LLM Simulations Face Robustness Concerns

Researchers at an unspecified location present evidence that small changes to LLM social simulations drastically alter outcomes. They find that minor adjustments to agent personas and instructions in simulations of the Prisoner’s Dilemma shifted cooperation rates by as much as 76 percentage points. Similar perturbations in social media echo chamber simulations affected polarization metrics.

The team notes that sensitivity to these changes varies significantly between different LLM models. One model showed a 76 percentage point shift, while another experienced only a 1 percentage point change from the same adjustment. This suggests that results may reflect the specific implementation of the simulation, not necessarily the social mechanisms being studied.

To help address this, the researchers introduce TRAILS, a taxonomy for auditing robustness across agent, interaction, and system levels. They argue that validating robustness must be a primary requirement before using these simulations to explain social phenomena or guide decisions. Further work continues to refine audit methods and improve simulation reliability.