Pith. sign in

REVIEW 2 major objections 5 minor 75 references

Bifrost, a self-supervised contrastive method, trains pretrained language models to encode the multi-level 'fallibility representations' of system logs—execution-flow, event, and component information—and reports average gains of 9.83% in a

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 03:23 UTC pith:BC33HDON

load-bearing objection Bifrost is a well-run, genuinely useful contrastive learning method for log diagnosis, but its synthetic negatives may be teaching lexical shortcuts rather than multi-level fallibility structure. the 2 major comments →

arxiv 2607.23169 v1 pith:BC33HDON submitted 2026-07-25 cs.SE

Bifrost: Empowering Pretrained Language Model with Fallibility Representation for Log-Based Fault Diagnosis

Minghua He , Tong Jia , Lingzhe Zhang , Chiming Duan , Xinlong Zhao , Leyi Pan , Cheng Wang , Kangjin Wang
show 4 more authors
Yinghao Yu Liping Zhang Yifan Wu Ying Li
This is my paper
classification cs.SE
keywords log-based fault diagnosisfallibility representationcontrastive learninganomaly detectionroot cause localizationfault identificationpretrained language modelslog representation learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that natural-language-pretrained models fail at log-based fault diagnosis because faults live in the multi-level structure of logs—unexpected execution flows, anomalous event semantics, and component identity—not in surface wording. It introduces Bifrost, a representation-learning method that injects synthetic fault cues into normal logs and uses three self-supervised contrastive objectives (execution-flow prediction, abnormal-event discrimination, component perception) to shape embeddings around those fault signals. A fourth mechanism, co-anchored fallibility representation learning (CARL), ties all objectives to real log events and shared negatives to stop stochastic sampling noise from accumulating in the model weights. Across three public log datasets and one industrial ML-as-a-Service system, Bifrost's embeddings improve anomaly-detection F1 by 9.83%, root-cause-localization HR@k by 18.28%, and fault-identification Macro-F1 by 20.88% on average over existing PLMs. The paper's overall claim is that log embeddings, not just larger models, are the bottleneck, and fault-aware contrastive objectives can fix them without needing labeled failures.

Core claim

The central claim is that pretrained language models encode natural-language semantics but systematically miss 'fallibility representations'—the fault-relevant structure that operators read at execution-flow, event, and component levels. The paper first shows empirically that in PLM embedding spaces, expected and unexpected events have high similarity, normal and anomalous events overlap heavily, and different system components collapse together. Bifrost then operationalizes fallibility as three relational contrastive objectives: Execution Flow Prediction aligns a flow-context anchor with flow-consistent evidence while contrasting against counterfactual continuations with injected fault phra

What carries the argument

The central object is the fallibility representation, learned through three self-supervised InfoNCE contrastive objectives (EFP, AED, SCP). Each objective defines anchors, positives, and negatives through injected counterfactuals: fault phrases from a corpus-grounded lexicon or alternative component keywords inserted into normal logs. CARL is the stabilizer: it reuses the real log event as a common anchor, isolates context-sampling randomness into a lower-weighted contrastive term, and shares a negative set across objectives, which the paper analyzes as reducing per-step gradient variance and bounding cumulative weight error.

Load-bearing premise

The paper's gains depend on the assumption that synthetic negatives—normal logs with fault phrases or alternative component keywords spliced in—approximate real system failures closely enough that contrastive learning on them transfers to genuine anomalies; if injected phrases are merely learnable lexical shortcuts, the reported improvements on real test logs would be inflated.

What would settle it

Take a set of genuine failure logs, remove or mask every phrase that appears in the constructed fault lexicon, and re-run Bifrost-trained anomaly detection and root-cause localization; if scores fall to baseline PLM levels, the gains are an artifact of injected phrase patterns. A complementary check: compare Bifrost's held-out performance on failures whose phrases never appear in the lexicon with its performance on failures that do—a large positive gap would indicate real transfer.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Anomaly detection, root-cause localization, and fault identification all improve together from the same embeddings, so one Bifrost encoder can serve all three diagnosis tasks in a production pipeline.
  • Because Bifrost learns from normal-operation logs plus an automatically generated fault lexicon, no labeled fault data is required to adapt a PLM to a new system's failure language.
  • Bifrost's 140M-parameter model matches or beats much larger embedding models on most metrics, making fault-aware representation learning a cheaper path to diagnosis accuracy than scaling model size.
  • The fallibility-jitter analysis implies that multi-objective contrastive learning can be destabilized by per-step sampling variance, and that anchoring objectives on real instances and sharing negatives reduces that variance.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The CARL negative-covariance analysis suggests a testable general principle for multi-task contrastive learning beyond logs: sharing negatives across objectives can lower gradient variance beyond what standard contrastive learning provides, which could be probed in multimodal or graph representation learning.
  • Because the fault lexicon is built from the training corpus through embedding retrieval, Bifrost's gains may depend on how well those embeddings already understand failure language; a system whose logs use rare proprietary error jargon might need a lexicon built from its own historical incidents to see the same benefit.
  • The industrial fault-identification result is the only evidence for the 20.88% Macro-F1 gain, so a natural next experiment is running Bifrost's representations on an incident corpus with fine-grained fault labels from a different platform to test whether that gain transfers beyond Platform-X.
  • A sharp test of whether Bifrost learns true event semantics rather than injection patterns would be to evaluate it on genuine failures whose phrases never appear in the constructed lexicon; the paper's cross-system results hint at transfer but do not report this directly.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes Bifrost, a log representation learning method that fine-tunes BART-Base with three self-supervised contrastive objectives intended to capture multi-level 'fallibility representations' of logs: Execution Flow Prediction (EFP), Abnormal Event Discrimination (AED), and System Component Perception (SCP). A Co-Anchored Representation Learning (CARL) strategy is introduced to reduce gradient variance ('fallibility jitter'), supported by a theoretical analysis in Appendices B and C. The method is evaluated on three public datasets (BGL, Hadoop, Thunderbird) and an industrial MLaaS dataset (Platform-X) for anomaly detection, root cause localization, and fault identification, reporting large gains over general PLMs, larger embedding models, and log-specific baselines, with a BART-Log control intended to isolate the effect of additional training.

Significance. If the reported results are taken at face value, Bifrost is a practically useful and parameter-efficient way to adapt PLMs to system logs for downstream fault diagnosis. The paper has clear strengths: a BART-Log control that isolates additional training, five-run mean±std reporting, ablations for each objective and for CARL, public code, and an explicit acknowledgment in the threats-to-validity section that injected counterfactuals are only an approximation. The main significance, however, rests on whether the synthetic injected negatives generalize to real failures rather than acting as a lexical shortcut. The current experimental design does not directly demonstrate that transfer, so the headline cross-system margin claims are not yet fully established.

major comments (2)
  1. [Sec. 4.2.2, 4.2.3, 4.2.5, RQ3 (Sec. 5.4)] The negative instances for EFP and AED are constructed by inserting phrases from a fault lexicon A that is seeded with warning/error/fatal and selected from normal training logs via embedding similarity. The central claim of transferable fallibility representations therefore depends on the assumption that these injected phrases approximate the distribution of real failure logs. RQ3 (Figure 5) varies only the size of the lexicon; it does not test whether real anomalous logs actually contain the lexicon phrases, nor whether Bifrost's advantage persists when those phrases are absent or masked. Without such a control, the margin over BART-Log could be explained by a detector of a small set of injected lexical cues rather than by multi-level fallibility representations. Please add (a) a non-failure lexical-injection control (e.g., random or generic words as negatives), and (b) an analysis of
  2. [Appendix C, Eqs. (24) and (28)] The CARL variance-reduction argument relies on two unproven assumptions: comparable gradient scales between EFP and LCL, and positive covariance Cov_i(g_A,i, g_B,i) > 0. The positivity is asserted from a shared 'push away' intuition, but the gradients act on different loss terms and different sampled instances; the sign of their covariance is not a formal consequence. Since the paper presents CARL as a theoretically motivated strategy, these assumptions should either be derived under stated conditions or verified empirically by computing the covariance during training. If they hold only under favorable conditions, the claim in Sec. 4.3.2 that CARL 'reduces parameter error through two structural mechanisms' should be softened accordingly.
minor comments (5)
  1. [Sec. 4.2.5 vs Sec. 5.1.4 and RQ3] Notation and defaults are inconsistent: Sec. 4.2.5 says m=30 by default, but Sec. 5.1.4 says the system failure corpus size is 10; RQ3 (Sec. 5.4) calls the varied quantity K, which Eqs. (3)-(5) use for the number of sampled negatives, not the lexicon size. Please reconcile the notation and state exactly what is varied in Figure 5.
  2. [Eq. (9)] The denominator of the SCP loss sums over k=1..K, but the negatives are indexed by z=1..Z. This is a typo that should be corrected to make the loss definition precise.
  3. [Sec. 3.1, Table 1] The two metrics in Table 1 conflict for some models: e.g., BERT-Base has higher Avg_Sim (worse) but higher Avg_Rank (better) than GloVe-300d. The text says it is 'readily apparent' that PLMs fail on unexpected flow identification, but the table does not give a consistent ordering. Please discuss or reconcile this, since it weakens the empirical-study claim.
  4. [Sec. 5.1.3] The construction of root-cause ground truth for BGL, Thunderbird, and Platform-X is not described. The reader cannot assess whether RCL labels are the alert lines, SRE annotations, or something else. Please specify the labeling protocol.
  5. [Sec. 4.2.5] The fault lexicon is built using Qwen3-Embedding and Qwen3-Reranker, which are external pretrained models. The 'self-supervised' characterization of the objectives is therefore qualified: the negative construction depends on natural-language embeddings. This should be acknowledged explicitly, ideally with an analysis of sensitivity to the choice of embedding/reranker.

Circularity Check

0 steps flagged

No significant circularity: the main AD/RCL/FI results are measured on held-out test logs with real labels, and the training objectives are self-supervised counterfactual contrasts rather than fitted versions of the test predictions.

full rationale

The derivation chain is self-contained. Bifrost's EFP/AED/SCP objectives are self-supervised contrasts built from normal training logs, and the reported AD, RCL, and FI numbers come from held-out test logs with real labels: public LogHub benchmarks for BGL/Hadoop/Thunderbird and a separately presented industrial case study for Platform-X. No reported prediction reduces to a training input by construction: test anomalies are observed logs, not the injected counterfactuals of Eqs. (3)/(5)/(8), and the FI classifier is trained on manually labeled Platform-X faults rather than read out from the SCP vocabulary. The RQ4 representation analysis does reuse the Section 3 probes, but 'fallibility representation' is defined by SRE multi-level analysis, not by those probes, so this is a construct-validity check rather than a logical reduction. The fallibility-jitter analysis in Appendices B and C is a standard SGD variance bound and does not assume CARL's conclusion. The paper's many author-overlapping references (LLMeLog, AIMS, FAMOS, etc.) support background claims that the paper also tests itself in Section 3 and are not load-bearing. The admitted limitation in Section 6 that 'injection is an approximation' and the proprietary nature of Platform-X are correctness/validity threats, not circularity. Overall, no load-bearing circular step was identified.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central claim depends on three families of inputs: (i) hyperparameters (λ, K, τ, w) fixed by the authors or validation; (ii) the operational assumption that synthetic fault phrases and component keywords inserted into normal logs teach representations that transfer to real failures; and (iii) two unproven technical assumptions in the CARL variance analysis (positive gradient covariance and comparable gradient scales). 'Fallibility representation' is a new construct with no independent evidence outside the paper's own tasks.

free parameters (5)
  • Loss weights λ1..λ4 = (1, 0.5, 1, 1)
    Set in Section 5.1.4 as implementation choices; validation set used for tuning; no sensitivity analysis shown.
  • Failure lexicon size K (m) = 10 (default; §4.2.5 says m=30)
    Number of fault phrases injected as negatives; RQ3 tests K=4..larger but the default value is internally inconsistent (§4.2.5 vs §5.1.4/§5.4).
  • Component vocabulary C = 8 keywords: cpu, gpu, disk, hardware, memory, kernel, network, file
    Predefined set used to construct SCP anchors/negatives; chosen by authors, not learned or validated across systems (Appendix A).
  • Contrastive temperature τ = not reported
    Appears in Eq (1)/(4)/(6)/(9) and affects all three objectives; no default value or tuning procedure given.
  • Sliding window w = 20
    Used for EFP context and downstream sequence grouping; inherited from prior work [20,33], not fitted here.
axioms (5)
  • standard math Gradient Lipschitz smoothness: ||G(W_a)-G(W_b)|| ≤ L_Lip ||W_a-W_b||
    Assumed in Appendix B to derive the fallibility-jitter bound; conventional for optimization analysis but unverified for this model.
  • domain assumption Injected counterfactual negatives approximate real failure distributions
    EFP/AED objectives train only on normal logs with inserted fault phrases (Eq 3,5); transfer to real anomalies is assumed, partially tested only by varying K.
  • domain assumption SRE three-level analysis (execution flow, event, component) is the right decomposition of log fallibility
    The entire objective design derives from this framing (Section 1, Figure 1); no evidence that these levels are sufficient or orthogonal.
  • ad hoc to paper Gradient terms g_A,i and g_B,i are positively correlated, Cov>0
    Appendix C Eq (28) asserts this without derivation or empirical verification; the negative-covariance variance reduction in Eq (31) depends on it.
  • ad hoc to paper Comparable gradient scales across EFP/LCL and λ2<λ1
    Variance isolation mechanism (Appendix C Eq 24) relies on this; no gradient-scale measurement provided.
invented entities (2)
  • Fallibility representation no independent evidence
    purpose: Conceptual target of Bifrost's three contrastive objectives; claimed to capture multi-level fault information in log embeddings.
    Defined operationally by the paper's own tasks (EFP/AED/SCP); no external measurement or falsifiable handle beyond the proposed training objectives.
  • Fallibility jitter no independent evidence
    purpose: Named phenomenon of optimization instability caused by stochastic sampling in the three objectives; motivates the CARL strategy.
    This is a re-description of standard SGD gradient noise with a new name; the paper's bound (Eq 23) is a standard variance-accumulation bound, not a new empirical entity.

pith-pipeline@v1.3.0-alltime-deepseek · 26651 in / 13369 out tokens · 116054 ms · 2026-08-01T03:23:36.310386+00:00 · methodology

0 comments
read the original abstract

Log-based fault diagnosis is crucial for runtime debugging and maintenance. Existing fault diagnosis methods use language models pre-trained on natural language (PLMs) for log representation. However, system faults are reflected in the multi-level structure of system logs. PLMs pre-trained on natural language struggle to comprehensively capture multi-level fault information, failing to meet the requirements of fault diagnosis. We refer to this information as fallibility representations. To address this problem, we propose a novel log representation learning method, Bifrost. It draws inspiration from the log analysis experience of Site Reliability Engineers and meticulously designs strategies based on self-supervised contrastive learning to learn the fallibility representations of logs. Across three public systems and one industrial ML-as-a-Service system, the log representations produced by Bifrost outperform existing PLMs by average margins of 9.83% in F1 for anomaly detection, 18.28% in HR@k for root cause localization, and 20.88% in Macro-F1 for fault identification.

Figures

Figures reproduced from arXiv: 2607.23169 by Cheng Wang, Chiming Duan, Kangjin Wang, Leyi Pan, Lingzhe Zhang, Liping Zhang, Minghua He, Tong Jia, XinLong Zhao, Yifan Wu, Yinghao Yu, Ying Li.

Figure 1
Figure 1. Figure 1: Practical log analysis experience of SREs. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: System-event Representations on BGL generated by PLMs. (a) BERT-Base (b) BART-Base [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: System-component Representations generated by PLMs [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Bifrost is initialized from the PLM BART-Base. It is then trained via a carefully designed representation learning process, which learns log fallibility representations from logs collected during normal system operation. After training, Bifrost encodes logs for fault diagnosis. that introduce failure cues and are unlikely under the normal execu￾tion flow. Specifically, we sample 𝐾 negatives { ˜ 𝑙𝑡,𝑘 } 𝐾 𝑘=… view at source ↗
Figure 5
Figure 5. Figure 5: Lexicon dependence of Bifrost with respect to the size of the system failure corpus used for injected failure phrases. (a) Avg_Sim as Metric (b) Avg_Rank as Metric [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Representational similarity in Bifrost. (a) Event-level (b) Component-level [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The log representation generated by Bifrost. 7 Related Work 7.1 Log-based Software System Fault Diagnosis Discovering and diagnosing software problems through log analysis has long been an active research area [13, 39, 51, 58, 64, 69]. Recent log-based fault diagnosis mainly follows two paradigms: LLM-based approaches that reason directly over raw logs or retrieved log con￾texts [13, 15, 31, 44, 52, 58, 63… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

75 extracted references · 8 linked inside Pith

  1. [1]

    Anunay Amar and Peter C Rigby. 2019. Mining historical test logs to predict bugs and localize faults in the test logs. In2019 IEEE/ACM 41st International Conference on Software Engineering (ICSE). IEEE, 140–151

    2019

  2. [2]

    Ching-Yao Chuang, Joshua Robinson, Yen-Chen Lin, Antonio Torralba, and Stefanie Jegelka. 2020. Debiased contrastive learning.Advances in neural information processing systems33 (2020), 8765–8775

    2020

  3. [3]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 NAACL. 4171–4186

    2019

  4. [4]

    Min Du, Feifei Li, Guineng Zheng, and Vivek Srikumar. 2017. Deeplog: Anomaly de- tection and diagnosis from system logs through deep learning. InProceedings of the 2017 ACM SIGSAC conference on computer and communications security. 1285–1298

    2017

  5. [5]

    Chiming Duan, Tong Jia, Minghua He, Pei Xiao, Lingzhe Zhang, Zhewei Zhong, Xin Zhang, and Ying Li. 2026. AIMS: A Content-Aware Resource Management Approach for AI Assistant Systems. InProceedings of the 34th ACM International Conference on the Foundations of Software Engineering. 1232–1236

    2026

  6. [6]

    Chiming Duan, Yong Yang, Tong Jia, Guiyang Liu, Jinbu Liu, Huxing Zhang, Qi Zhou, Ying Li, and Gang Huang. 2025. FAMOS: Fault diagnosis for Microservice Systems through Effective Multi-modal Data Fusion. InICSE 2025. IEEE Computer Society, 610–610

    2025

  7. [7]

    Hongcheng Guo, Jian Yang, Jiaheng Liu, Jiaqi Bai, Boyang Wang, Zhoujun Li, Tieqiao Zheng, Bo Zhang, Junran Peng, and Qi Tian. 2024. Logformer: A pre-train and tuning pipeline for log anomaly detection. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 135–143

    2024

  8. [8]

    Minghua He, Chiming Duan, Pei Xiao, Tong Jia, Siyu Yu, Lingzhe Zhang, Weijie Hong, Jin Han, Yifan Wu, Ying Li, et al . 2025. United We Stand: Towards End-to-End Log-based Fault Diagnosis via Interactive Multi-Task Learning.arXiv preprint arXiv:2509.24364(2025)

    arXiv 2025

  9. [10]

    Minghua He, Tong Jia, Chiming Duan, Huaqian Cai, Ying Li, and Gang Huang

  10. [11]

    Pinjia He, Jieming Zhu, Zibin Zheng, and Michael R Lyu. 2017. Drain: An online log parsing approach with fixed depth tree. In2017 ICWS. IEEE, 33–40

    2017

  11. [12]

    Junjie Huang, Minghua He, Jinyang Liu, Yintong Huo, Domenico Bianculli, and Michael R Lyu. 2025. CodeAD: Synthesize Code of Rules for Log-based Anomaly Detection with LLMs.arXiv preprint arXiv:2510.22986(2025)

    arXiv 2025

  12. [13]

    Junjie Huang, Zhihan Jiang, Zhuangbin Chen, and Michael Lyu. 2025. No More Labelled Examples? An Unsupervised Log Parser with LLMs.Proceedings of the ACM on Software Engineering2, FSE (2025), 2406–2429

    2025

  13. [14]

    Miao Jiang, Mohammad A Munawar, Thomas Reidemeister, and Paul AS Ward

  14. [15]

    Zhihan Jiang, Junjie Huang, Guangba Yu, Zhuangbin Chen, Yichen Li, Renyi Zhong, Cong Feng, Yongqiang Yang, Zengyin Yang, and Michael Lyu. 2025. L4: Diagnosing large-scale llm training failures via automated log analysis. InProceedings of the 33rd ACM International Conference on the Foundations of Software Engineering. 51–63

    2025

  15. [16]

    Mengyuan Jing, Yanmin Zhu, Tianzi Zang, and Ke Wang. 2023. Contrastive self-supervised learning in recommender systems: A survey.ACM Transactions on Information Systems42, 2 (2023), 1–39

    2023

  16. [17]

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. 2020. Supervised contrastive learning.Advances in neural information processing systems33 (2020), 18661–18673

    2020

  17. [18]

    Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. [n. d.]. ALBERT: A Lite BERT for Self-supervised Learning of Lan- guage Representations. InInternational Conference on Learning Representations

  18. [19]

    Van-Hoang Le and Hongyu Zhang. 2021. Log-based anomaly detection without log parsing. In2021 36th IEEE/ACM International Conference on Automated Software Engineering (ASE). IEEE, 492–504

    2021

  19. [20]

    Van-Hoang Le and Hongyu Zhang. 2022. Log-based anomaly detection with deep learning: How far are we?. InProceedings of the 44th international conference on software engineering. 1356–1367

    2022

  20. [21]

    Van-Hoang Le and Hongyu Zhang. 2024. Prelog: A pre-trained model for log analytics.Proceedings of the ACM on Management of Data2, 3 (2024), 1–28

    2024

  21. [22]

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 7871–7880

    2020

  22. [23]

    Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023. Towards general text embeddings with multi-stage contrastive learning.arXiv preprint arXiv:2308.03281(2023)

    Pith/arXiv arXiv 2023

  23. [24]

    Qingwei Lin, Tianci Li, Pu Zhao, Yudong Liu, Minghua Ma, Lingling Zheng, Murali Chintalapati, Bo Liu, Paul Wang, Hongyu Zhang, et al. 2023. Edits: An easy-to-difficult training strategy for cloud failure prediction. InCompanion Proceedings of the ACM Web Conference 2023. 371–375

    2023

  24. [25]

    Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. 2021. Self-supervised learning: Generative or contrastive.IEEE transactions on knowledge and data engineering35, 1 (2021), 857–876

    2021

  25. [26]

    Yuchen Liu, Yixuan Meng, Kaiyuan Xu, Zijun Xu, Tianyuan Wu, Yiwei Yang, and Shu Yin. 2021. Reproducibility: Performance Evaluation of MemXCT on Azure CycleCloud Platform.IEEE Transactions on Parallel and Distributed Systems33, 9 (2021), 2047–2049

    2021

  26. [27]

    Chuan Luo, Pu Zhao, Bo Qiao, Youjiang Wu, Hongyu Zhang, Wei Wu, Weihai Lu, Yingnong Dang, Saravanakumar Rajmohan, Qingwei Lin, et al. 2021. NTAM: neighborhood-temporal attention model for disk failure prediction in cloud platforms. InProceedings of the Web Conference 2021. 1181–1191

    2021

  27. [28]

    Lipeng Ma, Yixuan Li, Weidong Yang, Mingjie Zhou, Xinyi Liu, Ben Fei, Shuhao Li, Xiaoyan Sun, Sihang Jiang, and Yanghua Xiao. [n. d.]. LogReasoner: Empowering LLMs with Expert-like Coarse-to-Fine Reasoning for Automated Log Analysis. ACM Transactions on Software Engineering and Methodology([n. d.])

  28. [29]

    Lipeng Ma, Yixuan Li, Weidong Yang, Mingjie Zhou, Mingyu Zhao, Ben Fei, Shuhao Li, Sihang Jiang, and Yanghua Xiao. 2026. LogInstruct: Knowledge-Driven Instruction Synthesis for Enhancing LLM-Based Log Analysis.IEEE Transactions on Services Computing(2026)

    2026

  29. [30]

    Lipeng Ma, Weidong Yang, Bo Xu, Sihang Jiang, Ben Fei, Jiaqing Liang, Mingjie Zhou, and Yanghua Xiao. 2024. Knowlog: Knowledge enhanced pre-trained language model for log understanding. InProceedings of the 46th ieee/acm international conference on software engineering. 1–13

    2024

  30. [31]

    Zeyang Ma, An Ran Chen, Dong Jae Kim, Tse-Hsun Chen, and Shaowei Wang

  31. [32]

    Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. 2024. SFR-Embedding-Mistral:Enhance Text Retrieval with Transfer Learning. Salesforce AI Research Blog. https://www.salesforce.com/blog/sfr- embedding/

    2024

  32. [33]

    Weibin Meng, Ying Liu, Yichen Zhu, Shenglin Zhang, Dan Pei, Yuqing Liu, Yihao Chen, Ruizhi Zhang, Shimin Tao, Pei Sun, et al. 2019. Loganomaly: Unsupervised detection of sequential and quantitative anomalies in unstructured logs.. InIJCAI, Vol. 19. 4739–4745

    2019

  33. [34]

    George A Miller. 1995. WordNet: a lexical database for English.Commun. ACM 38, 11 (1995), 39–41

    1995

  34. [35]

    InProceedings of the IEEE/ACM 46th International Conference on Software Engineering

    Llmparser: An exploratory study on using large language models for log parsing. InProceedings of the IEEE/ACM 46th International Conference on Software Engineering. 1–13

  35. [36]

    Adam Oliner and Jon Stearley. 2007. What supercomputers say: A study of five system logs. In37th annual IEEE/IFIP international conference on dependable systems and networks (DSN’07). IEEE, 575–584

    2007

  36. [37]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748(2018)

    Pith/arXiv arXiv 2018

  37. [38]

    Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. InProceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). 1532–1543

    2014

  38. [39]

    Shuyi Niu, Jiawei Jin, Xiutian Huang, Yonggeng Wang, Wenhao Xu, and Youyong Kong. 2023. Locating Faulty Applications via Semantic and Topology Estimation. InCompanion Proceedings of the ACM Web Conference 2023. 528–532

    2023

  39. [40]

    Carl Martin Rosenberg and Leon Moonen. 2020. Spectrum-based log diagnosis. In Proceedings of the 14th ACM/IEEE International Symposium on Empirical Software Engineering and Measurement (ESEM). 1–12

    2020

  40. [41]

    Yicheng Sui, Yuzhe Zhang, Jianjun Sun, Ting Xu, Shenglin Zhang, Zhengdan Li, Yongqian Sun, Fangrui Guo, Junyu Shen, Yuzhi Zhang, et al . 2023. Logkg: Log failure diagnosis through knowledge graph.IEEE Transactions on Services Computing16, 5 (2023), 3493–3507

    2023

  41. [42]

    Jinrui Sun, Tong Jia, Minghua He, and Ying Li. 2026. VarParser: Unleashing the Neglected Power of Variables for LLM-based Log Parsing. InProceedings of the ACM Web Conference 2026. 6965–6976

    2026

  42. [43]

    Thomas Reidemeister, Mohammad A Munawar, and Paul AS Ward. 2010. Identifying symptoms of recurrent faults in log files of distributed information systems. In2010 IEEE Network Operations and Management Symposium-NOMS

    2010

  43. [44]

    Jiabo Wang, Guojun Chu, Jingyu Wang, Haifeng Sun, Qi Qi, Yuanyi Wang, Ji Qi, and Jianxin Liao. 2024. Logexpert: Log-based recommended resolutions generation using large language model. InProceedings of the 2024 ACM/IEEE 44th International Conference on Software Engineering: New Ideas and Emerging Results. 42–46

    2024

  44. [45]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2023. Improving Text Embeddings with Large Language Models.arXiv preprint arXiv:2401.00368(2023). Bifrost: Empowering Pretrained Language Model with Fallibility Representation for Log-Based Fault Diagnosis ASE ’26, October 12–16, 2026, Munich, Germany

    Pith/arXiv arXiv 2023

  45. [46]

    Lingzhi Wang, Nengwen Zhao, Junjie Chen, Pinnong Li, Wenchi Zhang, and Kaixin Sui. 2020. Root-cause metric location for microservice systems via log anomaly de- tection. In2020 IEEE international conference on web services (ICWS). IEEE, 142–150

    2020

  46. [47]

    Jinrui Sun, Tong Jia, Minghua He, Yihan Wu, Ying Li, and Gang Huang. 2025. Exploring Variable Potential for LLM-based Log Parsing Efficiency and Reduced Costs. InProceedings of the 33rd ACM International Conference on the Foundations of Software Engineering. 596–600

    2025

  47. [48]

    W Eric Wong, Ruizhi Gao, Yihao Li, Rui Abreu, and Franz Wotawa. 2016. A survey on software fault localization.IEEE Transactions on Software Engineering42, 8 (2016), 707–740

    2016

  48. [49]

    Tianyuan Wu, Wei Wang, Yinghao Yu, Siran Yang, Wenchao Wu, Qinkai Duan, Guodong Yang, Jiamang Wang, Lin Qu, and Liping Zhang. 2024. FALCON: Pinpointing and mitigating stragglers for large-scale hybrid-parallel training. arXiv preprint arXiv:2410.12588(2024)

    Pith/arXiv arXiv 2024

  49. [50]

    Xingfang Wu, Heng Li, and Foutse Khomh. 2023. On the effectiveness of log representation for log-based anomaly detection.Empirical Software Engineering 28, 6 (2023), 137

    2023

  50. [51]

    Thorsten Wittkopp, Philipp Wiesner, and Odej Kao. 2024. LogRCA: Log-Based Root Cause Analysis for Distributed Services. InEuropean Conference on Parallel Processing. Springer, 362–376

    2024

  51. [52]

    Pei Xiao, Chiming Duan, Minghua He, Tong Jia, Yifan Wu, Jing Xu, Gege Gao, Lingzhe Zhang, Weijie Hong, Ying Li, et al. 2025. CoorLog: Efficient-Generalizable Log Anomaly Detection via Adaptive Coordinator in Software Evolution. In2025 40th IEEE/ACM International Conference on Automated Software Engineering (ASE). IEEE, 1119–1131

    2025

  52. [53]

    Pei Xiao, Tong Jia, Chiming Duan, Minghua He, Weijie Hong, Xixuan Yang, Yihan Wu, Ying Li, and Gang Huang. 2025. Clslog: Collaborating large and small models for log-based anomaly detection. InProceedings of the 33rd ACM International Conference on the Foundations of Software Engineering. 686–690

    2025

  53. [54]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

    Pith/arXiv arXiv 2025

  54. [55]

    Wensheng Xia, Ying Li, Tong Jia, and Zhonghai Wu. 2019. Bugidentifier: An approach to identifying bugs via log mining for accelerating bug reporting stage. In2019 IEEE 19th International Conference on Software Quality, Reliability and Security (QRS). IEEE, 167–175

    2019

  55. [56]

    Qingyang Yu, Changhua Pei, Bowen Hao, Mingjie Li, Zeyan Li, Shenglin Zhang, Xianglin Lu, Rui Wang, Jiaqi Li, Zhenyu Wu, et al . 2023. Cmdiagnostor: An ambiguity-aware root cause localization approach based on call metric data. In Proceedings of the ACM web conference 2023. 2937–2947

    2023

  56. [57]

    Dun Zhang, Jiacheng Li, Ziyang Zeng, and Fulong Wang. 2025. Jasper and Stella: distillation of SOTA embedding models. arXiv:2412.19048 [cs.IR] https://arxiv.org/abs/2412.19048

    Pith/arXiv arXiv 2025

  57. [58]

    Lingzhe Zhang, Liancheng Fang, Chiming Duan, Minghua He, Leyi Pan, Pei Xiao, Shiyu Huang, Yunpeng Zhai, Xuming Hu, Philip S Yu, et al. 2025. A Survey on Parallel Text Generation: From Parallel Decoding to Diffusion Language Models. arXiv preprint arXiv:2508.08712(2025)

    arXiv 2025

  58. [59]

    Lin Yang, Junjie Chen, Zan Wang, Weijing Wang, Jiajun Jiang, Xuyuan Dong, and Wenbin Zhang. 2021. Semi-supervised log-based anomaly detection via probabilistic label estimation. In2021 IEEE/ACM 43rd International Conference on Software Engineering (ICSE). IEEE, 1448–1460

    2021

  59. [60]

    Lingzhe Zhang, Tong Jia, Kangjin Wang, Chiming Duan, Minghua He, Rongqian Wang, Xi Peng, Meiling Wang, Gong Zhang, Renhai Chen, et al. 2026. Towards In-Depth Root Cause Localization for Microservices with Multi-Agent Recursion- of-Thought.IEEE Transactions on Dependable and Secure Computing(2026)

    2026

  60. [61]

    Lingzhe Zhang, Tong Jia, Kangjin Wang, Weijie Hong, Chiming Duan, Minghua He, and Ying Li. 2025. Adaptive root cause localization for microservice systems with multi-agent recursion-of-thought.arXiv preprint arXiv:2508.20370(2025)

    Pith/arXiv arXiv 2025

  61. [62]

    Lingzhe Zhang, Tong Jia, Mingyu Wang, Weijie Hong, Chiming Duan, Minghua He, Rongqian Wang, Xi Peng, Meiling Wang, Nicholas Zhang, et al. 2026. Efficient failure management for multi-agent systems with reasoning trace representation. InProceedings of the 34th ACM International Conference on the Foundations of Software Engineering. 1222–1226

    2026

  62. [63]

    Lingzhe Zhang, Tong Jia, Weijie Hong, Mingyu Wang, Chiming Duan, Minghua He, Rongqian Wang, Xi Peng, Meiling Wang, Nicholas Zhang, et al . 2026. Runtimeslicer: Towards generalizable unified runtime state representation for failure management. InProceedings of the 34th ACM International Conference on the Foundations of Software Engineering. 1217–1221

    2026

  63. [64]

    Lingzhe Zhang, Tong Jia, Yunpeng Zhai, Leyi Pan, Chiming Duan, Minghua He, Pei Xiao, and Ying Li. 2026. Hypothesize-then-verify: Speculative root cause analysis for microservices with pathwise parallelism. InProceedings of the IEEE/ACM 48th International Conference on Software Engineering. 21–25

    2026

  64. [65]

    Lingzhe Zhang, Yunpeng Zhai, Tong Jia, Minghua He, Chiming Duan, Zhaoyang Liu, Bolin Ding, and Ying Li. 2026. E2e-reme: Towards end-to-end microservices auto-remediation via experience-simulation reinforcement fine-tuning. In Proceedings of the 34th ACM International Conference on the Foundations of Software Engineering. 322–333

    2026

  65. [66]

    Shenglin Zhang, Sibo Xia, Wenzhao Fan, Binpeng Shi, Xiao Xiong, Zhenyu Zhong, Minghua Ma, Yongqian Sun, and Dan Pei. 2024. Failure Diagnosis in Microservice Systems: A Comprehensive Survey and Analysis.arXiv preprint arXiv:2407.01710 (2024)

    Pith/arXiv arXiv 2024

  66. [67]

    Lingzhe Zhang, Tong Jia, Yunpeng Zhai, Leyi Pan, Chiming Duan, Minghua He, Mengxi Jia, and Ying Li. 2026. Agentic memory enhanced recursive reasoning for root cause localization in microservices. InProceedings of the IEEE/ACM 48th International Conference on Software Engineering: Software Engineering in Practice. 58–69

    2026

  67. [68]

    Xu Zhang, Yong Xu, Si Qin, Shilin He, Bo Qiao, Ze Li, Hongyu Zhang, Xukun Li, Yingnong Dang, Qingwei Lin, et al. 2021. Onion: identifying incident-indicating logs for cloud systems. InProceedings of the 29th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering. 1253–1263

    2021

  68. [69]

    Yingying Zhang, Zhengxiong Guan, Huajie Qian, Leili Xu, Hengbo Liu, Qingsong Wen, Liang Sun, Junwei Jiang, Lunting Fan, and Min Ke. 2021. CloudRCA: A root cause analysis framework for cloud computing platforms. InProceedings of the 30th ACM International Conference on Information & Knowledge Management. 4373–4382

    2021

  69. [70]

    Xinlong Zhao, Tong Jia, Minghua He, and Ying Li. 2026. Generality Is Not Enough: Zero-Label Cross-System Log-Based Anomaly Detection via Knowledge-Level Collaboration. InProceedings of the IEEE/ACM 48th International Conference on Software Engineering. 61–65

    2026

  70. [71]

    Xu Zhang, Yong Xu, Qingwei Lin, Bo Qiao, Hongyu Zhang, Yingnong Dang, Chunyu Xie, Xinsheng Yang, Qian Cheng, Ze Li, et al. 2019. Robust log-based anomaly detection on unstable log data. InProceedings of the 2019 27th ACM joint meeting on European software engineering conference and symposium on the foundations of software engineering. 807–817

    2019

  71. [72]

    Jieming Zhu, Shilin He, Pinjia He, Jinyang Liu, and Michael R Lyu. 2023. Loghub: A large collection of system log datasets for ai-driven log analytics. In2023 IEEE 34th International Symposium on Software Reliability Engineering (ISSRE). IEEE, 355–366

    2023

  72. [75]

    Xinlong Zhao, Tong Jia, Minghua He, Ying Li, and Gang Huang. 2025. ZeroLog: Zero-Label Generalizable Cross-System Log-based Anomaly Detection. In2025 IEEE 36th International Symposium on Software Reliability Engineering (ISSRE). IEEE, 550–561

    2025

  73. [2011]

    Efficient fault detection and diagnosis in complex software systems with information-theoretic monitoring.IEEE Transactions on Dependable and Secure Computing8, 4 (2011), 510–522

    2011

  74. [2024]

    In2024 IEEE 35th International Symposium on Software Reliability Engineering (ISSRE)

    LLMeLog: An Approach for Anomaly Detection based on LLM-enriched Log Events. In2024 IEEE 35th International Symposium on Software Reliability Engineering (ISSRE). IEEE, 132–143

  75. [2025]

    In2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE)

    Weakly-supervised Log-based Anomaly Detection with Inexact Labels via Multi-instance Learning. In2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE). IEEE Computer Society, 726–726