NetInspector: Measuring and Improving the Ability of LLMs to Reliably Generate Intent-Based Network Policies

NetInspector uses a three-layer agent framework to decouple network fact retrieval from reasoning, and verifies topology and group hierarchy with a real-time environment layer before approval, reducing the false negatives of an LLM checking whether intents violate security policies. The authors state that a fine-tuned LLM is good at intent translation but produces false negatives due to a lack of continuous grounding; on the 2,224-sample NetInspector-Bench, FNR is reduced by more than 30% compared with a no-grounding baseline.