ns3-GenAI: Integrating Large Language Models with ns3 for AI-Native Network Simulations

Abstract

Existing ns3/AI bridges target numeric reinforcement learning (RL) pipelines and cannot handle the text-centric prompt/response exchange, structured output validation, and multi-node orchestration that large language models (LLMs) require. We introduce ns3-GenAI, an open framework that augments the ns3 shared-memory interface with three bridge-level features: a hybrid text-and-numeric memory layout, a fail-safe validation gate prior to writing back to the simulator, and multi-instance synchronization to support coordinated LLM workflows. These mechanisms provide a common, validated simulation path in which hosted and local LLMs, RL, and deep learning modules consume the same ns3 states and return decisions without changing simulator-side logic. We instantiate the framework in two complementary settings: a data-plane semantic codec and a control-plane self-organizing network (SON) agent. Across both case studies, hosted and local LLMs and conventional baselines occupy distinct quality–latency and throughput–fairness operating points. Rather than claiming a universally superior model class, these controlled comparisons demonstrate ns3-GenAI’s hybrid, fail-safe, and synchronized shared-memory path for evaluating generative and conventional network intelligence under identical simulation conditions.