# CAIP: Detecting Router Misconfigurations with Context-Aware Iterative Prompting of LLMs

arXiv: https://arxiv.org/abs/2411.14283 · HTML: https://mlsanity.com/paper/2411.14283/

## Key facts

- Cited by 3 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 21 Nov 2024.

## Built on by (most cited first, 3 of 3)

- [NetConfArena: An Executable Benchmark for LLM Agents in Closed-Loop Network Configuration](https://mlsanity.com/paper/2608.23179/) (2026)
- [Towards Secure Cloud-Native Computing: Unveiling Kubernetes Misconfigurations with Large Language Models](https://mlsanity.com/paper/2609.20834/) (2026)
- [SoK: Semantic Decision Engines in Network Control Loops](https://mlsanity.com/paper/2610.06425/) (2026)

## Abstract (arXiv)

Model checkers and consistency checkers detect critical errors in router configurations, but these tools require significant manual effort to develop and maintain. LLM-based Q&A models have emerged as a promising alternative, allowing users to query partitions of configurations through prompts and receive answers based on learned patterns, thanks to transformer models pre-trained on vast datasets that provide generic configuration context for interpreting router configurations. Yet, current methods of partition-based prompting often do not provide enough network-specific context from the actual configurations to enable accurate inference. We introduce a Context-Aware Iterative Prompting (CAIP) framework that automates network-specific context extraction and optimizes LLM prompts for more precise router misconfiguration detection. CAIP addresses three challenges: (1) efficiently mining relevant context from complex configuration files, (2) accurately distinguishing between pre-defined and user-defined parameter values to prevent irrelevant context from being introduced, and (3) managing prompt context overload with iterative, guided interactions with the model. Our evaluations on synthetic and real-world configurations show that CAIP improves misconfiguration detection accuracy by more than 30% compared to partition-based LLM approaches, model checkers, and consistency checkers, uncovering over 20 previously undetected misconfigurations in real-world configurations.