Decoded
← Feed
arXiv:2608.04205AI4 aug 2026

MatrAIx: Simulating the World with 8.3 Billion Persona Agents

1 min read2 citationsHN ×6PDF ↗
SURVEY96%WEB95%CHAT92%OS-APP83%0100% PERSONA ADHERENCE
fig. persona adherenceNative apps break persona adherence most often

A pool of 8.3 billion simulated people for testing apps and chatbots before real users see them.

91.5%

Trials where agents followed persona

Simulated users must act like their assigned traits to be useful.

How it works

  1. Real users are costly

    Human studies take weeks. Offline benchmarks check task success but ignore how different people use and judge a product.

  2. Simulate diverse users

    Sample personas from 1,290 linked attributes, pair each with an LLM, and run them through surveys, chatbots, websites and apps.

  3. Personas mostly hold

    Across 18,189 trials, agents kept assigned behaviors in 366 of 400 controlled tests; native apps were the weakest.

queryassigninteractartifactsfindingsPersona8B poolCohortsamplingLLM personaagentsFourenvironmentsTaskverifierSubgroupreportsqueryassigninteractartifactsfindingsPersona8B poolCohortsamplingLLM personaagentsFourenvironmentsTaskverifierSubgroupreports

What they found

Which comes out on top on Persona adherence?

How many personas per study?

try it

Drag the cohort size up and watch subgroup estimates stabilize.

Personas per taskSubgroup reliabilitynoisy subgroup gapsstable subgroup patterns~1,000 personas per model
Subgroup reliability
low
here
noisy subgroup gaps
Small cohorts give noisy subgroup gaps; returns flatten, which is why tasks used about 1,000 personas.Illustrative model, not the paper's data

Terms

6 terms

Persona A record of up to 1,290 categorical traits describing one simulated person.

Like a casting agency with billions of actor profiles: pick a cast matching your audience, have them try your product, and collect their reactions.

Where it breaks
Actors are people; these are LLMs that may flatten or stereotype real behavior.

Limits

  • No check that simulated reactions match real human behavior in these tasks.
  • Results depend on which LLM powers the persona; findings varied by model.
  • Coreset matches only four demographic marginals, not the full joint population.
From the paper