New study: LLM social simulations need robustness audits before scientific claims

A new paper from researchers including Jinyi Ye and Emilio Ferrara cautions that LLM-based social simulations are dangerously brittle. In a repeated Prisoner's Dilemma experiment, altering persona format or game-instruction framing shifted cooperation rates by as much as 76 percentage points. Similarly, changes in network homophily and hub assignment produced consistent shifts in polarization metrics for an echo chamber simulation. These results demonstrate a 'butterfly effect' where small perturbations cascade into dramatically different macro-level outcomes.

To address this, the authors introduce TRAILS (Taxonomy for Robustness Audits In LLM Simulations), a framework that evaluates sensitivity at three levels: agent (micro), interaction (meso), and system (macro). Crucially, they found sensitivity unevenly distributed across model families—the same perturbation causing a 76 pp shift in one frontier model caused only a 1 pp shift in another. The paper argues that robustness must be measured per claim and per model, not assumed, and calls for such audits to become a first-order validation requirement before simulations inform policy or scientific understanding.