behaved as specified in 91.5% of trials.
Evaluating an AI product with real users is expensive, slow, and hard to scale, and different users ask, interact, and judge in very different ways that a focus group cannot cover. Offline benchmarks are cheaper and reproducible, but their metrics are fixed in advance and abstract that human diversity away entirely.
MatrAIx fills the gap by having large language models play many differently configured personas that run your product through surveys, chat, websites, and apps, then report how each population reacts. The example the paper returns to is a soda six-pack: after a $2 price increase, do shoppers of different income and motivation levels still buy?
Three pieces.
First, Persona 8B: 8.3 billion persona records, each described over 1,290 categorical dimensions in five groups (background 238, psychology 210, capability 331, behavior 124, lifestyle 387). Records come from two routes.
The synthetic route uses DAG sampling. Each dimension is a node; a directed edge means one dimension is sampled conditional on another (developers are more likely to code). A node's probability is the product of a prior, a compatibility ratio, and a marginal-match factor, which keeps correlated attributes tied.
The human-grounded route maps six sources onto the same schema: Wikipedia biographies (323,438), Amazon reviews (97,915), Stack Overflow Developer Survey (113,120), PRISM Alignment (1,487), General Social Survey (63,532), and a volunteer survey (355). Unsupported dimensions stay null.
The released set is a 1-million coreset: 599,847 human-grounded plus 400,000 synthetic. The authors note that because most human records are single-source, they give marginal distributions, not joint coverage over all 1,290 dimensions. So the 8.3B figure is the combinatorial expansion on the synthetic side; after quality filtering, only 400,000 synthetic records reach the public release.
Second, the Playground: four environments. Survey runs questionnaires for concept testing and price sensitivity. AI Chatbot holds conversations with assistants or support bots and can inject latency to measure satisfaction and willingness to continue. Web drives browser automation and computer-use agents on sites. App runs native apps in Docker, Linux, macOS, and iOS.
Third, Applications: 1,010 task templates across more than 25 domains (Commerce 207, Software 156, Finance 161, Healthcare 168, other 318), split by environment as 621 Survey, 371 Chatbot, 12 Web, 6 App. Each task is an executable study: pick a persona cohort, run it, record outcomes, verify with a checker. The 1,010 is a template count, not all executed; the paper actually ran eight representative tasks.
18,189 trials across eight tasks (two per environment), with personas driven by three models: Claude Opus 4.8, GPT 5.5, and Claude Haiku 4.5. The system under test is fixed; only the model behind the persona varies.
| Validation | Result |
| Persona adherence (400-trial study, 10 attributes x 4 environments) | 366/400 = 91.5%; 33 of 40 cells reached 4/5 success; App weakest at 6/10 |
| Extraction quality (100 human-grounded, 6 human raters) | Mean 4.135/5; Opus 4.8 within 1 point in 93.8% of comparisons, GPT 5.5 in 79.2% |
| OpenBB trust grouping (all three models) | Cramér's V = 0.228-0.363, q < 1e-8; all three models order the four trust levels identically |
| Meal-planning chat (1,000 personas per model) | 7.1 turns per conversation on average; empty nesters 66% vs career changers 46% likely to follow the plan |
The OpenBB row is the key check of whether the system recovers a known association: persona backgrounds separate trust subgroups stably, and the three models agree. The meal case shows the other side, persona dimensions explaining downstream behavior differences.
For teams building AI products, this is a scalable synthetic-user testing facility: before paying for human panels, run price sensitivity, feature discoverability, latency tolerance, retention, and recommendation across hundreds to thousands of personas by segment. Problems that hit particular user groups surface earlier.
Be honest about positioning, though. This is an infrastructure paper, not a benchmark-topping model. Its validation is self-referential: it checks whether the persona agent behaves as declared (91.5%) and recovers one known association (trust), not whether simulated purchase intent equals real purchase intent. The right use is as a hypothesis generator that flags subgroup differences worth following up with human studies.
The authors state this plainly: human studies remain necessary before applying conclusions to real populations or consequential decisions, and simulated-user evaluation need not be a perfect model of human behavior to be useful. They also flag that results depend on which LLM drives the persona, so important findings should be checked across models and traced to the underlying interactions, and that single-source human records give marginal, not joint, coverage.
A few things stood out on reading. The App environment is the clear weak spot: adherence only reaches 6/10, and it holds just 6 of the 1,010 tasks, because driving real applications is far harder than answering a survey. The OpenBB validation, while a known association, shows mostly that the LLMs internalize the same stereotyped correlation, not necessarily the real-world distribution. There were also integrity exceptions (the meal-planning GPT 5.5 arm matched the cohort manifest on only 43 of 370 trials), so execution consistency at scale is not yet solid. The community framing of precisely simulating all human behavior deserves skepticism: the paper shows personas behave as specified 91.5% of the time, not that they reproduce all of humanity.