Li & Ji 2026

When simulations look right but causal effects go wrong: LLMs as behavioral simulators. Li, Y., and Ji, X. (2026). arXiv:2604.02458.

Relationship to this project — Closest occupied territory

Establishes at survey scale (3 model families, 11 interventions, 59,508 participants, 62 countries) that descriptive realism and treatment-effect accuracy are only weakly related, and that prompt refinements selected for realism can worsen effect estimates. This paper therefore does not claim the broad divergence as new; it contributes one concrete strategic-interaction mechanism — a fixed persona panel passing coarse marginal checks largely through between-prompt composition of corner-concentrated policies — under prospective registration.