PaperlayerView exact revision on arXiv

Source-linked AI summary

Reasoning Language Models for Root Cause Analysis in 5G Wireless Networks

Mohamed Sana, Nicola Piovesan, Antonio De Domenico, Yibin Kang, Haozhe Zhang, Merouane Debbah, Fadhel Ayed

arXiv:2507.21974v1cs.AIcs.NI

TL;DR

Mobile-network RCA requires explanations grounded in domain knowledge and causal reasoning, yet open-source reasoning LLMs struggle with troubleshooting tasks. The paper introduces TeleLogs and trains reasoning LLMs with supervised fine-tuning followed by reinforcement learning; its approach outperforms instruction-tuned and state-of-the-art reasoning models, with over 95% accuracy using Qwen2.5-32B-Instruct on the TeleLogs test set.

  • Problem

    Mobile-network RCA requires interpretable, domain-informed causal explanations, while manually encoded rules depend on expertise and are difficult to scale.

    Source: §I · p. 1

  • Method

    The paper introduces TeleLogs and trains reasoning LLMs through supervised fine-tuning with structured traces followed by reinforcement learning.

    Source: §I · pp. 1–2

  • Results

    Over 95% accuracy with Qwen2.5-32B-Instruct on the TeleLogs test set; the approach outperforms instruction-tuned and state-of-the-art reasoning LLMs and generalizes to randomized variants.

    Source: §VII · p. 7

  • Takeaways & Limitations

    The results suggest that properly trained reasoning LLMs can serve as explainable diagnostic tools for complex systems.

    Source: §VII · p. 7

Abstract

from arXiv · show

Root Cause Analysis (RCA) in mobile networks remains a challenging task due to the need for interpretability, domain expertise, and causal reasoning. In this work, we propose a lightweight framework that leverages Large Language Models (LLMs) for RCA. To do so, we introduce TeleLogs, a curated dataset of annotated troubleshooting problems designed to benchmark RCA capabilities. Our evaluation reveals that existing open-source reasoning LLMs struggle with these problems, underscoring the need for domain-specific adaptation. To address this issue, we propose a two-stage training methodology that combines supervised fine-tuning with reinforcement learning to improve the accuracy and reasoning quality of LLMs. The proposed approach fine-tunes a series of RCA models to integrate domain knowledge and generate structured, multi-step diagnostic explanations, improving both interpretability and effectiveness. Extensive experiments across multiple LLM sizes show significant performance gains over state-of-the-art reasoning and non-reasoning models, including strong generalization to randomized test variants. These results demonstrate the promise of domain-adapted, reasoning-enhanced LLMs for practical and explainable RCA in network operation and management.

I. INTRODUCTION

Mobile-network RCA must explain underlying fault causes, but manually encoded expertise is difficult to scale and LLMs can lack the rigor required for decisions. The paper introduces TeleLogs and a two-stage training approach, addressing gaps in existing RCA methods and LLM-based explanations.

I. INTRODUCTION

  • Manual rule and heuristic encoding depends heavily on domain expertise and becomes prohibitive as mobile networks grow in scale, complexity, and heterogeneity.Other machine-learning approaches also face scalability, interpretability, and generalization limits, especially with multiple symptoms or high-dimensional data.Source: §I · p. 1
  • TeleLogs is a curated, expert-annotated dataset of network troubleshooting scenarios with training and testing data, publicly released to support RCA research.It provides the data resource used to benchmark troubleshooting capabilities.Source: §I · p. 1
  • Open-source reasoning LLMs struggle with TeleLogs troubleshooting tasks, while the proposed supervised fine-tuning and reinforcement-learning stages target diagnostic performance and reasoning quality.The supervised stage uses an LLM-driven multi-agent pipeline to generate structured chain-of-thought traces; the second applies GRPO.Source: §I · pp. 1–2

II. RELATED WORK

  • Graph-based RCA models dependencies, but cited approaches can remain limited to classification or a single base station and may lack comprehensive, human-readable rationales [5],.Fault trees also depend on expert-defined rules, generalize poorly to new scenarios, and struggle with complex causal relationships [4].Source: §II · p. 2
  • LLM-based RCA systems include alarm-driven agents [7], tool-using diagnostic systems, self-consistency-based RCAgent, and SOP-guided Flow-of-Action.The cited approaches use iterative code-based repair, external tools, multiple reasoning trajectories, or stepwise summaries, respectively.Source: §II · p. 2
  • The related-work gap is integrating rule-based explainability with LLM reasoning, whose natural-language outputs may lack formal interpretability.The paper positions its approach as combining domain-specific rules with structured, multi-step diagnostic explanations.Source: §I · p. 1, §II · p. 2

III. PROBLEM OVERVIEW

The paper formulates mobile-network RCA as identifying a likely cause from network settings, drive-test observations, and symptoms, while also producing a structured explanation. Its formulation focuses on a single throughput symptom and treats the cause-to-symptom mapping as noisy and potentially nonlinear.

A. Problem formulation

  • RCA has two objectives: identify the most probable cause of an observed symptom from predefined candidates and generate a structured, step-by-step supporting explanation.The input state combines network engineering parameters, drive-test observations, inferred symptoms, and possible root causes.Source: A. Problem formulation · p. 2
  • The network configuration is represented by cell-level parameters, while observations are categorical or numerical time series collected through drive testing.Parameters include cell ID, transmit and receive antenna counts, and beamforming scenario; observations represent user-plane measurements.Source: A. Problem formulation · p. 2
  • The formulation focuses on one symptom: user downlink throughput falling below a specified threshold at time t.The paper describes this as a throughput measurement inferred from the drive-test data.Source: A. Problem formulation · pp. 2–3
  • Potential root causes are latent factors that may explain observed symptoms, including misconfigured network parameters.The formulation represents symptom-to-cause relationships as a mapping from configurations, observations, and symptoms to a candidate cause.Source: A. Problem formulation · p. 3
  • The cause-to-symptom function is generally nonlinear and partially unknown, and its observations may include noise such as measurement errors.The function must therefore be inferred from data.Source: A. Problem formulation · p. 3

B. Root Cause Inference

The paper frames root-cause identification as probabilistic inference from network settings, observations, and symptoms. Because estimating this posterior is difficult in high-dimensional network data, it proposes training reasoning LLMs and develops TeleLogs as a domain-grounded dataset.

B. Root Cause Inference

  • RCA selects a root cause using network engineering parameters, observations, and the observed symptom, formulated as an inverse inference problem.The target is the cause with the greatest posterior probability given those inputs.Source: B. Root Cause Inference · p. 3
  • Estimating the root-cause posterior is difficult because real network data has high dimensionality and intricate dependencies [4].Source: B. Root Cause Inference · p. 3
  • The proposed remedy trains reasoning LLMs to give accurate troubleshooting and coherent, step-by-step explanations grounded in learned patterns and domain knowledge.The authors state that this capability requires high-quality data grounded in domain and system knowledge, motivating TeleLogs.Source: B. Root Cause Inference · p. 3

IV. THE TELELOGS TROUBLESHOOTING DATASET

TeleLogs is a synthetic drive-test dataset built around real network engineering parameters, with ground-truth annotations and controlled simulation variability. It captures network configuration and user-plane performance for RCA scenarios centered on low downlink throughput.

IV. THE TELELOGS TROUBLESHOOTING DATASET

  • TeleLogs simulates drive testing in a 5G network using real network engineering parameters, with full visibility into configuration and user-plane performance.Ground-truth annotations and controlled variability support systematic RCA evaluation under realistic 5G conditions; training and testing data are publicly released.Source: §IV · p. 3

A. Network engineering parameters

  • Network engineering parameters cover each base station’s deployment, antenna configuration, and beamforming scenario.Table I summarizes parameters associated with each base station, including location, height, cell ID, azimuth, downtilt, and transmit/receive mode.Source: A. Network engineering parameters · p. 3, Table I · p. 3

B. User plane data

  • User-plane observations include timestamped performance indicators and mobility context collected during the drive test.These include serving-cell downlink throughput, RSRP and SINR, neighboring-cell RSRP, and GPS speed.Source: B. User plane data · p. 3, Table II · p. 4

C. Observed symptom

  • Each diagnostic scenario centers on downlink throughput falling below 600 Mbps, with causes including mobility, handover, antenna-alignment, and deployment issues.The target symptom is a significant degradation in downlink throughput.Source: C. Observed symptom · p. 4

D. Root Causes

TeleLogs defines eight candidate root causes spanning mobility and handover, coverage and interference, and serving-cell resource allocation.

D. Root Causes

  • Mobility-related candidates include vehicle speed above 40 km/h, frequent handovers, and misconfigured handover thresholds.The listed effects include increased handover frequency and degraded user performance.Source: D. Root Causes · p. 4
  • Coverage and interference candidates include excessive downtilt or coverage distance, non-colocated co-frequency cells, and overlapping reference signals from matching PCI mod 30.The passages associate these conditions with weak coverage, poor RSRP, or interference.Source: D. Root Causes · p. 4
  • Insufficient serving-cell PRBs to reach the target throughput is another candidate root cause..Source: D. Root Causes · p. 4

V. TRAINING REASONING MODELS FOR NETWORK TROUBLESHOOTING

The training setup prompts an LLM to diagnose a network symptom from configuration and observation data, then explain its hypothesized cause. It combines supervised fine-tuning with reinforcement learning and a binary correctness reward.

V. TRAINING REASONING MODELS FOR NETWORK TROUBLESHOOTING

  • Each training sample pairs a structured prompt built from C, U, Yt, and st with its ground-truth root cause.The prompt template is designed for this RCA task.Source: §V · p. 4, Figure 2 · p. 4
  • Given a prompt, the LLM generates a reasoning trajectory that identifies and explains a hypothesized cause, which is parsed to obtain the predicted root cause.The reward indicates whether that predicted cause matches the ground truth.Source: §V · p. 4
  • The training objective includes a regularization term, typically the KL divergence between the learned policy and a reference policy.The regularization term is weighted by β.Source: §V · p. 5
  • Training first applies supervised fine-tuning to a base policy and then reinforcement learning to produce the final policy.The policies progress from πθ0 to πθ1 after SFT and πθ2 after RL.Source: §V · p. 5

A. Supervised Fine-Tuning

Supervised fine-tuning uses high-quality reasoning traces generated by multiple agents to teach RCA models structured diagnostic responses. An aggregator selects, condenses, and filters trajectories before they become SFT data.

A. Supervised Fine-Tuning

  • A multi-agent data-generation pipeline produces diverse training samples with explicit reasoning traces intended to provide correctness and reasoning depth.Agents independently analyze each prompt using tailored prompting strategies.Source: A. Supervised Fine-Tuning · p. 5
  • Agents use elimination-based prompting to rule out implausible causes or contradiction-based prompting to test candidate causes against observed data.These distinct strategies are intended to produce diverse diagnostic trajectories.Source: A. Supervised Fine-Tuning · p. 5
  • The aggregator selects a trajectory by majority vote, then reformulates it into a compact structured explanation when its predicted cause matches the ground truth.The resulting trace has significantly fewer tokens than the selected trajectory.Source: A. Supervised Fine-Tuning · p. 6
  • Only correct trajectories are retained as SFT data, and the model is trained on them with token-level cross-entropy loss.The retained traces are the aggregator’s reformulated explanations.Source: A. Supervised Fine-Tuning · p. 6

B. RL Fine-Tuning

The RL stage uses GRPO to update the SFT policy, with a clipped objective and advantages estimated relative to groups of sampled responses.

B. RL Fine-Tuning

  • GRPO replaces the first objective term with a clipped update using the ratio between current and old policy probabilities, token advantages, and a clipping constant.The clipping constant limits excessively large policy updates.Source: B. RL Fine-Tuning · p. 6
  • For each question, GRPO estimates advantages from the rewards of a group of sampled trajectories.The group contains N trajectories.Source: B. RL Fine-Tuning · p. 6
  • All tokens receive the trajectory-level reward, while each token’s effective reward also includes its contribution to the KL loss.The KL term regularizes the loss against a reference policy.Source: B. RL Fine-Tuning · p. 6
  • Figure 6 compares the performance of different training methodologies.Source: Figure 6 · p. 6

VI. RESULTS

The evaluation tests three Qwen model scales on TeleLogs using pass@1 and maj@4. Across those scales, the proposed SFT+RL method outperforms the listed training baselines.

VI. RESULTS

  • The evaluation uses Qwen2.5-1.5B-instruct, Qwen2.5-7B-instruct, and Qwen2.5-32B-instruct, fine-tuned for 10 epochs and 1500 training steps.Training uses batch size 128; the SFT data pipeline uses two reasoning agents and samples eight trajectories per question.Source: §VI · p. 6
  • pass@1 measures single-attempt correctness, while maj@4 checks whether majority voting across four answers produces a correct answer.Source: §VI · p. 6

A. Performance of different training methodologies

B. Performance comparison: Fine-tuned Small models outperform SoTA LLMs

Fine-tuned models outperform the cited base and reasoning LLMs on TeleLogs, including at small scale, and retain strong performance when superficial test cues are randomized. The conclusion frames the results as evidence for explainable RCA tools while identifying real-world data and multiple-root-cause scenarios as future work.

B. Performance comparison: Fine-tuned Small models outperform SoTA LLMs

C. Generalization to the Randomized Test Dataset

VII. CONCLUSION

  • The framework combines supervised fine-tuning and reinforcement learning with TeleLogs, a synthetic dataset of realistic 5G troubleshooting scenarios, to improve RCA accuracy and interpretability.The conclusion identifies multiple root-cause scenarios and real-world operational data as future work.Source: §VII · p. 7

APPENDIX

The appendix presents an example RCA problem and trace from TeleLogs. The author statement describes the trace structure, the task inputs, and dataset availability and privacy measures.

APPENDIX

  • The section presents an example of root-cause analysis (RCA) on a problem from TeleLogs, with the example shown in Figure 7 and the problem in Figure 8.Source: APPENDIX · p. 7

AUTHOR STATEMENT

  • TeleLogs is released under the MIT Licence, with data anonymized and aggregated for privacy and confidentiality.Source: AUTHOR STATEMENT · p. 7
  • The RCA trace proceeds through data interpretation, elimination of unlikely causes, and a justified final root-cause conclusion.Its analysis uses network KPIs, configuration parameters, and RCA criteria.Source: Figure 7 · p. 8
  • The example problem lists candidate causes C1–C8, domain rules about downtilt and beam scenarios, drive-test logs, and engineering parameters.The description states that the model receives the context needed to reason.Source: Figure 8 · p. 9

Related papers

Loading 2507.21974v1…