EvoCause Uses LLMs to Refine Causal Graphs for Root Cause Analysis

A new research paper introduces EvoCause, a system designed to improve Root Cause Analysis (RCA) in complex IT and telecommunication systems by using large language models (LLMs) to refine causal graphs. Traditional RCA methods often rely on fixed causal graphs learned from observational data, which cannot adapt to expert insights from past incidents. EvoCause closes this gap by allowing expert labels to constrain which alarms should be source nodes. While experts define the constraints, an LLM proposes semantically plausible edits to the causal graph. Deterministic code then validates these edits for node identity and acyclicity, retaining the best graph based on a labeled alignment set. Crucially, once refined, the graph operates independently, making predictions without further LLM intervention. The researchers also released TeleRCA, a new expert-annotated benchmark. On both synthetic data and TeleRCA, EvoCause significantly outperforms unrefined baselines, demonstrating improved F1 scores and reduced structural Hamming distance.