Retrieval-based and Fine-tuned LLM Approaches for Industrial Asset Health Monitoring and Decision Support

Seshu Kumar Damarla; Xiuli Zhu

REVIEW 3 major objections 5 minor 18 references

Retrieval-based and Fine-tuned LLM Approaches for Industrial Asset Health Monitoring and Decision Support

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read QLoRA fine-tuning of Qwen2.5-7B-Instruct lifts failure-sensor reasoning accuracy from 55% (zero-shot) to 88%, and meaning-based retrieval outperforms keyword matching.

desk verdict Solid benchmark comparison, but the headline fine-tuning result lacks a shuffle-only control and error bars; the domain-adaptation conclusion is currently overstated. read the letter →

arxiv 2607.24824 v1 pith:AVWHBNOA submitted 2026-07-18 cs.IR

Seshu Kumar Damarla , Xiuli Zhu This is my paper

classification cs.IR

keywords FailureSensorIQQLoRAQwen2.5-7B-Instructindustrialassethealthmonitoringoption-wiseretrievalsemanticsearchfailure-modereasoningmultiple-choiceQA

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

Industrial plants rely on human engineers to connect abnormal sensor readings to failure modes, a reasoning task that is hard to automate. This paper compares two automated routes on the FailureSensorIQ multiple-choice benchmark: an option-wise retrieval method that scores each answer choice against similar training records, and the open-source Qwen2.5-7B-Instruct language model used in zero-shot, few-shot, and QLoRA fine-tuned modes. The central result is that fine-tuning lifts accuracy to 88.01%, far above the 55.43% zero-shot and 54.49% few-shot scores, and that semantic and hybrid retrieval (70.97–73.41%) outperform keyword-based TF-IDF and BM25. These findings matter because they suggest general-purpose LLMs must be adapted to the specialized vocabulary and reasoning of asset health monitoring before they can serve as decision-support tools.

What carries the argument

The retrieval-only pipeline converts every training question into option-level records tagged correct or incorrect, then scores each test option separately as the mean similarity to the top-k support records minus λ times the mean similarity to top-k contradiction records (λ=0.75). The LLM pipeline uses QLoRA, a parameter-efficient fine-tuning method that freezes the 7B base weights and trains a small set of LoRA adapters on instruction-format failure-sensor reasoning examples.

What would settle it

Conduct the same zero-shot, few-shot, and QLoRA fine-tuned evaluation over multiple random splits and seeds on FailureSensorIQ; if the fine-tuned model's average advantage over zero-shot shrinks to a few points, the claimed necessity of fine-tuning would collapse.

Watch

Extended reading notes

Core claim

The paper claims that on the FailureSensorIQ benchmark for industrial failure-sensor relationships, a QLoRA fine-tuned Qwen2.5-7B-Instruct model achieves 88.01% accuracy, far above the zero-shot (55.43%) and few-shot (54.49%) versions of the same model, and that option-wise semantic and hybrid retrieval (70.97% and 73.41%) outperform keyword methods TF-IDF (56.18%) and BM25 (59.18%). The authors argue this shows that general-purpose LLMs require domain-specific adaptation for specialized industrial reasoning, and that meaning-based similarity is more important than exact word matching for finding engineering evidence.

Load-bearing premise

The comparison assumes that the single random 80/20 split, fixed hyperparameters, and chosen decoding settings yield stable accuracy differences; if the split or settings changed, the reported gaps may not replicate.

Editorial extensions

If this is right

Meaning-based (semantic and hybrid) retrieval beats exact keyword matching, indicating semantic similarity is the stronger evidence signal for failure-sensor reasoning.
General-purpose LLMs need task-specific fine-tuning to reach usable accuracy on specialized industrial diagnostics.
Accuracy drops as the number of answer options increases, and all methods lose ground under option shuffling, paraphrasing, changed labels, and added distractors.
The fine-tuned model stays most accurate under perturbation but still degrades considerably, so high test accuracy does not guarantee stability under practical changes.

Reading between the lines

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

The reported accuracy gaps come from a single random split and fixed hyperparameters; repeated splits and seeds would clarify whether the 88%-vs-55% gap is stable or partly due to split luck.
Combining option-wise retrieval with the fine-tuned LLM could exploit retrieval's speed and interpretability alongside the LLM's accuracy, but that blend is left untested here.
Sensitivity to option shuffling and label changes suggests that fixed-order test accuracy may overstate real-world reliability, making perturbation-based evaluation a necessary complement to accuracy reporting.
The hybrid search used a very low TF-IDF weight (α=0.1), so its performance is almost purely semantic; this balance may be dataset-specific and worth tuning per asset type.

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.

Desk Editor's Note

read the letter

The paper is a straightforward empirical comparison on the FailureSensorIQ benchmark. It does not introduce a new architecture, but the option-wise retrieval scoring — where each answer option gets support and contradiction evidence from training records — is a practical variant on retrieval-based QA, and the direct comparison of TF-IDF, BM25, semantic, and hybrid retrieval is useful for practitioners. The fine-tuned Qwen2.5-7B result (88%) is plausible, and the tables are internally consistent. The error analysis by question type and number of options is thorough, and the decision to include robustness checks is the right instinct.

The soft spots are real and one is load-bearing. There are no confidence intervals, no multiple seeds, and a single 80/20 split, so we have no idea whether the 88 vs 55 gap is stable. The paper does not compare against the original FailureSensorIQ baselines from the IBM paper, which would help calibrate. More importantly, the perturbed datasets are under-specified: ASP, ACP, and the "Consistency-Based Accuracy" are undefined, and the construction of the simple and complex perturbations is not described. The abstract and conclusions attribute the drops to shuffling, paraphrasing, label changes, and distractors, but Table VII mixes these factors. There is no shuffle-only control. That matters because the paper's central claim is that fine-tuning is necessary for reasoning. If the simple perturbation is just option shuffling, then the fine-tuned model dropping from 88% to 43% is direct evidence of position bias. If it isn't, that control is simply missing. Either way, the claim that domain adaptation is required is not yet secured, since the model may be relying on surface cues rather than sensor-failure reasoning.

Minor issues: the text says retrieval experiments ran on CPU while LLM experiments ran on GPU, then says all experiments ran on GPU — contradictory. Reference [17] is the InstructGPT paper, not the Qwen2.5 model card. These are easy fixes.

Who is this for? Researchers and engineers working on industrial decision support and FMEA-style QA. It is a benchmark evaluation, not a method breakthrough. A serious referee should see it, but only with the expectation of major revisions: add multiple seeds and confidence intervals, a shuffle-only evaluation, a detailed description of the perturbations, and the original baseline comparison. If the authors address those, the paper becomes a solid empirical reference. As it stands, I would not rely on the headline claim.

Referee Report

3 major / 5 minor

Summary. The paper studies industrial failure-sensor diagnostic reasoning on the FailureSensorIQ multiple-choice QA benchmark. It develops an option-wise retrieval method that scores each answer option against top-k support and contradiction records from the training set using TF-IDF, BM25, semantic search, and a hybrid of TF-IDF and semantic search. It also evaluates Qwen2.5-7B-Instruct under zero-shot prompting, few-shot prompting, and QLoRA fine-tuning. The main empirical claims are that hybrid retrieval outperforms pure keyword retrieval (73.41% vs. 56.18% / 59.18%) and that QLoRA fine-tuning substantially outperforms prompting (88.01% vs. 55.43% / 54.49%). The paper also reports error analysis by question type and number of options, disagreement analysis among methods, and robustness results on two perturbed test sets.

Significance. If the results are supported, the paper provides a useful comparison of retrieval-only and LLM-based approaches for a specialized industrial QA task, with an interesting option-wise retrieval formulation that avoids prompt-based RAG. Using an external IBM Research benchmark reduces circularity concerns. The paper also attempts robustness analysis, which is valuable for deployment-oriented research. However, the strength of the central conclusions currently exceeds the evidence: the perturbed-dataset construction is under-specified, no option-shuffle-only control is reported, and the single-split/no-confidence-interval design leaves the comparative claims less secure than the text suggests. The core idea is promising and the main conclusions are likely recoverable with additional experiments.

major comments (3)

[§IV, Table VII] The construction of the 'simple' and 'complex' perturbed datasets is not specified, and the metrics ASP, ACP, and ACC are never defined. The conclusions attribute the accuracy drops to 'option shuffling, changed labels, paraphrasing, and additional distractors,' but Table VII conflates these factors. In particular, no option-shuffle-only control is reported. Since QLoRA drops from 88.01% to 43.07% on the simple perturbed set, one cannot tell whether the original accuracy reflects semantic reasoning or position/surface-cue memorization. The same issue applies to the retrieval baselines: pure option reordering should not change their per-option scores, so their large drops (e.g., hybrid from 73.41% to 29.21%) imply that the perturbations include more than shuffling. Please define the perturbations, separate their effects (at minimum a shuffle-only condition), and define the scoring for the
[§IV experimental setup] All comparative claims rest on a single random 80:20 split with no seed, no repeated runs, and no confidence intervals. Hyperparameters (λ=0.75, k=2, α=0.1, BM25 k1/b, QLoRA rank/alpha/lr/epoch) are fixed without a validation-driven selection or sensitivity analysis. The reported gaps are large, but the central claim that domain adaptation is 'necessary' would be more secure with multiple seeds or splits, bootstrap intervals, and at least a local sensitivity check on λ, k, and α.
[Table III] The paper does not report the original FailureSensorIQ baselines from reference [16], the paper that introduced the benchmark. Without these published numbers, the reader cannot calibrate whether the zero-shot accuracy of 55.43% is the expected level or whether the fine-tuned accuracy of 88.01% is a substantive advance relative to prior work on this dataset. Please include the IBM baselines and any other published results on FailureSensorIQ.

minor comments (5)

[§IV] The text says 'the retrieval-based experiments were conducted on a CPU. All these experiments were carried out on a GPU.' Please resolve this contradiction and specify the hardware precisely.
[Table IV] The terms 'Negative Questions' and 'Positive Questions' are used without definition. Please explain what makes a question negative or positive in FailureSensorIQ.
[Table VII] The first data column is labelled 'Acc (%)' in the table body but the caption lists ACC as 'Consistency-Based Accuracy on the Complex Perturbed Dataset.' Rename the columns to avoid ambiguity, and define ASP, ACP, and ACC in the text.
[§III.A] The sentence 'The retrieval-only reasoning method neither need any LLM nor training' is grammatically unclear. Also, the display equations would benefit from consistent formatting and numbering.
[References] Reference [17] cites Ouyang et al. (InstructGPT/RLHF), but the text describes Qwen2.5-7B-Instruct. Please cite the Qwen2.5 technical report or model card instead.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all central results are measured against the external FailureSensorIQ benchmark with no definitional reduction.

full rationale

I examined each claimed inference in the paper. The retrieval-score formula (Section III.A) combines top-k support and contradiction similarities from the external FailureSensorIQ training split; the reported accuracies (Table III) are measured on a held-out 534-question test set, not recomputed from the method's own fitted parameters. The comparison between retrieval methods and QLoRA fine-tuning is an empirical benchmark evaluation against an externally authored dataset (IBM Research, reference [16]); no load-bearing claim is justified solely by a self-citation. The hand-set hyperparameters (lambda=0.75, k=2, alpha=0.1, QLoRA lr=2e-4, one epoch) are configuration choices, but the paper does not define the evaluation metric in terms of them, nor does it 'predict' a quantity that is merely a fitted parameter renamed. The robustness analysis (Table VII) reports independent perturbed test sets, and the fine-tuned model's accuracy is an external measurement, not an identity or a tautology. Thus no step reduces by construction to its inputs. Concerns about missing shuffle-only controls, undefined perturbation metrics, or single-split variance are methodological reliability issues, not circularity.

Assumptions & free parameters 12 free parameters · 4 assumptions · 0 invented entities

The paper introduces no invented entities. Its central method relies on a manually configured retrieval score (λ, k, α) and standard QLoRA/hyperparameter choices; none of these are fitted to the test labels, but they are hand-selected without sensitivity analysis. The benchmark itself is assumed to be a valid evaluation set.

free parameters (12)

λ (contradiction weight) = 0.75
Weight in retrieval score; hand-chosen, no sensitivity analysis reported.
k (top-k support/contradict records) = 2
Number of nearest records averaged; hand-chosen.
α (hybrid fusion weight) = 0.1
Weight of TF-IDF in hybrid = 0.1*TF-IDF + 0.9*semantic; hand-chosen.
BM25 k1 = 1.5
Standard BM25 smoothing parameter; default.
BM25 b = 0.75
Standard BM25 length-normalization parameter; default.
Embedding model = all-MiniLM-L6-v2
The semantic encoder; results are likely sensitive to this choice.
LoRA rank = 16
QLoRA adapter rank; fixed for all experiments.
LoRA alpha = 32
QLoRA scaling factor.
LoRA dropout = 0.05
Dropout in LoRA adapters.
Learning rate = 2e-4
Adam learning rate for QLoRA, 1 epoch.
Max sequence length = 768
Truncation length for LLM inputs.
Batch size / gradient accumulation = 1 / 8
Effective batch size 8.

assumptions (4)

domain assumption FailureSensorIQ ground-truth labels are correct and consistent.
The paper uses the IBM benchmark as ground truth without auditing label quality; all accuracy numbers inherit this assumption.
domain assumption The random 80:20 split yields a test set that is representative and leakage-free.
No seed is given; the perturbed datasets were generated from the test set, and the paper asserts no data leakage without details.
domain assumption Similarity between option-level texts is a valid proxy for failure-sensor reasoning evidence.
The retrieval-only method assumes surface/semantic similarity to training records indicates the correct answer.
standard math Standard TF-IDF, BM25, and sentence-embedding similarity computations are correct.
The paper relies on standard implementations; no proofs given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Retrieval-based and Fine-tuned LLM Approaches for Industrial Asset Health Monitoring and Decision Support." pith.science (2026). https://pith.science/paper/AVWHBNOA

@misc{pith2026260724824,
  author       = {Pith},
  title        = {Pith review of: Retrieval-based and Fine-tuned LLM Approaches for Industrial Asset Health Monitoring and Decision Support},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AVWHBNOA}},
  note         = {Machine review of arXiv:2607.24824}
}

read the original abstract

Industrial plants run many important machines such as pumps, turbines, and compressors. Although engineers can use their experience to identify and diagnose machine problems, transferring this reasoning ability to computer systems remains difficult. This work studies how well a retrieval-only method and an open-source large language model (LLM) perform failure-sensor diagnostic reasoning using the FailureSensorIQ benchmark, a multiple-choice question-answering task introduced by IBM Research. In the retrieval-only approach, each answer option is converted into an option-level query and scored using similar correct and incorrect records from the training data. TF-IDF, BM25, semantic search, and hybrid search are tested and compared. In the LLM-based approach, the Qwen2.5-7B-Instruct model is evaluated using zero-shot prompting, few-shot prompting, and QLoRA fine-tuning. The results show that semantic search and hybrid search perform better than pure keyword-matching techniques, indicating that meaning-based similarity is more important for industrial failure-sensor reasoning. Among the LLM-based methods, the fine-tuned model achieves the best performance and substantially improves over zero-shot and few-shot prompting. Error analysis shows that performance decreases as the number of answer options increases. Robustness analysis also shows that all methods are sensitive to option shuffling, changed labels, paraphrasing, and additional distractors.

Figures

Figures reproduced from arXiv: 2607.24824 by the authors.

**Figure 1.** LLM-Based Interpretation of Failure Mode and Sensor Relationships TABLE I. DISTRIBUTION OF QUESTIONS BY ASSET TYPE Asset No. of Questions Electric motor 234 Steam turbine 171 Aero gas turbine 336 Industrial gas turbine 240 Pump 152 Compressor 220 Reciprocating internal combustion engine 336 Electric generator 234 Fan 200 Power transformer 544 Total 2667 TABLE II. DISTRIBUTION OF QUESTIONS BY NO. OF ANSWER OPTIONS Nu… view at source ↗

**Figure 2.** Sample questions from FailureSensorIQ [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Overview of retrieval-only and LLM-based reasoning methods [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Configuration of Qwen2.5-7B-Instruct model III. THE PROPOSED METHODOLOGIES This section provides detailed description of the methodologies (retrieval-only reasoning and LLM-based reasoning) developed to answer the industrial failure-sensor multiple-choice questions [P…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 2 linked inside Pith

[17]

Training language models to follow instructions with human feedback,

L. Ouyang et al., “Training language models to follow instructions with human feedback,” arXiv preprint arXiv:2203.02155, 2022

arXiv 2022
[16]

FailureSensorIQ: A Multi -Choice QA Dataset for Understanding Sensor Relationships and Failure Modes,

C. Constantinides, D. Patel, S. Lin, C. Guerrero, S. D. Patil, and J. Kalagnanam, “FailureSensorIQ: A Multi -Choice QA Dataset for Understanding Sensor Relationships and Failure Modes,” arXiv preprint arXiv:2506.03278, 2025

arXiv 2025
[1]

A survey on potentials, pathways and challenges of large language models in new-generation intelligent manufacturing,

C. Zhang, Q. Xu, Y. Yu, G. Zhou, K. Zeng, F. Chang, and K. Ding, “A survey on potentials, pathways and challenges of large language models in new-generation intelligent manufacturing,” Robotics and Computer- Integrated Manufacturing, vol. 92, Art. no. 102883, 2025

2025
[2]

Human-in-the-loop and large language models in smart manufacturing: Current applications, challenges, and perspectives,

M. S. Bajestani, D. Mun, and D. B. Kim, “Human-in-the-loop and large language models in smart manufacturing: Current applications, challenges, and perspectives,” Journal of Manufacturing Systems, vol. 86, pp. 913–941, 2026

2026
[3]

Framework for LLM applications in manufacturing,

C. I. Garcia, M. A. DiBattista, T. A. Letelier, H. D. Halloran, and J. A. Camelio, “Framework for LLM applications in manufacturing,” Manufacturing Letters, vol. 41, pp. 253–263, 2024

2024
[4]

Enhancing Industry 5.0 assistance in assembling tasks with large language models and retrieval augmented generation: The DIA framework,

F. Bianchini, M. Calamo, S. Colabianchi, F. Costantino, M. Mecella, J. Rossi, and N. Sabetta, “Enhancing Industry 5.0 assistance in assembling tasks with large language models and retrieval augmented generation: The DIA framework,” Procedia Computer Science, vol. 277, pp. 1684– 1693, 2026

2026
[5]

Large language models for low -code process automation: Lowering the barriers for citizen developers,

P. A. François, S. Ciftci, C. Janiesch, and R. Plattfaut, “Large language models for low -code process automation: Lowering the barriers for citizen developers,” in Proc. Australasian Conference on Information Systems (ACIS 2025), 2025, Art. no. 242

2025
[6]

Utilizing LLMs for industrial process automation,

S. Fares, “Utilizing LLMs for industrial process automation,” arXiv preprint arXiv:2602.23331, 2026

arXiv 2026

Show all 18 references

[7]

Towards a benchmark dataset for large language models in the context of process automation,

T. Tizaoui and R. Tan, “Towards a benchmark dataset for large language models in the context of process automation,” Digital Chemical Engineering, vol. 13, Art. no. 100186, 2024

2024
[8]

Knowledge-enhanced industrial question-answering using large language models,

R. Liu, H. Ren, H. Ren, W. Rui, W. Cui, X. Liang, C. Yang, and W. Gui, “Knowledge-enhanced industrial question-answering using large language models,” Engineering, vol. 60, pp. 142–153, 2026

2026
[9]

An experiment of using a large language model to control a water tank system,

H. Wen, J. Roberts, A. Zaidi, and A. McLeod, “An experiment of using a large language model to control a water tank system,” Computers & Chemical Engineering, vol. 211, Art. no. 109656, 2026

2026
[10]

Autonomous industrial control using an agentic framework with large language models,

J. Vyas and M. Mercangöz, “Autonomous industrial control using an agentic framework with large language models,” IFAC-PapersOnLine, vol. 59, no. 6, pp. 349–354, 2025, doi: 10.1016/j.ifacol.2025.07.170

2025 doi
[11]

Large language model agent for user - friendly chemical process simulations,

J. Liang, N. Groll, and G. Sin, “Large language model agent for user - friendly chemical process simulations,” Digital Chemical Engineering, vol. 19, Art. no. 100312, 2026

2026
[12]

Improving process systems engineering with specialized multi -agent large language models,

F. A. R. D. Lima, A. Abdelrehim, A. Bharambe, M. Micluţa-Câmpeanu, D. Gandhi, A. Singhvi, et al., “Improving process systems engineering with specialized multi -agent large language models,” Chemical Engineering Journal Advances, vol. 26, Art. no. 101141, 2026

2026
[13]

A multi -expert framework for enhancing multimodal large language models in industrial anomaly detection,

Z. Chen and F. Imani, “A multi -expert framework for enhancing multimodal large language models in industrial anomaly detection,” Pattern Recognition, vol. 172, Art. no. 112752, 2026

2026
[14]

Oscillation analysis using retrieval-augmented generation,

A. Singh, M. M. Rousta, and B. Huang, “Oscillation analysis using retrieval-augmented generation,” Computers & Chemical Engineering, vol. 206, Art. no. 109489, 2026

2026
[15]

StictionGPT: Detecting valve stiction in process control loops using large vision language model,

T. Xue, C. Shang, D. Huang, and B. Huang, “StictionGPT: Detecting valve stiction in process control loops using large vision language model,” Control Engineering Practice, vol. 165, Art. no. 106588, 2025

2025
[18]

Dettmers, T., Pagnoni, A., Holtzman, A., & Zettlemoyer, L. (2023). QLoRA: Efficient Finetuning of Quantized LLMs. arXiv

2023

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [17]

Training language models to follow instructions with human feedback,

L. Ouyang et al., “Training language models to follow instructions with human feedback,” arXiv preprint arXiv:2203.02155, 2022

arXiv 2022

[2] [16]

FailureSensorIQ: A Multi -Choice QA Dataset for Understanding Sensor Relationships and Failure Modes,

C. Constantinides, D. Patel, S. Lin, C. Guerrero, S. D. Patil, and J. Kalagnanam, “FailureSensorIQ: A Multi -Choice QA Dataset for Understanding Sensor Relationships and Failure Modes,” arXiv preprint arXiv:2506.03278, 2025

arXiv 2025

[3] [1]

A survey on potentials, pathways and challenges of large language models in new-generation intelligent manufacturing,

C. Zhang, Q. Xu, Y. Yu, G. Zhou, K. Zeng, F. Chang, and K. Ding, “A survey on potentials, pathways and challenges of large language models in new-generation intelligent manufacturing,” Robotics and Computer- Integrated Manufacturing, vol. 92, Art. no. 102883, 2025

2025

[4] [2]

Human-in-the-loop and large language models in smart manufacturing: Current applications, challenges, and perspectives,

M. S. Bajestani, D. Mun, and D. B. Kim, “Human-in-the-loop and large language models in smart manufacturing: Current applications, challenges, and perspectives,” Journal of Manufacturing Systems, vol. 86, pp. 913–941, 2026

2026

[5] [3]

Framework for LLM applications in manufacturing,

C. I. Garcia, M. A. DiBattista, T. A. Letelier, H. D. Halloran, and J. A. Camelio, “Framework for LLM applications in manufacturing,” Manufacturing Letters, vol. 41, pp. 253–263, 2024

2024

[6] [4]

Enhancing Industry 5.0 assistance in assembling tasks with large language models and retrieval augmented generation: The DIA framework,

F. Bianchini, M. Calamo, S. Colabianchi, F. Costantino, M. Mecella, J. Rossi, and N. Sabetta, “Enhancing Industry 5.0 assistance in assembling tasks with large language models and retrieval augmented generation: The DIA framework,” Procedia Computer Science, vol. 277, pp. 1684– 1693, 2026

2026

[7] [5]

Large language models for low -code process automation: Lowering the barriers for citizen developers,

P. A. François, S. Ciftci, C. Janiesch, and R. Plattfaut, “Large language models for low -code process automation: Lowering the barriers for citizen developers,” in Proc. Australasian Conference on Information Systems (ACIS 2025), 2025, Art. no. 242

2025

[8] [6]

Utilizing LLMs for industrial process automation,

S. Fares, “Utilizing LLMs for industrial process automation,” arXiv preprint arXiv:2602.23331, 2026

arXiv 2026