Loom Framework Improves LLM-Based Root Cause Analysis Efficiency

Deploying Natural Language Processing (NLP) systems in industrial settings often faces the challenge of consolidating diverse and sometimes contradictory textual hypotheses into a reliable consensus. While large, monolithic Language Model (LLM) agents offer extensive expressive power for tasks like Root Cause Analysis (RCA), they are often hampered by context limitations, compounding errors (hallucinations), and slow inference times. Conversely, traditional weak supervision methods, though statistically robust, are confined to discrete classifications. The Loom framework bridges these two approaches, offering a generative consensus mechanism for RCA. It functions by taking open-form hypotheses generated by modular heuristics—which are diagnostic templates populated with specific incident details—and mapping them into a continuous embedding space. Conflicts among these signals are then resolved using an iterative centroid-based reweighting algorithm. This process culminates in a single, lightweight LLM step that synthesizes the weighted consensus. Evaluations on the OpenRCA benchmark demonstrate that Loom achieves an optimal balance between accuracy and efficiency. It matches the performance of state-of-the-art autonomous agents on two datasets (Bank and Market-2) and performs competitively on others (Market-1 and Telecom), all while requiring only one LLM call per incident. This translates to a substantial speed improvement, being approximately 26 times faster (or 33 times faster with an 8B-parameter synthesizer) than comparable methods. The deployment experience also highlighted the benefits of deterministic consensus in building trust with Subject Matter Experts.