A research collaboration led by Harvard University and the Massachusetts Institute of Technology has released MatrAIx, an open-source infrastructure that instantiates 8.3 billion synthetic personas as LLM-driven agents. These agents run through reproducible tasks across four digital environments.

The system is built on Persona 8B, a dataset of 8.3 billion unique profiles. Each profile is defined across 1,290 categorical dimensions covering demographics, psychology, capabilities, behavior, and lifestyle.

Each persona is constructed using a directed acyclic graph that preserves conditional dependencies between attributes. This prevents logically incoherent combinations such as a rural Kenyan villager with a Harvard Ph.D. who speaks only Icelandic.

The dataset blends dependency-aware synthetic generation with evidence-aware grounding. Sources include United Nations population statistics, the General Social Survey, Wikipedia biographies, Amazon verified purchase reviews, and Stack Overflow developer surveys.

A quality-filtered coreset of approximately one million personas is available on Hugging Face for immediate research use.

What's new

  • Scale: 8.3 billion personas matching current global population
  • Schema: 1,290 categorical dimensions across five core categories
  • Environments: Survey, AI Chatbot, Web (sandboxed browsing), and OS-App (native Linux, macOS, iOS control via virtual input)
  • Evaluation scope: 1,010 tasks across 25+ domains; 18,189 large-scale simulation runs
  • Validation: 91.5% trait adherence across 400 controlled trials (10 attributes × 4 environments); human expert rating of 4.135/5 for grounded personas
  • Model backends tested: Claude Opus 4.8 (93.8% within 1 point of human expert scores), GPT-5.5 (79.2% within 1 minute)
  • Release artifacts: GitHub repository (MIT license), Persona 1M coreset on Hugging Face, Playground UI, task library, and arXiv preprint 2608.04205

The MatrAIx Playground provides both a graphical interface and a CLI for sampling cohorts, selecting tasks, pinning model backends, and executing jobs. Docker is required for Web and OS-App environments. Survey and Chat tasks run without containers. The repository includes smoke tests for all four environment types and a handbook documenting persona schema, task specification, and environment runtime.

Why it matters

Traditional user research and product testing are constrained by recruitment cost, time, and sample diversity. MatrAIx offers a complementary instrument: the ability to stress-test product features, pricing changes, interface redesigns, and AI systems against billions of demographically and psychologically varied simulated users overnight on a single server.

Product teams can surface edge-case failures — such as how a price-sensitive novice abandons a broken checkout flow or how a conservative user reacts to an AI hallucination — before committing to expensive real-world launches.

Social scientists and policymakers gain a tool for exploring downstream effects of new technologies across highly diverse populations at a scale previously impossible.

Our take

The 91.5% aggregate consistency figure is impressive, but the drop to 83% in OS-App manipulation tasks signals that GUI-level interaction remains the weakest link. Until agents reliably operate native applications across platforms, the system's most ambitious use cases will require careful human-in-the-loop validation. The open release of the 1M coreset and task library lets the community probe exactly where the simulation holds and where it fractures.

Sources