Source-linked AI summary
Large Language Models (LLMs) for Telecom Root Cause Analysis (RCA): A Structured Reasoning Framework for Evidence-Grounded Diagnosis
Hao Zhou, Mandar Kulkarni, Hao Chen, Yan Xin, Charlie, Zhang
TL;DR
Telecom RCA is difficult in complex 5G and 6G networks because vanilla LLMs can reason unstably and fail to ground diagnoses in structured evidence. The paper proposes SEKA-FT, which canonicalizes inputs, supervises diagnostic paths, and grounds explanations; across TeleLogs and TelecomTS, it reports stronger accuracy and consistency than baseline approaches.
Problem
Complex cross-layer telecom networks require reliable RCA, but vanilla LLMs can produce hallucinated, unstable, and poorly evidence-aligned diagnoses.
Method
SEKA-FT organizes heterogeneous inputs into canonical contexts and jointly supervises evidence, staged diagnostic paths, root-cause decisions, and grounded explanations.
Results
Across TeleLogs and TelecomTS, SEKA-FT consistently outperforms ICL, non-LLM, and SFT-based baselines in diagnostic performance and decision consistency.
Takeaways & Limitations
The cross-dataset findings highlight structured reasoning design as important for reliable LLM-enabled RCA within the evaluated 5G RAN and observability scope.
Abstract
from arXiv · showhide
Root cause analysis (RCA) is a critical task in telecom network operations, but diagnosing performance degradations in modern 5G and emerging 6G networks remains challenging due to complex cross-layer dependencies. While large language models (LLMs) offer promising capabilities for reasoning and knowledge integration, directly applying vanilla LLMs to telecom RCA often leads to hallucination, unstable reasoning, and poor alignment with structured network evidence. This work first reviews the evolution of telecom RCA from rule-based and machine learning (ML) approaches to emerging LLM-enabled techniques, and provides an overview of recent paradigms, including structured reasoning, retrieval-augmented knowledge grounding, agentic orchestration, and verifiable reasoning. Building upon these insights, we propose a structured reasoning framework for LLM-enabled telecom RCA that aligns diagnostic reasoning with telecom-specific evidence and domain knowledge. The proposed approach first organizes heterogeneous network telemetry into canonical contexts, and then enforces decision-path reasoning during diagnosis, and finally generates evidence-grounded explanations for reliable fault identification. Experimental results on two 5G RCA datasets, TeleLogs and TelecomTS, demonstrate that the proposed framework consistently improves diagnostic accuracy and decision consistency compared with baseline techniques. These cross-dataset results highlight the importance of structured reasoning design for practical LLM-based RCA systems in next-generation telecom networks.
I. INTRODUCTION
Modern telecom networks are difficult to diagnose because heterogeneous cross-layer systems make failures operationally consequential, while conventional RCA lacks explicit diagnostic reasoning. The paper proposes SEKA-FT to align evidence, decision paths, and explanations for telecom RCA.
- Motivation: 5G-Advanced and emerging 6G networks integrate heterogeneous radio, transport, and core components, increasing operational difficulty and outage consequences.Reported incidents disrupted over 100 million devices and caused more than $100 million in loss.
- Prior approaches: Rule-based RCA uses predefined alarms and thresholds, whereas ML models detect and classify faults through statistical pattern recognition.These approaches process large telemetry volumes but can struggle with unseen failure modes.
- LLM challenge: Vanilla LLMs offer multi-step reasoning but may produce unstable diagnostic paths and fail to align heterogeneous structured evidence with telecom decisions.Relevant factors include mobility behavior, resource scheduling, and coverage geometry.
- Proposed framework: SEKA-FT reorganizes heterogeneous inputs into canonical contexts, stages diagnostic checks, and supervises explanations grounded in concrete indicators and domain knowledge.Its design follows an evidence-to-path-to-decision structure.
- Evaluation: Case studies on TeleLogs and TelecomTS report consistent improvements over ICL, non-LLM, and SFT-based baselines.The framework jointly encodes structured evidence, diagnostic paths, root-cause decisions, and explanations.
II. RCA: FROM RULE-BASED AND ML TO LLMS
Telecom RCA has progressed from deterministic rules to statistical ML and then to LLM-based evidence-grounded reasoning. Each transition improves flexibility or integration, while structured evidence alignment remains necessary for reliable decisions.
- Rule-based RCA: Rule-based systems provide transparent deterministic behavior but depend on predefined alarms, thresholds, and manually maintained troubleshooting logic.Static rules become difficult to scale and update as network scenarios diversify.
- ML-based RCA: ML methods improve scalability by learning patterns from telemetry, but usually map inputs to labels without explicitly showing how evidence supports candidate causes.Their performance also depends on training-data coverage and distribution stability.
- LLM-enabled RCA: LLMs can combine structured telemetry, textual descriptions, engineering knowledge, and operational instructions in a unified reasoning interface.They can generate diagnostic traces and support multi-step analysis across heterogeneous evidence sources.
- Reliability challenge: Without structured evidence alignment, LLMs may generate plausible but unsupported explanations and follow unstable reasoning paths.The central challenge is controlling how heterogeneous evidence connects to diagnostic decisions.
III. LLM-ENABLED TECHNIQUES FOR TELECOM ROOT CAUSE ANALYSIS
LLM-enabled telecom RCA techniques progressively strengthen diagnostic organization and grounding. CoT structures reasoning steps, while RAG grounds diagnoses in retrieved authoritative knowledge.
- Chain-of-Thought: CoT prompts LLMs to generate intermediate reasoning steps before reaching a final answer.For telecom RCA, staged checks can proceed from physical-layer indicators to MAC/RLC evidence and then RRC or handover logs.
- Chain-of-Thought: Structured CoT organizes RCA as an explicit diagnostic path instead of direct label prediction, helping evaluate cross-layer dependencies.The example checks throughput degradation across physical, MAC/RLC, and signalling evidence.
- Retrieval-Augmented Generation: RAG retrieves relevant information from an external authoritative knowledge base before generating a response.Telecom sources may include 3GPP specifications, vendor manuals, incident reports, and verified troubleshooting rules.
- Retrieval-Augmented Generation: Retrieved telecom evidence can constrain diagnoses and expose transport, core-network, or neighboring-cell dependencies beyond local KPIs.This grounding is intended to reduce hallucination during RCA.
C. Agentic Orchestration: Environment-interactive Planning
Agentic orchestration extends LLM-based RCA through autonomous tool use and multi-step hypothesis verification, while RLVR evaluates reasoning through network outcomes. These paradigms motivate SEKA-FT’s structured evidence and decision-path control.
- Agentic orchestration: Agentic LLM systems execute multi-step plans and interact with external systems through API calls.In telecom RCA, agents can coordinate specialized tools such as simulations or self-organizing network modules.
- Agentic orchestration: Agentic orchestration treats the LLM as a coordinator that verifies diagnostic hypotheses using specialized tools.The cited examples include ns-3 simulation and queries for antenna-tilt or power information.
- RLVR: RLVR evaluates proposed diagnostic paths and mitigation actions against checkable network outcomes such as throughput, packet loss, or handover failure rate.Successful trajectories can be reinforced and ineffective ones penalized.
- Paradigm progression: CoT, RAG, agentic orchestration, and RLVR form progressively stronger grounding and environmental coupling paradigms for telecom RCA.The progression moves from internal reasoning structure to outcome-based alignment.
- SEKA-FT: SEKA-FT focuses on structured evidence alignment and decision-path control as a practical step toward reliable LLM-based RCA.Its method is introduced as a structured fine-tuning framework for telecom RCA decision-path control.
A. Canonical Context Structuring for Stable RCA Fine-tuning
Canonical context structuring normalizes heterogeneous telecom inputs into consistent evidence blocks and semantic slots, helping stabilize diagnostic reasoning and standardized RCA reporting.
- A. Canonical Context Structuring for Stable RCA Fine-tuning: Canonical context structuring normalizes heterogeneous RCA inputs into consistent evidence blocks before LLM-based diagnostic reasoning.The inputs may include troubleshooting tickets, user-plane tables, and OSS logs.
- A. Canonical Context Structuring for Stable RCA Fine-tuning: The framework separates user-plane indicators, control-plane events, topology relationships, and configuration parameters into explicit blocks.Examples include Bottleneck Snapshot, UE Global State, and Trajectory-Level Events.
- A. Canonical Context Structuring for Stable RCA Fine-tuning: Consistent semantic slots keep serving-cell metrics, neighbor relationships, and mobility indicators in comparable structural positions across samples.This places Scheduled RBs and Serving SS-RSRP within a fixed diagnostic context rather than presenting them as isolated numeric tokens.
- A. Canonical Context Structuring for Stable RCA Fine-tuning: The method reduces redundant or unrelated telemetry while preserving signals that support root-cause inference.This creates a stable interface between heterogeneous telemetry and LLM-based reasoning.
- A. Canonical Context Structuring for Stable RCA Fine-tuning: Structured antenna and radio measurements can associate aggressive antenna configuration with edge-coverage loss rather than scheduling or mobility.The resulting behavior reduces output randomness and promotes standardized RCA reporting.
B. CoT-enabled Decision-Path Control for Reliable RCA
SEKA-FT replaces sparse label prediction with supervised diagnostic paths that extract interpretable checks, refine hypotheses using domain knowledge, and maintain evidence-grounded explanations.
- B. CoT-enabled Decision-Path Control for Reliable RCA: Final-label supervision alone can encourage shortcut correlations instead of teaching how to evaluate evidence before diagnosis.A categorical root-cause label does not specify the intended troubleshooting logic.
- B. CoT-enabled Decision-Path Control for Reliable RCA: CoT-enabled decision-path control makes the model generate intermediate diagnostic assessments before producing a root-cause label.The path is organized from complex evidence to structured checks and then to a final decision.
- B. CoT-enabled Decision-Path Control for Reliable RCA: Diagnostic checks convert high-dimensional telecom measurements into compact, verifiable abstractions and progressively prune unsupported hypotheses.Examples include mobility speed, scheduled resources, handover counts, and UE-to-serving-cell distance.
- B. CoT-enabled Decision-Path Control for Reliable RCA: Evidence- and knowledge-anchored explanations supervise both the final answer and the evidence path leading to it.The design uses a two-step evidence-based explanation structure.
- B. CoT-enabled Decision-Path Control for Reliable RCA: Step 1 extracts interpretable checks, while Step 2 refines hypotheses using the outcomes of those checks and domain knowledge.In the illustrative case, false checks eliminate several candidate causes before final reasoning.
- B. CoT-enabled Decision-Path Control for Reliable RCA: Anchoring explanations to observable indicators reduces hallucinated reasoning and improves RCA traceability.The framework explicitly connects evidence extraction with knowledge-conditioned decision refinement.
D. Integrated SEKA-FT Framework and Alignment Objective
The integrated SEKA-FT framework aligns raw telecom evidence, diagnostic paths, and explanations through three stages, then jointly trains them with a causal language-modeling objective.
- D. Integrated SEKA-FT Framework and Alignment Objective: Stage 1 normalizes heterogeneous telemetry into canonical context blocks that provide a consistent evidence interface.The blocks include bottleneck snapshots, UE global states, trajectory-level events, and engineering parameters.
- D. Integrated SEKA-FT Framework and Alignment Objective: Stage 2 applies CoT-based decision-path control instead of directly mapping structured evidence to a root-cause label.This addresses unstable or shortcut-based reasoning from sparse supervision.
- D. Integrated SEKA-FT Framework and Alignment Objective: Stage 3 expands supervision from one categorical label to check-based evidence abstraction and knowledge-conditioned hypothesis refinement.The structured target is intended to support both label accuracy and traceable diagnostic conclusions.
- D. Integrated SEKA-FT Framework and Alignment Objective: SEKA-FT optimizes token-level causal language modeling over a structured target with prompt tokens masked out.The target couples the diagnostic path, evidence-grounded refinement, and final RCA decision.
V. CASE STUDY: PERFORMANCE EVALUATION IN 5G THROUGHPUT DEGRADATION
SEKA-FT is evaluated on TeleLogs and TelecomTS against sequence-classification and supervised fine-tuning baselines, with repeated runs, confidence intervals, and paired significance tests.
- V. CASE STUDY: PERFORMANCE EVALUATION IN 5G THROUGHPUT DEGRADATION: SEKA-FT is evaluated on two 5G RCA datasets: TeleLogs and TelecomTS.TeleLogs contains structured KPIs, mobility statistics, engineering parameters, and root-cause labels; TelecomTS contains high-resolution testbed KPI records.
- V. CASE STUDY: PERFORMANCE EVALUATION IN 5G THROUGHPUT DEGRADATION: The structured supervision target combines the original RCA label, benchmark-defined checks, automatically generated explanations, and the intermediate diagnostic path.The primary base model is Qwen2.5-1.5B-Instruct.
- V. CASE STUDY: PERFORMANCE EVALUATION IN 5G THROUGHPUT DEGRADATION: Baselines include LSTM, Vanilla SFT, SFT + Structured Input, and SFT + Explanation, alongside the full SEKA-FT model.Additional model-scaling evaluations use Qwen2.5-7B-Instruct and Qwen3-32B.
- V. CASE STUDY: PERFORMANCE EVALUATION IN 5G THROUGHPUT DEGRADATION: All reported results are averaged over ten independent random-seed runs with 95% confidence intervals.The intervals use Student’s t-intervals based on the sample mean and standard deviation.
- V. CASE STUDY: PERFORMANCE EVALUATION IN 5G THROUGHPUT DEGRADATION: Paired tests compare predictions on identical held-out samples using exact McNemar’s test for Accuracy and a paired permutation test for Macro-F1.These tests evaluate paired differences across the reported baselines.
B. Experiment Results on TeleLogs
On TeleLogs, SEKA-FT delivers the strongest accuracy, Macro-F1, convergence, and decision-path consistency among evaluated approaches. Its residual errors mainly involve ambiguous coverage and neighbor-cell evidence, while sequential checks reduce the effective diagnosis space.
- SEKA-FT achieves 0.942 ± 0.006 Accuracy and 0.937 ± 0.007 Macro-F1, outperforming Vanilla SFT, structured-input SFT, and the LSTM baseline.
- SEKA-FT improves rapidly and maintains consistently high Accuracy and Macro-F1, whereas other fine-tuning strategies remain near random-level or saturate lower.The comparison attributes this learning behavior to structured supervision of final labels and intermediate diagnostic logic.
- Qwen3-32B reaches only 0.126 ± 0.010 Accuracy and 0.109 ± 0.009 Macro-F1 in regular ICL, while lightweight 1.5B SEKA-FT substantially outperforms all ICL baselines.
- SEKA-FT achieves near-perfect consistency across overshooting-coverage, frequent-handover, high-mobility-speed, and insufficient-resource-block checks.These checks use deterministic evidence abstractions and domain-rule-based hypothesis refinement targets.
- Most residual errors occur among coverage- and neighbor-cell-related categories because excessive downtilt, stronger neighbor selection, and overlapping coverage can produce similar RSRP/SINR patterns.
- Sequential checks reduce the average candidate-set size from 8 to 6.25, 4.75, 3.5, and 2.5 after four checks under a uniform root-cause distribution.A positive check identifies its corresponding root cause, while a negative result prunes that hypothesis.
C. Experiment Results on TelecomTS
On TelecomTS, SEKA-FT preserves the TeleLogs ranking pattern and achieves the best performance, with statistically significant gains over the strongest competing baseline. The results support generalization beyond the TeleLogs-specific input format within the evaluated settings.
- SEKA-FT achieves 0.647 ± 0.004 Accuracy and 0.615 ± 0.005 Macro-F1 on TelecomTS, exceeding all reported baselines.
- Compared with SFT + Structured Input, SEKA-FT improves Accuracy by 0.515 and Macro-F1 by 0.477.
- Paired tests find SEKA-FT significantly outperforms SFT + Structured Input on both Accuracy and Macro-F1 across TeleLogs and TelecomTS, with p < 0.001 for both metrics.
- Across both datasets, raw SFT and ICL are ineffective, LSTM provides limited improvement, structured input helps but remains insufficient, and SEKA-FT ranks strongest.The authors define this as cross-dataset robustness within the evaluated 5G RAN/observability scope.
VI. CONCLUSION
The paper proposes structured reasoning for LLM-enabled telecom RCA by aligning diagnosis with telecom-specific evidence and decision paths. Experiments report improved diagnostic accuracy and decision consistency over conventional SFT and ICL baselines.
- The proposed framework aligns diagnostic reasoning with telecom-specific evidence and decision paths.
- Experimental results demonstrate significantly improved diagnostic accuracy and decision consistency compared with conventional SFT and ICL baselines.
- Future evaluation will extend the framework to core- and transport-network faults, previously unseen failures, and real operator incidents.