LINEADAPTER: Parameter-Efficient Fine-Tuning for Log Anomaly Detection and Root Cause Analysis
The DOI matches its Crossref record; 34 of the 34 references it cites were found in a registry or an index; 6 of the 15 sources its claims cite were checked for support.
Abstract
The growing scale and complexity of distributed systems such as Hadoop produce massive volumes of complex log data, making automated anomaly detection and root cause analysis both essential and increasingly challenging. Traditional rule-based approaches, relying on static thresholds or manual heuristics, struggle to scale due to limited adaptability and high maintenance costs. Transformer-based language models have emerged as powerful tools for modelling log sequences by capturing contextual patterns. To enable efficient adaptation with fewer parameters, techniques such as In-Context Learning (ICL) and Low-Rank Adaptation (LoRA) have been proposed. However, applying these methods to log analysis with Small Language Models (SLMs) introduces several challenges, including high memory and computational overhead (in the case of ICL), limited fine-tuning capacity (with LoRA on smaller models), and poor generalisation across heterogeneous environments. To address these limitations, we propose Lineadapter, a parameter-efficient fine-tuning framework tailored for SLMs in log anomaly detection and root cause analysis. Lineadapter extends convolutional adapter principles to sequential data by integrating lightweight 1D convolutional layers within transformer blocks. This design enables SLMs to adapt effectively to log-specific patterns with minimal computational cost, while preserving the backbone model's representational power. We evaluate LINEADAPTER on multiple real-world system log datasets, comparing its performance with ICL, LoRA, and rule-based baselines. Results show that Lineadapter achieves higher F1-scores and better precision-recall trade-offs, particularly on medium-scale models, establishing it as a scalable, robust, and practical solution for log-based anomaly detection.
Overview
CoChat Summary
TL;DR
LINEADAPTER is a parameter-efficient fine-tuning framework for small language models that inserts a 1D convolutional bottleneck adapter after transformer feed-forward/MLP layers while keeping the backbone frozen, targeting log anomaly detection and root cause analysis. Across LogHub and Automate datasets, it improves F1 and precision-recall balance over rule-based and ICL baselines, with especially strong RCA semantic similarity for larger SLMs.
Key findings
- The rule-based baseline achieves high precision above 0.96 across all datasets but low recall of 0.34–0.40, giving F1-scores of 0.51–0.56; neural approaches (ICL, LoRA, LineAdapter) achieve significantly higher F1-scores.
- LineAdapter on Llama-3.2-1B reaches F1-scores of 0.852 on HDFS and 0.979 on BGL, while LoRA on Llama-3.2-1B shows low HDFS recall of 0.308; Bloom-3B with LoRA and LineAdapter approaches or exceeds F1 0.95.
- On Automate RCA, a linear regression detector achieves accuracy 0.997, and LineAdapter fine-tuned models reach cosine similarity scores of 0.857 for BERT-base-uncased, 0.959 for LLaMA-3.2-1B, and 0.977 for Bloom-3B.
- LineAdapter costs O(2dd′ + kd′) versus LoRA’s O(2Mdr), yielding roughly a 4× reduction in FLOPs and parameter count when M ≥ 4 and d′ is comparable to r.
Methods
- Evaluates BERT-base-uncased, LLaMA 3.2-1B, and Bloom-3B on LogHub datasets HDFS, Hadoop, and BGL, plus the Automate RCA dataset, comparing rule-based detection, ICL, LoRA, and LineAdapter.
- LogHub data are merged from original train/test partitions, binary-labeled, stratified-shuffled, with BGL and HDFS stratified 10% sampling and the third dataset retained fully; an 8:2 stratified split is used. Preprocessing removes prefixes before the first comma, replaces nonnumeric characters with spaces, collapses whitespace, deduplicates lines, and discards lines shorter than three tokens.
- LineAdapter projects hidden states to a bottleneck d′ ≪ d, applies ReLU and a depthwise 1D convolution with kernel size k = 3, restores hidden size, and adds a residual update; it is inserted after FFN/MLP layers, with only adapters and task head trained while the backbone stays frozen. RCA is evaluated by cosine similarity between generated explanations and annotated ground-truth causes.
Limitations
- Stated by the paper: all training and inference were executed on a single compute node, limiting scalability and not exploiting distributed computing environments; future work aims to integrate distributed fine-tuning.
- Evident from the text: Table II reports ICL results only for Llama-3.2-1B and Bloom-3B, with BERT-base-uncased ICL entries marked “*”; evaluation covers three LogHub datasets plus Automate and three SLMs.
- Evident from the text: RCA evaluation uses correlation-based cosine similarity because direct comparison of generative root-cause outputs against ground truth is considered unreliable due to stochasticity.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
5 Valid · 2 Not valid · 4 Not checked · 51 Own results and interpretations
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“For a transformer block with hidden size d , LineAdapter introduces two linear projections of size d × d ′ and a depthwise 1D convolution with kernel size k , leading to an additional cost of O (2 dd ′ + kd ′ ) .”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“By restricting computations to the bottleneck dimension d ′ , the number of trainable parameters is reduced from O ( d 2 ) to O (2 dd ′ )”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“First, anomaly detection is performed using a lightweight linear regression baseline, which achieves an accuracy of 0.997, demonstrating that traditional machine learning models are sufficient for reliable anomaly identification in this dataset.”
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Our evaluation highlights the trade-offs between accuracy, efficiency, and scalability across diverse log anomaly detection scenarios.”
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The results, summarised in Figure 3, show a clear improvement in explanatory power as the model size increases.”
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
17 more claims
Create a free account to see all 62 claims and to verify the 5 references still open.
- It may state two separate claims as one. 33
- Probably not a substantial claim. 13
- The quote only partly states this claim. 1
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ B. Root-cause Analysis
“0.996 | 0.974”
§ B. Root-cause Analysis
“0.974 | 0.994”
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ B. Root-cause Analysis
“0.952 | 0.900”
§ B. Root-cause Analysis
“0.900 | 1.000”
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ B. Root-cause Analysis
“0.730 | 1.000”
§ B. Root-cause Analysis
“1.000 | 0.308”
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ B. Root-cause Analysis
“0.852 | 0.939”
§ B. Root-cause Analysis
“0.939 | 0.549”
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ B. Root-cause Analysis
“| BGL | LineAdapter | 0.952”
§ B. Root-cause Analysis
“0.996 | 0.980”
2 more findings
Create a free account to see every finding of this check.
Citations
Cites
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325
LoRA: Low-Rank Adaptation of Large Language Models
Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2021) · arXiv · doi:10.48550/arxiv.2106.09685 · Cited by 13,453 · cited 2×
“To further adapt the language model to the downstream task while maintaining computational efficiency, we employ LoRA [7].”
Self-citationDetecting large-scale system problems by mining console logs
Xu, W., Huang, L., Fox, A., Patterson, D., & Jordan, M. I. (2009) · Proceedings of the ACM SIGOPS 22nd symposium on Operating systems principles · doi:10.1145/1629575.1629587 · Cited by 1,332
Drain: an online log parsing approach with fixed depth tree
He, P., Zhu, J., Zheng, Z., & Lyu, M. R. (2017) · 2017 IEEE International Conference on Web Services (ICWS) · doi:10.1109/icws.2017.13 · Cited by 943 · cited 2×
“Corresponding author: Yinhao Li (yinhao.li@newcastle.ac.uk) features [4], [5].”
A Survey on In-context Learning
Dong, Q., Li, L., Dai, D., Zheng, C., Ma, J., Li, R., Xia, H., Xu, J., Wu, Z., Chang, B., et al. (2024) · Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/2024.emnlp-main.64 · Cited by 649 · cited 2×
“In this work, we adopt an ICL [6] framework, wherein task-specific behaviour is elicited by conditioning a language model on a set of retrieved demonstration examples, without any gradient-based optimisation.”
What Supercomputers Say: A Study of Five System Logs
Oliner, A., & Stearley, J. (2007) · 37th Annual IEEE/IFIP International Conference on Dependable Systems and Networks (DSN'07) · doi:10.1109/dsn.2007.103 · Cited by 620
Tools and benchmarks for automated log parsing
Zhu, J., He, S., Liu, J., He, P., Xie, Q., Zheng, Z., & Lyu, M. R. (2019) · 2019 IEEE/ACM 41st International Conference on Software Engineering: Software Engineering in Practice (ICSE-SEIP) · doi:10.1109/icse-seip.2019.00021 · Cited by 574 · cited 2×
“Corresponding author: Yinhao Li (yinhao.li@newcastle.ac.uk) features [4], [5].”
MicroRCA: Root Cause Localization of Performance Issues in Microservices
Wu, L., Tordsson, J., Elmroth, E., & Kao, O. (2020) · NOMS 2020 - 2020 IEEE/IFIP Network Operations and Management Symposium · doi:10.1109/noms47738.2020.9110353 · Cited by 265
Sage: practical and scalable ML-driven performance debugging in microservices
Gan, Y., Liang, M., Dev, S., Lo, D., & Delimitrou, C. (2021) · Proceedings of the 26th ACM International Conference on Architectural Support for Programming Languages and Operating Systems · doi:10.1145/3445814.3446700 · Cited by 174
Research works printed but not linked: 1
- [27] Meta, 'Meta llama 3.1,' 2024, Meta AI Blog . -
Cited by
Murugadoss, R., Pranesh, M., Pragadeeshwaran, N., & Eswar, V. (2026) · International Conference on Data Science for Cyber-Physical Systems Resilience using Advanced Applications (ICDCA) · doi:10.1109/icdca69396.2026.11620031
Metadata
- Authors
- Liu, X., Bao, W., Zhang, Y., Li, Y., Ranjan, R., & Jha, D. N.
- Affiliations
- Newcastle University (GB)
- Venue
- 2025 IEEE 31th International Conference on Parallel and Distributed Systems (ICPADS)
- Publisher
- IEEE
- Published
- 2025-12-14
- Pages
- 1-8
- OpenAlex
- W7124170655
- Semantic Scholar
- d8b1cff846b6976f13d79201882316e0c77f6a69
- Cited by
- 0
- References
- 23
- FWCI
- 0.00
- Type
- conference paper
- Language
- en
- Open access
- closed
- Locations
- IEEE International Conference on Parallel and Distributed Systems (ICPADS) (publishedVersion)
- Topics
- Physical Sciences › Computer Science › Computer Networks and Communications › Software System Performance and Reliability
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Anomaly Detection Techniques and Applications
- Keywords
- log anomaly detection, root cause analysis, parameter-efficient fine-tuning, small language models, transformer language models, in-context learning, low-rank adaptation
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: LINEADAPTER: Parameter-Efficient Fine-Tuning for Log Anomaly Detection and Root Cause Analysis. https://app.cochat.ai/papers/10.1109/icpads67057.2025.11322907
- The paper
- Liu, X., Bao, W., Zhang, Y., Li, Y., Ranjan, R., & Jha, D. N. (2025). LINEADAPTER: Parameter-Efficient Fine-Tuning for Log Anomaly Detection and Root Cause Analysis. 2025 IEEE 31th International Conference on Parallel and Distributed Systems (ICPADS). https://doi.org/10.1109/icpads67057.2025.11322907
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.