Let’s Synthesize Step-by-Step: Generating Network Configurations with Chain of Thought

Configuring networks is a highly-demanding and error-prone process. Recently, large language models (LLMs) have been used to synthesize network configurations. Even showing great potential in automatically configuring networks, we find that the accuracy of the generated configuration is still unsatisfactory, due to two reasons: (1) generating the entire configuration in a single-step is too hard, which overwhelms the model’s ability to reason over complex topologies and device roles; (2) LLMs lack sufficient domain knowledge and procedural understanding to handle structured tasks like protocol inference and parameter allocation. This paper proposes CoTNet, an LLM-based network synthesizer, which consists of two major steps: (1)CoTNet decomposes configuration generation into a two-stage process—first generating configuration templates, then assigning concrete parameters—to ensure correctness and consistency; (2)it further fine-tunes a lightweight LLM on a large-scale Chain-of-Thought (CoT) inference dataset, enabling the model to internalize domain-specific knowledge and structured reasoning patterns. Experimental results demonstrate that CoTNet achieves 94.5% device-level and 89% network-level accuracy in generating configurations for connectivity intents. Ablation studies further confirm that both CoT inference and two-stage task decomposition are critical to ensuring the correctness of configuration synthesis.