Why Most Crisis AI Simulations Are Wrong About Human Timing — And How a New Model Fixes It
LLM agents finally mimic human bursty activity patterns during disasters
A new paper on arXiv tackles a blind spot in LLM-based social simulation: temporal realism. While existing simulators can generate plausible individual actions, they fail to reproduce the bursty, heavy-tailed timing seen in real human crises—short intense activity followed by long quiet intervals. Researchers analyzed a city-scale log of offline volunteering in Shenzhen during COVID-19 and confirmed that this burstiness is endogenous and self-exciting, not driven by daily cycles. Yet a standard LLM-only simulator produced near-regular activity (median burstiness -0.14), missing the self-excitation channel entirely.
To fix this, the team built a simulator with a data-calibrated self-excitation gate and a crisis-period regime that decides when each agent acts. The LLM is queried only at those moments to determine which task to join and whether to commit. This decoupling lifted per-agent timing above the burst threshold (median burstiness 0.37) without degrading content decisions. The results suggest that achieving realistic crisis simulations requires explicit modeling of when humans act, separate from what they choose to do.
- Standard LLM simulations produce near-regular timing (burstiness -0.14) instead of real-world bursty patterns
- Using a multi-year Shenzhen COVID-19 volunteer dataset, researchers found timing is self-exciting and endogenous
- A data-calibrated self-excitation gate lifted burstiness to 0.37 without affecting LLM task decisions
Why It Matters
Enables accurate crisis simulations that predict not just actions but when people act.