Title: PropLLM: Enabling Accurate Network Fault Diagnosis Through Propagation-Aware Scene Reconstruction

Abstract:

While network faults cascade sequentially through topology and protocol dependencies, operational systems usually only detect symptomatic alerts at the final stages of these propagation chains. In these end-point scenarios, different root-cause issues often manifest as nearly identical symptoms. Current diagnostic methods—whether they rely on rules, machine learning (ML), or large language models (LLMs)—typically attempt to map an alert set to a diagnosis in a single step. This structural limitation renders them incapable of resolving the ambiguity inherent in end-point observations.

To address this, we introduce PropLLM, the first approach to combine the hop-by-hop scene reconstruction paradigm with the generative reasoning power of LLMs. Beginning with end-point alerts, PropLLM retraces the propagation path step-by-step. At each stage, it gathers verifiable factual evidence from a dual-layer knowledge graph (KG). Furthermore, our proposed Temporal Causal Propagation Attention (TCPA) mechanism embeds known topological causal priors directly into the attention computation. This guides the model along the correct causal trajectory, allowing it to localize the root cause and identify the fault type via a fully evidenced causal chain.

Evaluated on a real-world Wi-Fi multimodal fault dataset, PropLLM outperforms the strongest baseline by increasing fault type diagnosis accuracy by 3.9% and root cause localization accuracy by 4.7%, while simultaneously cutting the hallucination rate by 50.8%. Additional experiments conducted on the TeleLogs 5G dataset further validate the method’s effectiveness across diverse network environments.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC