Preprint reports faster LLM-serving simulation framework

Researchers introduced Simthesizer, a preprint framework in which a coding agent converts natural-language feature requests into changes to an LLM-serving simulator under guardrails and fidelity validation. The authors report that extensions built with the framework averaged 2.51% throughput error against a vLLM-based system, compared with 6.03% for extensions built with existing simulators. On identical workloads, they report simulations up to 284.96 times faster than LLMServingSim2.0 and 23.19 times faster than Vidur.