PropLLM: Propagation-Aware Scene Reconstruction for Network Fault Diagnosis

PropLLM is a novel AI system that diagnoses network faults by tracing propagation paths backward from symptomatic alerts using large language models combined with knowledge graphs. The approach achieves 3.9% improvement in fault diagnosis accuracy and reduces hallucinations by 50.8% compared to existing methods, with validation across Wi-Fi and 5G networks.

PropLLM addresses a fundamental challenge in network operations: distinguishing between distinct root-cause faults that produce identical end-point symptoms. Traditional diagnostic systems—whether rule-based, machine learning, or LLM-based—process alerts in a single pass, inherently limiting their ability to resolve ambiguity when multiple failure scenarios converge at the same symptomatic endpoint. This research introduces a structural innovation by integrating hop-by-hop scene reconstruction with generative LLM reasoning.

The system's technical architecture combines two key mechanisms: a dual-layer knowledge graph that provides verifiable factual evidence at each step of the propagation chain, and a Temporal Causal Propagation Attention (TCPA) mechanism that encodes topological causal priors directly into the model's attention computation. This design ensures the LLM follows network physics rather than relying on statistical pattern matching, addressing a critical weakness in applying LLMs to infrastructure diagnostics where hallucinations can lead to costly misdiagnosis.

The performance improvements—3.9% accuracy gain in fault typing, 4.7% in root cause localization, and 50.8% hallucination reduction—demonstrate that integrating domain-specific causal structures with generative models produces measurable operational benefits. Validation across Wi-Fi and 5G datasets suggests the approach generalizes across network types.

For infrastructure operators, this represents progress toward automated fault diagnosis that reduces mean-time-to-resolution. For AI researchers, PropLLM exemplifies how constraining LLM reasoning with structured domain knowledge and causal priors can improve both accuracy and interpretability, a pattern likely to emerge across other critical infrastructure domains.

• →PropLLM traces network faults backward from alerts using LLMs plus knowledge graphs, improving diagnosis accuracy by 3.9% and reducing hallucinations by 50.8%.
• →The Temporal Causal Propagation Attention mechanism encodes network topology directly into attention computation to guide reasoning along correct causal paths.
• →Backward hop-by-hop reconstruction resolves the fundamental limitation of single-pass alert-to-diagnosis mapping when different root causes produce identical symptoms.
• →Validated performance across Wi-Fi and 5G networks indicates the approach generalizes across different infrastructure scenarios.
• →Structural integration of domain knowledge with generative AI reduces hallucination rates, critical for high-stakes infrastructure diagnostic applications.