Benchmarks / Twin-2K-500

Twin-2K-500

Released
Digital twin fidelity

A large dataset for building and evaluating digital twins based on the responses of real people.

Updated 2025
2,000+ Participants
500+ Questions
19 Evaluation domains

Background

Digital twins aim to simulate the behavior and preferences of specific people. Twin-2K-500 supports large-scale study of when these simulations preserve genuine individual variation and when they collapse toward stereotypes or population averages.

What it evaluates

  • Individual-level response prediction
  • Relative heterogeneity across people
  • Fidelity across different subject domains
  • Bias introduced by richer persona descriptions

Dataset

The dataset contains responses from more than 2,000 people to more than 500 questions. Follow-on evaluation spans 19 domains.

Evaluation methodology

A model receives persona information and predicts held-out responses for the corresponding participant. Analysis compares aggregate accuracy with the more demanding goal of preserving individual-level behavior and variation.

Representative task

Given a participant’s available background and prior responses, predict how that same participant will answer a held-out question.

Results

Published findings show that digital twins can capture relative heterogeneity but still struggle with precise individual prediction and systematic persona-induced bias.