Root Cause Analysis Method Based on Large Language Models with Residual Connection Structures

Researchers specializing in cloud computing systems published a paper on the arXiv preprint server on February 14, 2025, introducing RC-LLM, a novel root cause analysis method based on Large Language Models with residual connection structures designed to improve fault localization within complex microservice architectures. The proposal addresses the growing difficulty of identifying the origins of system failures in large-scale environments where intricate dependencies and high-dimensional telemetry data often overwhelm traditional diagnostic tools. By leveraging advanced AI architectures, the researchers aim to automate the detection of system anomalies that are otherwise obscured by the massive volume of metrics, logs, and traces generated by modern software applications. The core of the RC-LLM framework is its unique residual-like hierarchical fusion mechanism. This structure is inspired by residual connections found in deep neural networks, which allow information to bypass certain layers to prevent data loss or gradient degradation. In the context of root cause analysis (RCA), this architectural choice enables the model to effectively integrate multi-modal data—such as performance metrics and textual logs—while maintaining the integrity of the failure signals as they propagate through the model's analytical layers. This approach significantly mitigates the 'noise' typically associated with complex fault propagation between interconnected microservices. Traditional RCA methods frequently struggle with the 'dimensionality curse,' where the sheer variety and scale of data points make it nearly impossible to pinpoint a single point of failure in real-time. By utilizing Large Language Models (LLMs), the RC-LLM method can interpret semantic information within system logs and correlate it with numerical metrics in a way that previous heuristic or statistical models could not. This advancement represents a significant shift toward 'AIOps' (Artificial Intelligence for IT Operations), promising to reduce system downtime and operational costs for technology companies managing global-scale digital infrastructure.