Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

TN-AutoRCA: Benchmark Construction and Agentic Framework for Self-Improving Alarm-Based Root Cause Analysis in Telecommunication Networks

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An agentic repair loop lifts alarm root-cause F1 from 59% to 92% on a new 530-scenario telecom benchmark.

desk verdict TN-RCA530 is a genuinely useful new benchmark, but the 91.79% F1 from Auto-RCA is not credible as a measure of general RCA skill until the authors rule out graph-traversal shortcuts and stop optimizing on the test set. read the letter →

arxiv 2507.18190 v2 pith:CXYTYBAX submitted 2025-07-24 cs.CL

Keyu Wu , Qianjin Yu , Manlin Mei , Ruiting Liu , Jun Wang , Kailai Zhang , Yelun Bao This is my paper
classification cs.CL
keywords rootcauseanalysistelecommunicationnetworksalarmcorrelationknowledgegraphLLMagentbenchmarkconstructioniterativeprogramrepairF1score
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Telecom alarm root-cause analysis is a graph-based, multi-label diagnostic task, and the field has lacked a realistic public benchmark for measuring AI systems on it. This paper tries to fill that gap with TN-RCA530, a set of 530 fault scenarios constructed backward from expert-verified root causes into knowledge graphs of alarms, equipment, and candidate causes. It also argues that an agentic framework, Auto-RCA, can far outperform direct LLM prompting on this benchmark: the framework iteratively evaluates a code solution, aggregates all failure cases, and feeds contrastive feedback to an LLM so it repairs systemic logic flaws instead of isolated errors. The authors report that direct Gemini-2.5-Pro scores 58.99% macro-averaged F1, while the best Auto-RCA-refined solution reaches 91.79%. The paper concludes that verifiable, self-improving agentic loops are a practical route to mastering complex domain-specific reasoning tasks.

What carries the argument

The load-bearing machinery is the 'cycle consistency check' used to build and grade TN-RCA530, plus the Auto-RCA 'evaluate-analyze-repair' loop. The cycle consistency check constructs each scenario backward from a known root cause to the alarms it triggers, then runs a forward inference pass from those alarms back to candidate causes; a one-to-one cycle marks a Simple scenario and a one-to-many cycle marks a Difficult one, automatically splitting the 530 cases into 29 Simple and 501 Difficult scenarios. The Auto-RCA loop then runs a code solution against the whole benchmark, clusters incorrect outputs into Bad Cases, and uses the dominant failure pattern as contrastive feedback to prompt the LLM to repair the code's logic, with the Orchestrator rejecting any round whose F1 does not improve. This turns the benchmark itself into a training signal: the LLM improves the solver instead of being asked to reason better in a single prompt.

What would settle it

Take a held-out subset of TN-RCA530 scenarios, strip the AlarmCause nodes and causedBy edges from their input graphs, run the Auto-RCA repair loop only on the remaining scenarios, and measure F1 on the stripped held-out set; a large drop from 91.79% would show the answer is largely encoded in the provided graph structure.

Watch

Extended reading notes

Core claim

The paper's central claim is that TN-RCA530 is the first real-world, publicly accessible benchmark for alarm-based telecom RCA, and that it exposes a clear performance ceiling: the best direct LLM reaches only 0.6254 mixed F1, with most strong models below 0.62. Against that ceiling, Auto-RCA orchestrates an 'evaluate-analyze-repair' loop around a Python solution. The Evaluator runs the code over all 530 scenarios and collects the wrong outputs as Bad Cases; the Bad Case Analyzer identifies the dominant failure pattern; the LLM Agent rewrites the code to remove that systemic flaw; and the Orchestrator accepts a change only if the benchmark F1 improves. With Gemini-2.5-Pro as the reasoning engine, the system lifts the solution from a 0.5899 direct baseline to 0.9179 F1 on the mixed set, with recall 0.9767 and precision 0.8951. The paper concludes that the effective unit for complex domain reasoning is not the LLM alone but the LLM embedded in an iterative, evaluator-driven system.

Load-bearing premise

The load-bearing premise is that the benchmark's ground-truth labels are correct, complete, and not leaked by the input representation, since each scenario already lists candidate AlarmCause nodes and causedBy edges that could give away the answer.

Editorial extensions

If this is right

  • Direct LLM evaluation on TN-RCA530 gives the community a reusable baseline: no tested model exceeds 0.6254 mixed F1, so future telecom RCA methods can be ranked on a fixed, public yardstick.
  • Auto-RCA with Gemini-2.5-Pro reaches 0.9179 mixed F1 in one accepted round, and the acceptance-only Orchestrator keeps subsequent rounds monotonic, scoring rejected candidates as 0.0000.
  • With 94.5% of scenarios graded Difficult, the benchmark mostly measures a model's ability to disambiguate among plausible causes rather than to trace single unbroken paths.
  • Because the loop is not tied to telecom-specific code, the paper's stated next step is to apply Auto-RCA to other graph-based reasoning problems with verifiable solutions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the candidate AlarmCause nodes and causedBy edges were removed or paraphrased in the input graphs, the refined code's F1 advantage might shrink sharply; rerunning the same loop on such a variant would separate learning the benchmark's format from learning to do causal reasoning.
  • The one-round jump from 0.5899 to 0.9179 suggests the largest gain comes from discovering the benchmark's generation rules and label conventions, so a held-out scenario split is needed to test whether the loop generalizes or overfits to recurring root-cause patterns.
  • Because matching is exact on (cause description, equipment id), near-miss answers count as false, so part of the remaining 8.21% gap could be lexical rather than diagnostic; a semantic-matching evaluation would quantify that.
  • The same evaluate-analyze-repair architecture should transfer to other alarm-dense, verifiable domains such as cloud microservice incident diagnosis, provided a simulator can produce trustworthy ground truth.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces TN-RCA530, a benchmark of 530 telecommunication alarm fault scenarios constructed from expert-validated knowledge graphs, and proposes Auto-RCA, an agentic framework that iteratively refines a Python-based solution by evaluating it on the full benchmark, analyzing bad cases, and regenerating code. The main empirical claim is that Auto-RCA lifts macro-averaged F1 from a direct-LLM baseline of 58.99% (Gemini-2.5-Pro) to 91.79% on TN-RCA530. The paper also reports baseline results for several LLMs and a difficulty grading scheme that classifies 501 of the 530 scenarios as 'Difficult'.

Significance. If the benchmark is released and the evaluation protocol is validated, TN-RCA530 would be a useful domain resource, and the Auto-RCA 'evaluate-analyze-repair' loop is an interesting contribution to agentic program repair. The idea of learning from aggregate failure patterns rather than individual errors is sensible and worth pursuing. However, the significance of the headline results is currently undermined by two load-bearing issues: the input representation may reveal the answer via pre-listed AlarmCause nodes and causedBy edges, and the iterative optimization is performed and selected using the full benchmark ground truth, so the reported 91.79% is likely an overfit to the evaluation set rather than an unbiased estimate of generalization. The paper also does not provide artifact access, error bars, or a trivial-baseline comparison, all of which are needed to establish the claims.

major comments (3)
  1. [Data Representation and Statistics (Knowledge Graph Structure)] The input.json representation explicitly contains Root Cause Nodes (AlarmCause) as candidate nodes and causedBy edges that link a root cause to an alarm. The ground truth in label.json is one of these pre-listed candidates. As a result, the core task can be reduced to following causedBy edges from TargetAlarm nodes to AlarmCause nodes. The paper provides no ablation that hides or removes these cues and no trivial baseline (e.g., a script that outputs all AlarmCause nodes reachable from a TargetAlarm) to show that the reported 91.79% F1 is not simply graph traversal. Please add such an ablation and a graph-walking baseline, and report whether their F1 approaches the claimed result.
  2. [The Auto-RCA Agentic Framework and Evaluation Protocol] The iterative loop uses all 530 scenarios and their ground-truth labels as the training signal: the Evaluator computes F1 on the entire benchmark, and the Orchestrator accepts a code change only if it improves F1 on that same set. Consequently, the final 91.79% is an optimized fit to the evaluation set, and the comparison with direct LLM baselines is not an apples-to-apples generalization comparison. The paper should be revised to use a held-out split (e.g., 80/20 or k-fold), with iteration selection performed only on the training/validation portion and the final score reported on the held-out portion. Repeated runs with different seeds should also be reported, since single-run agentic trajectories can be highly variable.
  3. [Table 1 (RQ1 baseline results)] The F1 values in Table 1 are internally inconsistent with the stated macro-averaged harmonic-mean definition. For example, DeepSeek-R1-671B Mixed has Precision@1=0.4471 and Recall@1=0.9537, whose harmonic mean is 0.6088, while the table reports 0.5897; Gemini-2.5-Pro Difficult has Precision@1=0.4311 and Recall@1=0.9658, whose harmonic mean is 0.5961, while the table reports 0.5824; Qwen3-32B Difficult similarly reports 0.5789 instead of the harmonic mean 0.5947. Please reconcile the metric calculations and recheck whether the reported baseline F1 values, and hence the claimed improvement of Auto-RCA over the best baseline, are correct.
minor comments (4)
  1. [Abstract and Introduction] The paper claims TN-RCA530 is 'publicly accessible' but does not provide a URL or repository for the benchmark or the evaluation code. Please add an artifact link so the community can verify the results.
  2. [Table 1 (Simple scenarios)] For Simple scenarios, many models report an identical Precision@1, Recall@1, and F1 of 0.8621, which corresponds to 25 correct out of 29. It would be helpful to explain why even the best models fail on 4 of the 29 unambiguous scenarios, especially if the exact-match criterion for equipment_id is the cause.
  3. [Evaluation Framework and Metrics] The matching criterion is an exact tuple match of (cause_description, equipment_id). The paper should clarify how minor textual differences in cause_description (e.g., synonyms, spacing, case) are normalized, since this materially affects F1 and cross-model comparability.
  4. [Figure 5 caption] The caption states that a score of 0.0000 signifies that a code modification was rejected, but the figure itself has no explicit legend or axis explanation showing this convention. Please add a note to the figure or its caption making this explicit.

Circularity Check

2 steps flagged · score 8.0 of 10

Auto-RCA's 91.79% is an optimized fit to the benchmark it is evaluated on, and the benchmark's ground truth is embedded in the input graph by construction.

  1. self definitional [Methodology > Benchmark Construction > Verifiability; Data Representation and Statistics]
    "Since every case is constructed by tracing from a known, expert-verified root cause to its resulting alarms, the correct answer (the root cause) is embedded in the data by design."

    The input graph contains AlarmCause nodes as candidate root causes and causedBy edges 'between a root cause and an alarm'; the prompt's Step 2 instructs the model to 'traverse all possible paths from it to the root cause nodes AlarmCause'. Because each scenario is generated backward from the true root cause, the correct AlarmCause is connected to the TargetAlarm by construction. The benchmark's target is therefore an element of the input representation, and F1 measures selection from pre-wired candidates rather than independent diagnosis.

  2. fitted input called prediction [The Auto-RCA Agentic Framework > System Architecture and Iterative Workflow; Experiments RQ2]
    "The Evaluator (Test Engineer): Quantifies the performance of a given code solution. It runs the code against the entire TN-RCA530 benchmark, calculates a precise F1-Score, and compiles all incorrect outputs into a set of Bad Cases for analysis. ... Evaluation & Decision: The candidate code is sanitized and re-evaluated on the benchmark. The Orchestrator compares the new F1-score to the previous best. If the score improves, the new code is accepted and becomes the basis for the next iteration."

    The reported final F1 of 91.79% is the value of the same TN-RCA530 objective used to accept or reject every code revision across the five iterative rounds. No held-out split or external evaluation is reported, so the headline number is an optimized fit to the evaluation set, not an independent estimate of the framework's problem-solving ability. The claim of 'state-of-the-art problem-solving capability' reduces by construction to maximizing F1 on the very benchmark used for scoring.

full rationale

The paper is transparent that Auto-RCA iteratively optimizes code against the full TN-RCA530 benchmark, but that transparency does not remove the circularity: the final 91.79% is the objective value that the Orchestrator selected for, so presenting it as evidence that the framework 'masters' the benchmark is a fitted-input-called-prediction. Independently, the benchmark construction makes the answer structurally recoverable from the input: scenarios are generated backward from an expert-verified root cause, the input graph lists that root cause as an AlarmCause candidate, and causedBy edges plus the prompt's explicit path-traversal instruction connect the TargetAlarm to the ground truth. The direct-LLM baselines are legitimate measurements, and there is no load-bearing self-citation chain; the circularity lies in the benchmark's self-definitional answer encoding and in the optimize-and-report-on-the-same-set evaluation protocol.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities requiring independent evidence. The central claim depends on benchmark label quality, metric fidelity, absence of leakage, and the five-round optimization budget.

free parameters (1)
  • Auto-RCA iteration budget = 5 rounds
    The system stops after five refinement rounds with no convergence criterion; the reported 91.79% depends on this stopping point.
assumptions (4)
  • domain assumption Expert-defined backward rules from root cause to alarms enumerate all and only the alarms each fault triggers, so the benchmark labels are complete and correct.
    The Verifiability section says ground truth is embedded by design and expert-verified, but no independent validation of rule completeness or coverage is provided.
  • domain assumption The exact tuple match on (cause description, equipment id) is a faithful measure of correct RCA; paraphrases or partial matches are not counted.
    Evaluation Framework, Matching: a true positive requires exact match, so semantically correct answers can be scored as false negatives.
  • domain assumption Candidate AlarmCause nodes and causedBy edges in the input graphs do not trivially reveal the ground truth.
    Data Representation states input.json includes root cause candidate nodes and causedBy edges; the paper does not test whether these leak labels.
  • domain assumption Single-run LLM evaluations are stable enough to support the reported F1 differences.
    Tables 1 and 2 report one F1 per model without repeated runs or intervals.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TN-AutoRCA: Benchmark Construction and Agentic Framework for Self-Improving Alarm-Based Root Cause Analysis in Telecommunication Networks." pith.science (2026). https://pith.science/paper/CXYTYBAX

@misc{pith2026250718190,
  author       = {Pith},
  title        = {Pith review of: TN-AutoRCA: Benchmark Construction and Agentic Framework for Self-Improving Alarm-Based Root Cause Analysis in Telecommunication Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CXYTYBAX}},
  note         = {Machine review of arXiv:2507.18190}
}
read the original abstract

Root Cause Analysis (RCA) in telecommunication networks is a critical task, yet it presents a formidable challenge for Artificial Intelligence (AI) due to its complex, graph-based reasoning requirements and the scarcity of realistic benchmarks.

Figures

Figures reproduced from arXiv: 2507.18190 by the authors.

Figure 1
Figure 1. The framework for LLM-based root cause analy [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the cycle consistency check for auto [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Distribution and Complexity Stratification of Root Causes in TN-RCA530. The chart displays the frequency of major [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The architecture of the Auto-RCA agentic frame [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Performance Improvement on TN-RCA530 using [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

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

    cs.AI 2025-07 conditional novelty 5.0 of 10

    A two-stage SFT+RL fine-tuning pipeline for reasoning LLMs achieves 95.86% root cause accuracy on a new synthetic 5G troubleshooting benchmark, TeleLogs.

Reference graph

Works this paper leans on

34 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Amayuelas, A.; Sain, J.; Kaur, S.; and Smiley, C. 2025. Grounding LLM Reasoning with Knowledge Graphs. arXiv:2502.13247

    arXiv 2025

  4. [4]

    Arora, D.; Sonwane, A.; Wadhwa, N.; Mehrotra, A.; Utpala, S.; Bairi, R.; Kanade, A.; and Natarajan, N. 2024. MASAI: Modular Architecture for Software-engineering AI Agents. arXiv:2406.11638

    arXiv 2024

  5. [5]

    Barriah, L.; De Domenico, A.; Powell, L.; Sana, M.; Wang, P.; Piovesan, N.; and Debbah, M. 2025. GSMA Open-Telco LLM Benchmarks. https://huggingface.co/blog/otellm/gsma-benchmarks

    work page 2025

  6. [6]

    Bouloutas, A.; Calo, S.; and Finkel, A. 1994. Alarm correlation and fault identification in communication networks. Communications, IEEE Transactions on, 42: 523 -- 533

    work page 1994

  7. [7]

    Campos, V.; Shariffdeen, R.; Ulges, A.; and Noller, Y. 2025. Empirical Evaluation of Generalizable Automated Program Repair with Large Language Models. arXiv:2506.03283

    work page arXiv 2025

  8. [8]

    DeepSeek-AI; Guo, D.; Yang, D.; Zhang, H.; Song, J.; Zhang, R.; Xu, R.; Zhu, Q.; Ma, S.; Wang, P.; Bi, X.; Zhang, X.; Yu, X.; Wu, Y.; Wu, Z. F.; Gou, Z.; Shao, Z.; Li, Z.; Gao, Z.; Liu, A.; Xue, B.; Wang, B.; Wu, B.; Feng, B.; Lu, C.; Zhao, C.; Deng, C.; Zhang, C.; Ruan, C.; Dai, D.; Chen, D.; Ji, D.; Li, E.; Lin, F.; Dai, F.; Luo, F.; Hao, G.; Chen, G.; ...

    arXiv 2025

Show all 34 references
  1. [9]

    S.; Sharma, M.; and Rajabi, E

    Dehal, R. S.; Sharma, M.; and Rajabi, E. 2025. Knowledge Graphs and Their Reciprocal Relationship with Large Language Models. Machine Learning and Knowledge Extraction, 7(2)

    2025

  2. [10]

    Ethiraj, V.; Vijay, D.; Menon, S.; and Berscilla, H. 2025. Efficient Telecom Specific LLM: TSLAM-Mini with QLoRA and Digital Twin Data. arXiv:2505.07877

    2025 arXiv

  3. [11]

    Ibrahim, N.; Aboulela, S.; Ibrahim, A.; and Kashef, R. 2024. A survey on augmenting knowledge graphs (KGs) with large language models (LLMs): models, evaluation metrics, benchmarks, and challenges. Discover Artificial Intelligence, 4(1)

    2024

  4. [12]

    Jakobson, G.; and Weissman, M. 1995. Real-time telecommunication network management: extending event correlation with temporal constraints. In Proceedings of the Fourth International Symposium on Integrated Network Management IV, 290–301. GBR: Chapman & Hall, Ltd. ISBN 0412715708

    1995

  5. [13]

    Ji, S.; Liu, L.; Xi, J.; Zhang, X.; and Li, X. 2024. KLR-KGC: Knowledge-Guided LLM Reasoning for Knowledge Graph Completion. Electronics, 13(24)

    2024

  6. [14]

    Just, R.; Jalali, D.; and Ernst, M. D. 2014. Defects4J : A Database of Existing Faults to Enable Controlled Testing Studies for Java Programs. In Proceedings of the 2014 International Symposium on Software Testing and Analysis, ISSTA '14, 437--440. New York, NY, USA: ACM

    2014

  7. [15]

    Kamoi, R.; Zhang, Y.; Zhang, N.; Han, J.; and Zhang, R. 2024. When Can LLM s Actually Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLM s. Transactions of the Association for Computational Linguistics, 12: 1417--1440

    2024

  8. [16]

    Kong, J.; Cheng, M.; Xie, X.; Liu, S.; Du, X.; and Guo, Q. 2024. ContrastRepair: Enhancing Conversation-Based Automated Program Repair via Contrastive Test Case Pairs. arXiv:2403.01971

    2024

  9. [17]

    Li, M.; Yang, M.; and Chen, P. 2023. Alarm reduction and root cause inference based on association mining in communication network. Frontiers in Computer Science, Volume 5 - 2023

    2023

  10. [18]

    Liang, R.; Liu, F.; and Liu, J. 2020. A Belief Network Reasoning Framework for Fault Localization in Communication Networks. Sensors (Basel, Switzerland), 20

    2020

  11. [19]

    Ma, X.; Zhen, W.; Ren, H.; Zhang, G.; Zhang, K.; and Dong, H. 2024. A Method for Fault Localization in Distribution Networks with High Proportions of Distributed Generation Based on Graph Convolutional Networks. Energies, 17(22)

    2024

  12. [20]

    Ma, Y.; Li, Y.; Dong, Y.; Jiang, X.; Cao, R.; Chen, J.; Huang, F.; and Li, B. 2025. Thinking Longer, Not Larger: Enhancing Software Engineering Agents via Scaling Test-Time Compute. arXiv:2503.23803

    2025 arXiv

  13. [21]

    C.; Ying, R.; and Tassiulas, L

    Maatouk, A.; Ampudia, K. C.; Ying, R.; and Tassiulas, L. 2025. Tele-LLMs: A Series of Specialized Large Language Models for Telecommunications. arXiv:2409.05314

    2025 arXiv

  14. [22]

    Pan, L.; Saxon, M.; Xu, W.; Nathani, D.; Wang, X.; and Wang, W. Y. 2024. Automatically Correcting Large Language Models: Surveying the Landscape of Diverse Automated Correction Strategies. Transactions of the Association for Computational Linguistics, 12: 484--506

    2024

  15. [23]

    Pham, L.; Zhang, H.; Ha, H.; Salim, F.; and Zhang, X. 2025. RCAEval: A Benchmark for Root Cause Analysis of Microservice Systems with Telemetry Data. arXiv:2412.17015

    2025 arXiv

  16. [24]

    Roy, D.; Zhang, X.; Bhave, R.; Bansal, C.; Las-Casas, P.; Fonseca, R.; and Rajmohan, S. 2024. Exploring LLM-Based Agents for Root Cause Analysis. In Companion Proceedings of the 32nd ACM International Conference on the Foundations of Software Engineering, FSE 2024, 208–219. Ne...

    2024

  17. [25]

    Shen, T.; Wang, J.; Zhang, X.; and Cambria, E. 2025. Reasoning with Trees: Faithful Question Answering over Knowledge Graph. In Rambow, O.; Wanner, L.; Apidianaki, M.; Al-Khalifa, H.; Eugenio, B. D.; and Schockaert, S., eds., Proceedings of the 31st International Conference on...

    2025

  18. [26]

    Shen, X.; Wang, F.; and Xia, R. 2025. Reason-Align-Respond: Aligning LLM Reasoning with Knowledge Graphs for KGQA. arXiv:2505.20971

    2025 arXiv

  19. [27]

    Team, K.; Du, A.; Gao, B.; Xing, B.; Jiang, C.; Chen, C.; Li, C.; Xiao, C.; Du, C.; Liao, C.; Tang, C.; Wang, C.; Zhang, D.; Yuan, E.; Lu, E.; Tang, F.; Sung, F.; Wei, G.; Lai, G.; Guo, H.; Zhu, H.; Ding, H.; Hu, H.; Yang, H.; Zhang, H.; Yao, H.; Zhao, H.; Lu, H.; Li, H.; Yu, ...

    2025 arXiv

  20. [28]

    Team, Q. 2025. QwQ-32B: Embracing the Power of Reinforcement Learning

    2025

  21. [29]

    Wang, Q.; Zhang, X.; Li, M.; Yuan, Y.; Xiao, M.; Zhuang, F.; and Yu, D. 2025. TAMO:Fine-Grained Root Cause Analysis via Tool-Assisted LLM Agent with Multi-Modality Observation Data in Cloud-Native Systems. arXiv:2504.20462

    2025

  22. [30]

    Wang, Z.; Huang, B.; Zhou, B.; Chen, J.; and Wang, Y. 2024. An Enhanced Fault Localization Technique for Distribution Networks Utilizing Cost-Sensitive Graph Neural Networks. Processes, 12(11)

    2024

  23. [31]

    Xu, J.; Zhang, Q.; Zhong, Z.; He, S.; Zhang, C.; Lin, Q.; Pei, D.; He, P.; Zhang, D.; and Zhang, Q. 2025. OpenRCA: Can Large Language Models Locate the Root Cause of Software Failures? In The Thirteenth International Conference on Learning Representations

    2025

  24. [32]

    Yang, A.; Li, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Gao, C.; Huang, C.; Lv, C.; Zheng, C.; Liu, D.; Zhou, F.; Huang, F.; Hu, F.; Ge, H.; Wei, H.; Lin, H.; Tang, J.; Yang, J.; Tu, J.; Zhang, J.; Yang, J.; Yang, J.; Zhou, J.; Zhou, J.; Lin, J.; Dang, K.; Bao, K.; ...

    2025 arXiv

  25. [33]

    Zhang, B. 2023. Root Cause Analysis of Communication Network Based on Deep Fuzzy Neural Network. IEEE Access, PP: 1--1

    2023

  26. [34]

    Zhang, K.; Kalander, M.; Zhou, M.; Zhang, X.; and Ye, J. 2021. An Influence-based Approach for Root Cause Alarm Discovery in Telecom Networks. arXiv:2105.03092

    2021 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.