User behavior simulation increasingly leverages large language models (LLMs) to support multi‑agent ecosystem simulations. Existing simulators rely on static user profiles inferred from historical observations, which are insufficient in live‑streaming contexts where interactions continuously reshape behavior. To address this, we introduce LiveSim, an LLM‑based framework for simulating live‑stream ecosystems. It treats users as editable behavioral hypotheses and refines them through trajectory‑grounded interactions. By comparing simulated trajectories with observed ones, discrepancies reveal missing environment‑shaping effects. These signals are extracted as transferable environment‑behavior patterns and stored in a collective behavioral memory, improving fidelity at the user level and enabling ecosystem‑wide simulation. Experiments on real live‑stream risk‑control data demonstrate LiveSim’s effectiveness in enhancing user‑level fidelity and facilitating ecosystem‑level analysis of risk evolution and platform interventions.
Blogger's Review: LiveSim’s iterative hypothesis refinement markedly improves simulation realism, offering actionable insights for platform risk management.