Pith. sign in

REVIEW 3 major objections 4 minor 57 references

On Sequence-to-Sequence Models for Automated Log Parsing

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A controlled study of 396 sequence-to-sequence models claims that Transformer and Mamba outperform bidirectional and monodirectional LSTMs for automated log parsing, with Mamba near-Transformer accuracy at a fraction of the compute cost.

desk verdict A serious, compute-heavy empirical study whose headline architecture ranking is not reproducible from its own Table 13 — tokenization is confounded with architecture, and the abstract's ordering doesn't match any defined aggregation. read the letter →

arxiv 2602.07698 v2 pith:JL24JK2R submitted 2026-02-07 cs.SE cs.CL

Adam Sorrenti , Andriy Miranskyy This is my paper
classification cs.SEcs.CL
keywords logparsingsequence-to-sequencemodelsTransformerMambastatespaceLSTMLevenshteineditdistancetokenization
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

This paper tries to establish which sequence-modeling architecture should be used for automated log parsing, and under what settings. By training 396 encoder–decoder models on a synthetic Apache log corpus with controlled format difficulty, it claims that a Transformer achieves the lowest mean relative edit distance (0.111), Mamba is close behind (0.145) while costing far less to train and run, and both clearly outperform mono- and bi-directional LSTMs (0.186 and 0.265). It also claims that character-level tokenization generally helps, that longer Transformer input sequences buy negligible practical accuracy, and that Transformer and Mamba need less training data than recurrent models. If these claims hold, teams can choose Mamba for compute-constrained parsing and Transformer for robustness to heterogeneous formats, with character-level tokenization and short sequences as sensible defaults.

What carries the argument

The central mechanism is per-character sequence-to-sequence mapping: each character of a log record is assigned a semantic field label (IP address, timestamp, request, status, and so on), and the model is trained to reproduce that label sequence. Evaluation uses relative Levenshtein edit distance between the inverse-tokenized prediction and ground truth after truncation. The controlled sweep — four training dataset difficulties, two tokenization schemes, three sequence lengths, three training-data fractions, and four architectures — is what lets the paper separate architecture effects from representation, context-window, and data-availability effects.

What would settle it

Run the same trained models under a field-level or span-overlap metric (e.g., per-field F1 or boundary IoU). If word-level tokenization matches or beats character-level on that metric, the reported tokenization ranking—and with it part of the architecture ranking—is a metric artifact. Alternatively, evaluate Transformers on logs whose effective sequences exceed their training length; a sharp error increase would falsify the claim that longer sequences bring negligible benefit.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that automated log parsing, framed as a per-character sequence-to-sequence mapping from raw log text to semantic field labels, is better served by attention-based and state-space architectures than by recurrent models, but not uniformly so. The Transformer achieves the lowest mean relative edit distance overall (0.111), Mamba is competitive (0.145) with substantially lower computational cost, and both are more sample-efficient than LSTM baselines. The ranking is not static: Mamba dominates on structured, low-diversity training datasets, while the Transformer gains the advantage when log formats are highly heterogeneous and distribution-shift

Load-bearing premise

The load-bearing assumption is that relative Levenshtein edit distance, measured on characters after truncating predictions, faithfully reflects parsing quality; because it counts every character mismatch, it can punish word-level predictions for tiny boundary errors even when the assigned field is correct.

Editorial extensions

If this is right

  • Practitioners can deploy Mamba for production log parsing when compute, latency, or budget is tight, since it gives near-Transformer accuracy at orders-of-magnitude lower cost.
  • Transformer remains the safer choice when training and deployment log formats diverge strongly, such as field reordering, missing fields, or formatting changes.
  • Character-level tokenization is a strong default for edit-distance-driven evaluation; word-level tokenization adds out-of-vocabulary tokens and boundary-mismatch penalties.
  • Transformer input sequences can be kept near the 70th percentile of observed lengths without meaningful accuracy loss, avoiding quadratic attention costs.
  • Mamba and Transformer retain stable performance with only 10% of training data, so labelled-data collection can be reduced for these architectures.
  • Sequence length has negligible practical effect on Transformer parsing accuracy, so longer contexts are unlikely to be worth their computational cost for similar log workloads.

Reading between the lines

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

  • The character-level edit-distance metric likely inflates the apparent advantage of character-level tokenization; a field-level or span-level metric could narrow or even reverse that gap, so the tokenization ranking is metric-dependent.
  • The aggregate Transformer-versus-Mamba gap (0.111 vs 0.145) is small in absolute terms and mostly driven by hard format-shift settings; for many operational logs, cost may dominate the architecture choice.
  • Subword tokenization, listed by the authors as future work, is a natural testable extension that may combine character-level robustness with word-level efficiency under the same evaluation protocol.
  • The reported sample-efficiency advantage suggests a cheap practical recipe: train Mamba or Transformer on a few thousand labelled examples with character tokens and short sequences, before committing to larger data collection.
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 reports a large empirical comparison of four sequence-to-sequence architectures (Transformer, Mamba, mono-directional LSTM, and bi-directional LSTM) for automated log parsing on the HTTPd-parse benchmark. The authors train 396 models, varying training dataset difficulty, tokenization, sequence length, and training-data fraction, and evaluate with relative Levenshtein edit distance plus Wilcoxon tests with Benjamini–Hochberg correction and effect sizes. The headline claims are that Transformer achieves the lowest mean relative edit distance (0.111), followed by Mamba (0.145), mono-LSTM (0.186), and bi-LSTM (0.265); that character tokenization generally helps; that sequence length has negligible practical effect; and that Mamba offers competitive accuracy at much lower computational cost.

Significance. If the empirical conclusions held, the paper would be a useful practical guide for selecting sequence models for log parsing, and it would provide one of the first systematic evaluations of Mamba in this application. The study has notable strengths: a large, systematically varied experimental grid (396 trained models); careful attention to statistical testing with effect sizes and multiple-comparison correction; detailed appendices with per-configuration results; and explicit cost profiling. These features make the underlying experimental framework a valuable contribution. However, the central architecture ranking is currently undermined by a tokenization confound in RQ4 and by aggregate numbers that are not reproducible from the reported tables. The paper is therefore not acceptable in its present form, but the issues are localizable and fixable with a re-analysis or substantially revised claims.

major comments (3)
  1. [Abstract; Section 4.4, Table 13] The abstract's aggregate mean DR values (MT 0.111, MM 0.145, ML 0.186, MB 0.265) cannot be reproduced from Table 13, which is the table that corresponds to RQ4. Equal weighting across TT, TE, TM, and TH gives MM≈0.085, MT≈0.163, ML≈0.185, MB≈0.22 — a different ordering in which Mamba, not Transformer, is best. No aggregation rule is stated anywhere in the paper. The headline ranking is therefore unsupported as written. The authors must either specify and apply a transparent aggregation rule, or report the RQ4 results per training/validation dataset and derive the abstract claim from those numbers.
  2. [Section 4.4; Section 3.2.1; Section 4.2] RQ4 is not a controlled architecture comparison. The text states that ML, MB, and MM are evaluated with character-level tokenization while MT is evaluated with word-level tokenization at sequence length 256. Since Section 4.2 demonstrates that tokenization has a large effect on DR (e.g., Table 9 shows MM char mean 0.08 vs MM word mean 0.15, and multi-fold differences for ML/MB), architecture and tokenization are completely confounded in the abstract's ranking. The same issue affects RQ3 (Section 4.3), where MT uses word tokens while the recurrent/Mamba models use character tokens. Please run a matched design (e.g., MT with character tokens, or all models with both tokenizations) or explicitly reframe the claims as comparisons of recommended configurations, not architectures.
  3. [Section 4.2.5; Abstract; Conclusion] The paper acknowledges in Section 4.2.5 that character-level edit distance interacts with tokenization, structurally favoring character-level tokenization because word-boundary mismatches incur multiple character edits even when the field-level prediction is semantically correct. This caveat is not adequately reflected in the abstract's statement that 'character-level tokenization generally improves performance.' The RQ2 result is therefore not established as a general property of tokenization; it may be an artifact of the evaluation metric. The authors should either add a field- or span-level metric, or substantially temper the practical guidance drawn from RQ2.
minor comments (4)
  1. [Section 3.2.1] The sequence lengths are introduced as '256, 765, and 4188' but all later tables and text use 756. Please correct the typo.
  2. [Section 4.5.1 vs Section 3.2.1] The threats-to-validity section says all stochastic procedures were repeated with 'ten different random seeds,' while Section 3.2.1 states partial training datasets were sampled 'five times with different seeds.' Please align these statements.
  3. [Section 3.4] The cost comparison in Table 7 mixes frameworks: LSTM and Transformer use TensorFlow/Keras, while Mamba uses PyTorch. Since computational cost is a key practical claim, please clarify whether the wall-clock measurements are comparable across frameworks and batch configurations, or add a caveat.
  4. [Conclusion / reproducibility] The paper says the experimental framework is reproducible, but I did not find an artifact or code/data availability statement. If code and data are available, please cite the repository; if not, please add an explicit availability section.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity: the load-bearing results are empirical measurements, with at most a minor non-load-bearing self-citation for benchmark and baseline; the metric/tokenization interaction and aggregation mismatch are correctness risks, not circularity.

full rationale

The paper's central claims are direct empirical measurements of relative edit distance across trained models. No parameter is fitted to the reported outcomes and then renamed as a prediction, and no claimed result is defined into existence by a self-referential equation. The closest candidate is the RQ2 comparison of character vs word tokenization under a character-level metric. Section 4.2.5 explicitly acknowledges this interaction: 'edit distance operates at the character level, which introduces an inherent interaction between the evaluation metric and the representation choice,' and cautions that 'the observed superiority of character-based tokenization should be interpreted in the context of this metric choice.' That is an honest construct-validity caveat, not a hidden circular step. The RQ4 design also evaluates ML/MB/MM with character tokenization and MT with word tokenization, which is a confounding of architecture and representation; however, this is a controlled-comparison design choice and reproducibility concern, not an equivalence-by-construction. The abstract's aggregate DR values (0.111, 0.145, 0.186, 0.265) do not obviously follow from Table 13 under equal weighting; this is an internal-consistency/aggregation problem, not circularity. The prior work [26] supplies the HTTPd-parse benchmark, the LSTM baseline implementation, and the Levenshtein metric; because one author overlaps, this is a self-citation. It is load-bearing in the sense that the experiments use that infrastructure, but it does not determine the new Transformer/Mamba/LSTM comparisons, which are newly trained and reported here. Thus the paper contains no material circular step; the self-citation is minor and non-load-bearing, consistent with score 2.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities or fitted constants; its central claims are empirical comparisons. The main carried assumptions are the synthetic benchmark, the edit-distance metric, and the chosen hyperparameters, several inherited from prior work by one of the authors.

free parameters (6)
  • LSTM hidden units / dropout = 512 cells, dropout 0.2
    Chosen as best-performing from prior work [26]; sets recurrent baseline capacity (Section 3.1).
  • Transformer dimensions = embedding 256, FF 2048, 8 heads, dropout 0.2
    Taken from the original Transformer [24]; not tuned for log parsing (Section 3.1).
  • Mamba embedding and defaults = embedding 128; state expansion 16, conv width 4, block expansion 2
    Uses mamba_ssm v1.2.0.post1 default parameters with embedding 128 (Section 3.1).
  • Vocabulary size cap = 15,000 tokens
    Used for one-hot encoding; affects word-tokenization OOV rates (Section 3.2.1).
  • MT sequence-length thresholds = 256, 765 (also written 756), 4188 tokens
    Chosen to approximate 70th/99th/100th percentiles of validation log lengths; defines RQ1 treatment levels (Section 3.2.1).
  • Training protocol = 300 epochs, batch 64, Adam lr 1e-3, betas 0.9/0.999
    Fixed protocol for all models; influences convergence, stability, and the sample-efficiency conclusions (Section 3.3).
assumptions (4)
  • domain assumption HTTPd-parse synthetic datasets with ELF/CLF/random compositions represent realistic Apache log parsing workloads.
    All training and validation data are synthetically generated using the heuristic from [26]; validation sets are 100% ELF, which is the basis for distribution-shift claims (Section 3.2, Table 6).
  • domain assumption Relative Levenshtein edit distance DR after truncation and inverse tokenization accurately measures parsing quality.
    Primary metric; authors note it operates at character level and can bias tokenization comparisons (Sections 2.2.1, 4.2.5).
  • standard math Pairwise Wilcoxon signed-rank tests with Benjamini-Hochberg correction are valid for the paired model-output comparisons.
    Used for all significance tests; assumes paired observations and exchangeability under the null (Section 4.1.2).
  • domain assumption mamba_ssm v1.2.0.post1 implements Mamba as described in [30].
    Mamba results depend on correctness of the external PyTorch implementation (Sections 3.1, 3.4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of On Sequence-to-Sequence Models for Automated Log Parsing." pith.science (2026). https://pith.science/paper/JL24JK2R

@misc{pith2026260207698,
  author       = {Pith},
  title        = {Pith review of: On Sequence-to-Sequence Models for Automated Log Parsing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JL24JK2R}},
  note         = {Machine review of arXiv:2602.07698}
}
read the original abstract

Context: Log parsing is a critical standard operating procedure in software systems, enabling monitoring, anomaly detection, and failure diagnosis. However, automated log parsing remains challenging due to heterogeneous log formats, distribution shifts between training and deployment data, and the brittleness of rule-based approaches. Objectives: This study aims to systematically evaluate how sequence modelling architecture, representation choice, sequence length, and training data availability influence automated log parsing performance and computational cost. Methods: We conduct a controlled empirical study comparing four sequence modelling architectures: Transformer, Mamba state-space, monodirectional LSTM, and bidirectional LSTM models. In total, 396 models are trained across multiple dataset configurations and evaluated using relative Levenshtein edit distance with statistical significance testing. Results: Transformer achieves the lowest mean relative edit distance (0.111), followed by Mamba (0.145), mono-LSTM (0.186), and bi-LSTM (0.265), where lower values are better. Mamba provides competitive accuracy with substantially lower computational cost. Character-level tokenization generally improves performance, sequence length has negligible practical impact on Transformer accuracy, and both Mamba and Transformer demonstrate stronger sample efficiency than recurrent models. Conclusion: Overall, Transformers reduce parsing error by 23.4%, while Mamba is a strong alternative under data or compute constraints. These results also clarify the roles of representation choice, sequence length, and sample efficiency, providing practical guidance for researchers and practitioners.

Figures

Figures reproduced from arXiv: 2602.07698 by the authors.

Figure 1
Figure 1. A sample sequence-to-sequence mapping of an Apache web server log. The first line contains the raw log message. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Box-plots of the relative edit distance of [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Box-plots of the relative edit distance by tokenization method and training dataset for each model architecture. [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 4 canonical work pages

  1. [1]

    D. Yuan, H. Mai, W. Xiong, L. Tan, Y. Zhou, S. Pasupathy, Sherlog: error diagnosis by connecting clues from run-time logs, in: J. C. Hoe, V. S. Adve (Eds.), Proceedings of the 15th International Conference on Architectural Support for Programming Languages and Operating Systems, ASPLOS 2010, Pittsburgh, Pennsylvania, USA, March 13-17, 2010, ACM, 2010, pp....

    arXiv 2010

  2. [2]

    S. Lu, B. Rao, X. Wei, B. Tak, L. Wang, L. Wang, Log-based abnormal task detection and root cause analysis for spark, in: I. Altintas, S. Chen (Eds.), 2017 IEEE International Conference on Web Services, ICWS 2017, Honolulu, HI, USA, June 25-30, 2017, IEEE, 2017, pp. 389–396.doi:10.1109/ICWS.201 7.135. URLhttps://doi.org/10.1109/ICWS.2017.135

    work page doi:10.1109/icws.201 2017

  3. [3]

    A.Das, F.Mueller, C.Siegel, A.Vishnu, Desh: deeplearningforsystemhealthpredictionofleadtimesto failure in HPC, in: M. Zhao, A. Chandra, L. Ramakrishnan (Eds.), Proceedings of the 27th International Symposium on High-Performance Parallel and Distributed Computing, HPDC 2018, Tempe, AZ, USA, June 11-15, 2018, ACM, 2018, pp. 40–51.doi:10.1145/3208040.3208051...

    arXiv 2018

  4. [4]

    G. Lee, J. Lin, C. Liu, A. Lorek, D. V. Ryaboy, The unified logging infrastructure for data analytics at twitter, Proc. VLDB Endow. 5 (12) (2012) 1771–1780.doi:10.14778/2367502.2367516. URLhttp://vldb.org/pvldb/vol5/p1771_georgelee_vldb2012.pdf

    arXiv 2012

  5. [5]

    M. Du, F. Li, G. Zheng, V. Srikumar, Deeplog: Anomaly detection and diagnosis from system logs through deep learning, in: B. Thuraisingham, D. Evans, T. Malkin, D. Xu (Eds.), Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, CCS 2017, Dallas, TX, USA, October 30 - November 03, 2017, ACM, 2017, pp. 1285–1298.doi:10.1145...

    arXiv 2017

  6. [6]

    Zhang, Y

    X. Zhang, Y. Xu, Q. Lin, B. Qiao, H. Zhang, Y. Dang, C. Xie, X. Yang, Q. Cheng, Z. Li, J. Chen, X. He, R. Yao, J. Lou, M. Chintalapati, F. Shen, D. Zhang, Robust log-based anomaly detection on unstable log data, in: M. Dumas, D. Pfahl, S. Apel, A. Russo (Eds.), Proceedings of the ACM Joint Meeting on European Software Engineering Conference and Symposium ...

    arXiv 2019

  7. [7]

    Zhang, H

    B. Zhang, H. Zhang, P. Moscato, A. Zhang, Anomaly detection via mining numerical workflow relations from logs, in: International Symposium on Reliable Distributed Systems, SRDS 2020, Shanghai, China, September 21-24, 2020, IEEE, 2020, pp. 195–204.doi:10.1109/SRDS51746.2020.00027. URLhttps://doi.org/10.1109/SRDS51746.2020.00027

    arXiv 2020

  8. [8]

    M. S. Islam, W. Pourmajidi, L. Zhang, J. Steinbacher, T. Erwin, A. V. Miranskyy, Anomaly detection in a large-scale cloud platform, in: 43rd IEEE/ACM International Conference on Software Engineering: Software Engineering in Practice, ICSE (SEIP) 2021, Madrid, Spain, May 25-28, 2021, IEEE, 2021, pp. 150–159.doi:10.1109/ICSE-SEIP52600.2021.00024. URLhttps:/...

    arXiv 2021

Show all 57 references
  1. [9]

    M. S. Islam, M. S. Rakha, W. Pourmajidi, J. Sivaloganathan, J. Steinbacher, A. V. Miranskyy, Anomaly detection in large-scale cloud systems: An industry case and dataset, in: 47th IEEE/ACM International Conference on Software Engineering: Software Engineering in Practice, SEIP...

    2025

  2. [10]

    H. Mi, H. Wang, Y. Zhou, M. R. Lyu, H. Cai, Toward fine-grained, unsupervised, scalable performance diagnosis for production cloud computing systems, IEEE Trans. Parallel Distributed Syst. 24 (6) (2013) 1245–1255.doi:10.1109/TPDS.2013.21. URLhttps://doi.org/10.1109/TPDS.2013.21

    2013 doi

  3. [11]

    A. V. Miranskyy, A. Hamou-Lhadj, E. Cialini, A. Larsson, Operational-log analysis for big data systems: Challenges and solutions, IEEE Softw. 33 (2) (2016) 52–59.doi:10.1109/MS.2016.33. URLhttps://doi.org/10.1109/MS.2016.33

    2016 doi

  4. [12]

    Pourmajidi, J

    W. Pourmajidi, J. Steinbacher, T. Erwin, A. Miranskyy, On challenges of cloud monitoring, in: Pro- ceedings of the 27th Annual International Conference on Computer Science and Software Engineering, 2017, pp. 259–265

    2017

  5. [13]

    Pourmajidi, A

    W. Pourmajidi, A. Miranskyy, J. Steinbacher, T. Erwin, D. Godwin, Dogfooding: Using ibm cloud ser- vices to monitor ibm cloud infrastructure, in: Proceedings of the 29th Annual International Conference on Computer Science and Software Engineering, 2019, pp. 344–353

    2019

  6. [14]

    Pourmajidi, L

    W. Pourmajidi, L. Zhang, A. Miranskyy, J. Steinbacher, D. Godwin, T. Erwin, The challenging land- scape of cloud monitoring, in: Knowledge Management in the Development of Data-Intensive Systems, CRC Press, 2021, pp. 157–189

    2021

  7. [15]

    33–40.doi:10.1109/ICWS.2017.13

    P.He, J.Zhu, Z.Zheng, M.R.Lyu, Drain: Anonlinelogparsingapproachwithfixeddepthtree, in: 2017 IEEE International Conference on Web Services (ICWS), 2017, pp. 33–40.doi:10.1109/ICWS.2017.13

    2017 doi

  8. [16]

    Z. M. Jiang, A. E. Hassan, P. Flora, G. Hamann, Abstracting execution logs to execution events for enterprise applications (short paper), in: 2008 The Eighth International Conference on Quality Software, 2008, pp. 181–186.doi:10.1109/QSIC.2008.50

    2008 doi

  9. [17]

    H. Dai, H. Li, C. Chen, W. Shang, T. Chen, Logram: Efficient log parsing using $n$n-gram dictionaries, IEEE Trans. Software Eng. 48 (3) (2022) 879–892.doi:10.1109/TSE.2020.3007554. URLhttps://doi.org/10.1109/TSE.2020.3007554

    2022

  10. [18]

    Vaarandi, M

    R. Vaarandi, M. Pihelgas, Logcluster - A data clustering and pattern mining algorithm for event logs, in: M. Tortonesi, J. Schönwälder, E. R. M. Madeira, C. Schmitt, J. Serrat (Eds.), 11th International Conference on Network and Service Management, CNSM 2015, Barcelona, Spain,...

    2015

  11. [19]

    L. Tang, T. Li, C.-S. Perng, Logsig: generating system events from raw textual logs, in: Proceedings of the 20th ACM International Conference on Information and Knowledge Management, CIKM ’11, Association for Computing Machinery, New York, NY, USA, 2011, p. 785–794.doi:10.1145...

    2011

  12. [20]

    M. Du, F. Li, Spell: Online streaming parsing of large unstructured system logs, IEEE Transactions on Knowledge and Data Engineering 31 (11) (2019) 2213–2227.doi:10.1109/TKDE.2018.2875442

    2019

  13. [21]

    H. Guo, S. Yuan, X. Wu, Logbert: Log anomaly detection via bert, in: 2021 International Joint Conference on Neural Networks (IJCNN), 2021, pp. 1–8.doi:10.1109/IJCNN52387.2021.9534113

    2021

  14. [22]

    Nedelkoski, J

    S. Nedelkoski, J. Bogatinovski, A. Acker, J. Cardoso, O. Kao, Self-supervised log parsing, in: Y. Dong, D. Mladenić, C. Saunders (Eds.), Machine Learning and Knowledge Discovery in Databases: Applied Data Science Track, Springer International Publishing, Cham, 2021, pp. 122–138. 20

    2021

  15. [23]

    Z. Ma, A. R. Chen, D. J. Kim, T.-H. Chen, S. Wang, Llmparser: An exploratory study on using large language models for log parsing, in: Proceedings of the IEEE/ACM 46th International Conference on Software Engineering, ICSE ’24, Association for Computing Machinery, New York, NY...

    2024

  16. [24]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, I. Polosukhin, Attention is all you need, in: I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vish- wanathan, R. Garnett (Eds.), Advances in Neural Information Processing System...

    2017

  17. [26]

    J. Rand, A. Miranskyy, On automatic parsing of log records, in: 2021 IEEE/ACM 43rd International Conference on Software Engineering: New Ideas and Emerging Results (ICSE-NIER), 2021, pp. 41–45. doi:10.1109/ICSE-NIER52604.2021.00017

    2021

  18. [27]

    A. Gu, K. Goel, C. Ré, Efficiently modeling long sequences with structured state spaces, CoRR abs/2111.00396 (2021).arXiv:2111.00396. URLhttps://arxiv.org/abs/2111.00396

    2021 arXiv

  19. [28]

    M. Beck, K. Pöppel, M. Spanring, A. Auer, O. Prudnikova, M. Kopp, G. Klambauer, J. Brandstetter, S. Hochreiter, xlstm: Extended long short-term memory, in: A. Globersons, L. Mackey, D. Belgrave, A.Fan, U.Paquet, J.M.Tomczak, C.Zhang(Eds.), AdvancesinNeuralInformationProcessing...

    2024

  20. [29]

    Danieli, P

    F. Danieli, P. Rodriguez, M. Sarabia, X. Suau, L. Zappella, Pararnn: Unlocking parallel training of nonlinear rnns for large language models (2025).arXiv:2510.21450. URLhttps://arxiv.org/abs/2510.21450

    2025

  21. [30]

    A. Gu, T. Dao, Mamba: Linear-time sequence modeling with selective state spaces, CoRR abs/2312.00752 (2023).arXiv:2312.00752,doi:10.48550/ARXIV.2312.00752. URLhttps://doi.org/10.48550/arXiv.2312.00752

    2023 doi arXiv

  22. [31]

    Pecchia, M

    A. Pecchia, M. Cinque, G. Carrozza, D. Cotroneo, Industry practices and event logging: Assessment of a critical software development process, in: 2015 IEEE/ACM 37th IEEE International Conference on Software Engineering, Vol. 2, 2015, pp. 169–178.doi:10.1109/ICSE.2015.145

    2015 doi

  23. [32]

    J. Zhu, S. He, J. Liu, P. He, Q. Xie, Z. Zheng, M. R. Lyu, Tools and benchmarks for automated log parsing, in: 2019 IEEE/ACM 41st International Conference on Software Engineering: Software Engineering in Practice (ICSE-SEIP), 2019, pp. 121–130.doi:10.1109/ICSE-SEIP.2019.00021

    2019

  24. [33]

    URLhttps://httpd.apache.org/docs/current/mod/mod_log_config.html

    Apache HTTP Server Project, Apache modulemod_log_config, available athttps://httpd.apache .org/docs/current/mod/mod_log_config.html(2025). URLhttps://httpd.apache.org/docs/current/mod/mod_log_config.html

    2025

  25. [34]

    J. Zhu, S. He, P. He, J. Liu, M. R. Lyu, Loghub: A large collection of system log datasets for ai-driven log analytics, in: 2023 IEEE 34th International Symposium on Software Reliability Engineering (ISSRE), 2023, pp. 355–366.doi:10.1109/ISSRE59848.2023.00071. 21

    2023

  26. [35]

    Messaoudi, A

    S. Messaoudi, A. Panichella, D. Bianculli, L. Briand, R. Sasnauskas, A search-based approach for accurate identification of log message formats, in: Proceedings of the 26th Conference on Program Comprehension, 2018, pp. 167–177

    2018

  27. [36]

    K. Cho, B. van Merrienboer, Ç. Gülçehre, D. Bahdanau, F. Bougares, H. Schwenk, Y. Bengio, Learning phrase representations using RNN encoder-decoder for statistical machine translation, in: A. Moschitti, B. Pang, W. Daelemans (Eds.), Proceedings of the 2014 Conference on Empiri...

    2014 doi

  28. [37]

    Sutskever, O

    I. Sutskever, O. Vinyals, Q. V. Le, Sequence to sequence learning with neural networks (2014).arXiv: 1409.3215. URLhttps://arxiv.org/abs/1409.3215

    2014 arXiv

  29. [38]

    Hochreiter, J

    S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Comput. 9 (8) (1997) 1735–1780. doi:10.1162/NECO.1997.9.8.1735. URLhttps://doi.org/10.1162/neco.1997.9.8.1735

    1997 doi

  30. [39]

    Hochreiter, Y

    S. Hochreiter, Y. Bengio, P. Frasconi, J. Schmidhuber, et al., Gradient flow in recurrent nets: the difficulty of learning long-term dependencies (2001)

    2001

  31. [40]

    Bahdanau, K

    D. Bahdanau, K. Cho, Y. Bengio, Neural machine translation by jointly learning to align and translate, in: Y. Bengio, Y. LeCun (Eds.), 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. UR...

    2015 arXiv

  32. [41]

    Y. Kim, C. Denton, L. Hoang, A. M. Rush, Structured attention networks, in: 5th International Con- ference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings, OpenReview.net, 2017. URLhttps://openreview.net/forum?id=HkE0Nvqlg

    2017

  33. [42]

    Y. Tay, M. Dehghani, D. Bahri, D. Metzler, Efficient transformers: A survey, ACM Comput. Surv. 55 (6) (2023) 109:1–109:28.doi:10.1145/3530811. URLhttps://doi.org/10.1145/3530811

    2023 doi

  34. [43]

    A. Gu, I. Johnson, K. Goel, K. Saab, T. Dao, A. Rudra, C. Ré, Combining recurrent, convolutional, and continuous-time models with linear state-space layers, CoRR abs/2110.13985 (2021).arXiv:2110 .13985. URLhttps://arxiv.org/abs/2110.13985

    2021 arXiv

  35. [44]

    A. Gu, T. Dao, mamba-ssm (ver. 1.2.0.post1),https://pypi.org/project/mamba-ssm/1.2.0.post1/ (Mar. 2024)

    2024

  36. [45]

    D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, in: Y. Bengio, Y. LeCun (Eds.), 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. URLhttp://arxiv.org/abs/1412.6980

    2015 arXiv

  37. [46]

    Abadi, A

    M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, et al., Tensorflow: Large-scale machine learning on heterogeneous distributed systems, arXiv preprint arXiv:1603.04467 (2016)

    2016 arXiv

  38. [47]

    Chollet, et al., Keras,https://keras.io(2015)

    F. Chollet, et al., Keras,https://keras.io(2015)

    2015

  39. [48]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al., Pytorch: An imperative style, high-performance deep learning library, Advances in neural information processing systems 32 (2019). 22

    2019

  40. [49]

    Yadan, Hydra - a framework for elegantly configuring complex applications, Github (2019)

    O. Yadan, Hydra - a framework for elegantly configuring complex applications, Github (2019). URLhttps://github.com/facebookresearch/hydra

    2019

  41. [50]

    A. B. Yoo, M. A. Jette, M. Grondona, Slurm: Simple linux utility for resource management, in: Work- shop on job scheduling strategies for parallel processing, Springer, 2003, pp. 44–60

    2003

  42. [51]

    Benjamini, Y

    Y. Benjamini, Y. Hochberg, Controlling the false discovery rate: a practical and powerful approach to multiple testing, Journal of the Royal statistical society: series B (Methodological) 57 (1) (1995) 289–300

    1995

  43. [52]

    Wohlin, P

    C. Wohlin, P. Runeson, M. Höst, M. C. Ohlsson, B. Regnell, A. Wesslén, Experimentation in Software Engineering, Springer, 2012.doi:10.1007/978-3-642-29044-2. URLhttps://doi.org/10.1007/978-3-642-29044-2

    2012 doi

  44. [53]

    R. K. Yin, Case study research: Design and methods, Vol. 5, sage, 2009

    2009

  45. [54]

    The Git Development Team, Git: Fast, distributed version control system,https://git-scm.com/ (2005)

    2005

  46. [55]

    Wieringa, M

    R. Wieringa, M. Daneva, Six strategies for generalizing software engineering theories, Science of com- puter programming 101 (2015) 136–152

    2015

  47. [56]

    Sennrich, B

    R. Sennrich, B. Haddow, A. Birch, Neural machine translation of rare words with subword units, CoRR abs/1508.07909 (2015).arXiv:1508.07909. URLhttp://arxiv.org/abs/1508.07909

    2015 arXiv

  48. [57]

    Y. Wu, M. Schuster, Z. Chen, Q. V. Le, M. Norouzi, W. Macherey, M. Krikun, Y. Cao, Q. Gao, K. Macherey, J. Klingner, A. Shah, M. Johnson, X. Liu, L. Kaiser, S. Gouws, Y. Kato, T. Kudo, H. Kazawa, K. Stevens, G. Kurian, N. Patil, W. Wang, C. Young, J. Smith, J. Riesa, A. Rudnic...

    2016 arXiv

  49. [58]

    OPTIONS 46rpb HTTP/1.1

    T. Kudo, J. Richardson, Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing, CoRR abs/1808.06226 (2018).arXiv:1808.06226. URLhttp://arxiv.org/abs/1808.06226 23 Appendix A. Extended Evaluation Results Appendix A.1. Seque...

    2018 arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.