Root cause analysis (RCA) is a cornerstone of telecom network operations. With 5G and emerging 6G systems exhibiting intricate cross‑layer dependencies, traditional rule‑based or pure machine‑learning approaches struggle to pinpoint performance degradations quickly. Large language models (LLMs) bring powerful reasoning and knowledge‑integration capabilities, yet naïvely applying vanilla LLMs to telecom RCA often yields hallucinations, unstable reasoning chains, and poor alignment with concrete network evidence.
This paper first surveys the evolution of telecom RCA: from rule engines, through feature‑based ML models, to the latest LLM‑enabled techniques. It then outlines four emerging paradigms—structured reasoning, retrieval‑augmented grounding, agentic orchestration, and verifiable reasoning.
Building on these insights, we propose a structured reasoning framework tailored for LLM‑driven telecom RCA. The workflow consists of:
- Canonical context construction – mapping heterogeneous telemetry (KPIs, alarm graphs, topology snapshots) into a unified representation;
- Decision‑path reasoning – enforcing chain‑of‑thought steps where the LLM must explicitly cite the supporting evidence for each intermediate judgment;
- Evidence‑grounded explanation generation – producing fault diagnoses together with traceable evidence links.
Experiments on two public 5G RCA datasets, TeleLogs and TelecomTS, show that the proposed framework consistently outperforms vanilla LLMs, retrieval‑augmented LLMs, and traditional ML baselines, improving diagnostic accuracy by 7%‑12% and decision consistency by ~15%. The cross‑dataset gains highlight that structured reasoning is essential for practical LLM‑based RCA systems in next‑generation telecom networks.
Review: By structuring network telemetry and enforcing evidence‑driven chain‑of‑thought, the authors effectively curb LLM hallucinations and deliver a viable path toward intelligent, trustworthy telecom operations.