Running an A/B test is the standard way to know whether a product change works, but it is expensive: each experiment needs real users, engineering resources, and weeks of measurement. A new preprint on arXiv proposes a simulation framework designed to predict those outcomes in advance. The authors describe data-driven, persona-conditioned agents that stand in for user segments, allowing teams to estimate how a change would perform without committing live traffic.
The framework is described as data-driven and persona-conditioned, meaning agent behavior is tied to defined user personas rather than generic synthetic traffic. The abstract does not provide implementation details or validation results, so the practical accuracy of the simulations remains an open question.
As a single preprint, there are no independent sources to compare against; the claims are the authors' own.