Pith. sign in

REVIEW 3 major objections 5 minor 35 references

Fine-tuning only the language layer of a lightweight multimodal model on a few expert-style chain-of-thought pairs grades power-transmission defects better than commercial systems.

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 ↗

Fine-tuning only the LLM layer of Qwen3-VL-8B on a handful of commercial-MLLM-generated CoT Q&As yields SOTA accuracy on three few-shot power-equipment defect-grading tasks and supports multi-task use.

T0 review reviewed 2026-07-13 challenge →

load-bearing objection Solid first-of-kind industrial result: LLM-layer LoRA on a few dozen teacher CoTs beats commercial MLLMs on three few-shot DGPTE tasks, with the usual private-data and verification caveats. the 3 major comments →

arxiv 2605.28822 v1 pith:ZUSRMPXZ submitted 2026-03-25 cs.CL

Lightweight Multimodal LLM-Enabled Cost-Effective Defect Grading of Power Transmission Equipment

Tao Wang , Lipeng Zhu , Jiayong Li , Feng Gao , Siwen Liang This is my paper
classification cs.CL
keywords defect gradingpower transmission equipmentmultimodal large language modelin-context learningLoRAchain-of-thoughtfew-shot learningQwen3-VL
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.

The reading

Refined defect grading on power lines has stayed mostly manual because samples are heavily imbalanced and expert logic is hard to encode in ordinary detectors. This paper shows that commercial multimodal models, prompted with decision-tree rules and a few reference images, can already reach usable accuracy; a second prompt then forces them to write short, verifiable reasoning traces for each training image. Those traces become the entire training set for a far smaller open model. Updating only its language layer with low-rank adaptation produces accuracies of roughly 93 percent, 86 percent and 77 percent on three graded tasks—higher than the best commercial models—and the same model can be jointly trained on all three tasks with only a small drop. The method therefore turns expensive expert knowledge into a cheap, interpretable, offline pipeline.

Core claim

When a commercial multimodal model is first steered by in-context learning and decision-tree prompts into state-of-the-art defect grading, then asked to regenerate the full reasoning path given the true grade label, the resulting small set of verified question-answer pairs is enough to fine-tune Qwen3-VL-8B. Restricting the update to the language-model layer alone yields the highest accuracies (92.63 percent, 85.85 percent, 77.04 percent) on three few-shot power-transmission grading tasks, surpassing every commercial baseline and remaining competitive under multi-task joint fine-tuning.

What carries the argument

Decision-tree-based chain-of-thought (DT-based CoT) generation: a commercial model is first maximized via in-context learning, then re-queried with the true grade so that it writes the intermediate reasoning steps; those verified pairs drive LoRA supervised fine-tuning restricted to the language-model layer of the lightweight multimodal model.

Load-bearing premise

The commercial model, given the correct grade and the expert decision tree, produces reasoning traces that remain factually correct after light human review and that still capture the visual cues needed for the held-out long-tail images.

What would settle it

Rewrite the machine-generated chain-of-thought traces by hand on the same few dozen training images, retrain the identical LoRA setup, and check whether test accuracy falls by more than a few points; or show that full-module fine-tuning on the same tiny set suddenly outperforms language-layer-only adaptation.

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

If this is right

  • A single offline 8-billion-parameter multimodal model can replace multi-stage visual pipelines for several defect-grading tasks at once.
  • Expert knowledge written as decision trees can be transferred into model reasoning without collecting large annotated datasets.
  • Annotation cost collapses to light verification of machine-written reasoning traces rather than full manual labeling.
  • Long-tail industrial visual tasks become competitive with human grading once only the language layer is adapted.
  • Joint multi-task fine-tuning remains feasible, so one lightweight model can serve multiple equipment types.

Where Pith is reading between the lines

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

  • The same commercial-to-lightweight distillation of expert decision-tree traces could transfer to other safety-critical inspections that suffer from rare defect classes, such as rail or pipeline monitoring.
  • Because the visual encoder already appears to extract the needed features, future industrial adapters may freeze vision entirely and train only language-side low-rank modules.
  • The performance ceiling on the hardest four-class nest-location task suggests that pure text chain-of-thought still misses fine three-dimensional geometric cues that a specialized detector could supply as additional tokens.
  • Multi-turn dialogue fine-tuning, listed by the authors as next work, would let field engineers query intermediate decision-tree steps on demand without leaving the model.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 5 minor

Summary. The paper proposes a cost-effective MLLM pipeline for refined defect grading of power transmission equipment (DGPTE) under few-shot, long-tail conditions. Expert standards are cast as decision-tree CoT; commercial MLLMs are prompted with ICL (reference images + DT logic) to select a SOTA teacher (Eq. 2), which is then re-queried with the true grade label to synthesize a small set of CoT Q&As (Eq. 3, ~30 images per class). These Q&As are used for LoRA-SFT of Qwen3-VL-8B. On three tasks of increasing difficulty (binary crimping, 3-class corrosion, 4-class bird-nest impact), fine-tuning only the LLM layer yields reported accuracies of 92.63 %, 85.85 % and 77.04 %, surpassing the best commercial models under the same ICL prompt (90.26 %, 56.58 %, 53.80 %). Multi-task joint fine-tuning remains competitive (drop <3 %), and ablations examine module choice, image token position, and resolution.

Significance. If the reported gains hold under independent verification, the work supplies a practical, offline-deployable recipe for industrial visual grading that integrates expert logic without large-scale annotation or multi-model pipelines. The explicit finding that only the LLM layer needs adaptation under few-shot conditions, together with the multi-task joint result, is useful for resource-constrained power-inspection settings. Strengths include a clear three-task ladder, systematic commercial-model comparison (Table V), module/position/resolution ablations (Figs. 4–5, Tables VI–VII), and an interpretable DT-CoT format. The contribution is primarily empirical and engineering-oriented rather than theoretical.

major comments (3)
  1. Section III-B Steps 2–3 and Eqs. (2)–(3): the training CoTs are generated by first selecting m* on the evaluation distribution via R_dg accuracy, then conditioning the same model on the true grade R_GR. Human review is described only as light inspection “to ensure full correctness.” Because the large accuracy jumps (especially Task 2: 56.58 % → 85.85 %; Task 3: 53.80 % → 77.04 %) rest on these label-conditioned rationales being both factually accurate at every DT step and representative of held-out long-tail visual modes, the manuscript needs quantitative evidence of CoT fidelity (e.g., step-level error rates before/after review, inter-annotator agreement, or a control that trains on pure labels without CoT). Without it the central claim that the student generalizes rather than memorizes a brittle mapping remains under-supported.
  2. Section IV / Figs. 4–5 and Tables V–VII: all headline numbers are point estimates with no error bars, confidence intervals, or statistical significance tests, despite a free-parameter set that includes epochs, batch size, LoRA rank/alpha and learning rate, and a stratified 30-per-class train split. At minimum, report results over multiple random seeds or stratified folds so that the claimed SOTA margins and the “<3 % multi-task drop” can be assessed for stability.
  3. Section II and the evaluation protocol: the paper asserts that the fine-tuned model reaches “human-like” / “comparable to manual grading” performance, yet no human baseline accuracy, inter-rater reliability, or expert-vs-model confusion analysis is provided on the same test images. Without that comparison the claim that automated grading now matches or exceeds manual work cannot be verified.
minor comments (5)
  1. Throughout: inconsistent acronyms (DGPTE / PTEDG / GDPTE / DDPTE) and repeated typos (“state-of-te-art”, “Fint-tuning”, “Rusults”, “Mecro-F1”, “Language Language Model layer”) should be cleaned.
  2. Eq. (1) and Table III: image resizing and cutoff-length choices are reasonable but the interaction between max-side 1280 and the 12 280-token cutoff is never quantified; a short note on token budget would help reproducibility.
  3. Fig. 3 / prompt construction: the exact JSON schema and the full text of R_dg / R_qa are only partially illustrated; releasing the complete prompt templates would strengthen reproducibility.
  4. Table V Case 4 vs. fine-tuned results: commercial models are evaluated zero/few-shot while the student is fine-tuned; a brief discussion of this asymmetry (and whether commercial models could also be lightly adapted) would clarify the comparison.
  5. References: several recent industrial MLLM / defect-grading works are cited, but a short related-work paragraph contrasting the present DT-CoT distillation approach with pure detection or knowledge-graph methods would improve positioning.

Circularity Check

0 steps flagged

No circularity: empirical few-shot SFT pipeline with independent held-out evaluation; no derivation reduces to its inputs by construction.

full rationale

The paper is a standard empirical MLLM fine-tuning study. Commercial models are ranked by ICL accuracy under four prompt variants (Table V, Eq. 2), the best is then prompted with true labels to emit DT-based CoT rationales only for the stratified 30-per-class training images (Eq. 3, Sec. III-B Step 3), those rationales are lightly human-checked, and Qwen3-VL-8B is LoRA-SFT’d (only LLM layer) before accuracy/MF1 is measured on the remaining held-out images. Nothing is defined in terms of the quantity later claimed as a result; no free parameter is fitted to a subset and then re-reported as a prediction; no uniqueness theorem or ansatz is imported via self-citation; and the multi-task joint-fine-tuning numbers remain ordinary empirical measurements. The mild methodological risk that teacher selection may have used the same distribution later used for comparison is ordinary experimental practice, not a circular reduction of any equation or claim to its own inputs. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central empirical claim rests on a small number of free hyper-parameters chosen for the few-shot regime, on the domain assumption that expert decision-tree logic plus commercial-MLLM CoT generation yields transferable supervision, and on no newly invented physical entities.

free parameters (6)
  • shots_per_class = 30
    Fixed at 30 images per defect grade for training; the entire few-shot claim depends on this hand-chosen sample size.
  • LoRA_rank = 32
    Set to 32; controls the capacity of the adapter that produces the reported SOTA numbers.
  • LoRA_alpha = 128
    Set to 128; scales the adapter update magnitude.
  • learning_rate = 5e-5
    Fixed at 5e-5 for all LoRA runs.
  • image_max_side = 1280
    Resized so longest side equals 1280 px; ablation shows accuracy is sensitive to this choice.
  • epochs = variable
    Dynamically chosen from {5,50,100,150,300,500} per module configuration to fit memory; final reported numbers depend on the selected epoch.
axioms (4)
  • domain assumption A decision-tree decomposition of expert grading rules can be faithfully encoded as a chain-of-thought prompt that commercial MLLMs can follow.
    Invoked in Section III-A and Table II; without it the generated Q&As would not transfer expert knowledge.
  • domain assumption When a commercial MLLM is given the true grade label, the CoT it produces (after light human verification) is a high-quality training target for a smaller student model.
    Core of the Fine-tuning Dataset Construction module (Eq. 3 and Step 3).
  • ad hoc to paper The visual encoder and multimodal alignment layers of Qwen3-VL-8B already extract the features needed for defect grading; only the language layer needs adaptation under few-shot conditions.
    Empirical finding of Section IV-C and Fig. 4; treated as an operating assumption for the recommended fine-tuning recipe.
  • domain assumption Stratified sampling of 30 images per class yields a training distribution representative enough for the long-tail test set.
    Stated in Section III-A; the few-shot claim rests on this sampling assumption.

reviewed 2026-07-13 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Lightweight Multimodal LLM-Enabled Cost-Effective Defect Grading of Power Transmission Equipment." pith.science (2026). https://pith.science/paper/ZUSRMPXZ

@misc{pith2026260528822,
  author       = {Pith},
  title        = {Pith review of: Lightweight Multimodal LLM-Enabled Cost-Effective Defect Grading of Power Transmission Equipment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZUSRMPXZ}},
  note         = {Machine review of arXiv:2605.28822}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Defect grading of power transmission equipment (DGPTE) is crucial to the stability of electric energy transmission. Although existing machine learning methods exhibit strong capabilities in defect detection, they are plagued by difficulties in integrating expert experience and facing class imbalance in more refined defect grading field. To address this issue, this paper introduces a novel defect grading framework based on multimodal large language model (MLLM). Specifically, this approach maximizes the commercial MLLMs' potential of DGPTE through in-context learning and obtains the state-of-te-art (SOTA) model. By sending a secondary request to this model, a small number of chain of thought-based question-answer pairs (Q\&As) are generated, which effectively reduces the cost of manual annotation. In this way, these high-quality interpretable Q\&As are used to train Qwen3-VL-8B via Low-Rank Adaption-based supervised fine-tuning (SFT). Experimental results on three DGPTE tasks demonstrate that fine-tuning only the language model layer yields the SOTA performance. Furthermore, multi-task joint fine-tuning verifies the feasibility of handling multiple grading tasks within only a single lightweight MLLM.

Figures

Figures reproduced from arXiv: 2605.28822 by Feng Gao, Jiayong Li, Lipeng Zhu, Siwen Liang, Tao Wang.

Figure 2
Figure 2. Figure 2: It concludes three steps below: • Step 1: Input all images to YOLOv8 for object detection, where the positions of objective equipment need to be labeled in the images. It provides precise guidance for subsequent defect grading. • Step 2: Design defect grading prompt Rdg, feed it into the commercial MLLMs and select the SOTA MLLM for the target tasks: m∗ = argmax m∈M ACCm(Rdg) (2) where m∗ and M denote the … view at source ↗
Figure 2
Figure 2. Figure 2: The framework of proposed fine-tuning strategy [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: An example of Q&A for MLLM fine-tuning in Task 1 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy comparison of different fine-tuning strategies [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Accuracy comparison of different fine-tuning modules [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: two generated defect grading examples for Task 3 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] 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

35 extracted references

  1. [1]

    Accurate industrial anomaly detection and localization using weakly-supervised residual transformers,

    H. Li, J. Wu, D. Liu, L. Y . Wu, H. Chen, and C. Shen, “Accurate industrial anomaly detection and localization using weakly-supervised residual transformers,”IEEE Transactions on Image Processing, vol. 35, pp. 1551–1566, 2026

    2026

  2. [2]

    Pga-net: Pyramid feature fusion and global context attention network for automated surface defect detection,

    H. Dong, K. Song, Y . He, J. Xu, Y . Yan, and Q. Meng, “Pga-net: Pyramid feature fusion and global context attention network for automated surface defect detection,”IEEE Transactions on Industrial Informatics, vol. 16, no. 12, pp. 7448–7458, 2020. 9 TABLE VII PERFORMANCEOVERHEAD BYDIFFERENTINPUTSIZES Size ACC MF1 Training Time (h) Training VRAM (GB) Infer...

    2020

  3. [3]

    Vlcim: A vision-language cyclic interaction model for industrial defect detection,

    X. Shen, L. Li, Y . Ma, S. Xu, J. Liu, Z. Yang, and Y . Shi, “Vlcim: A vision-language cyclic interaction model for industrial defect detection,” IEEE Transactions on Instrumentation and Measurement, vol. 74, pp. 1–13, 2025

    2025

  4. [4]

    Joint knowledge graph and large language model for fault diagnosis and its application in aviation assembly,

    P. LIU, L. Qian, X. Zhao, and B. Tao, “Joint knowledge graph and large language model for fault diagnosis and its application in aviation assembly,”IEEE Transactions on Industrial Informatics, vol. 20, no. 6, pp. 8160–8169, 2024

    2024

  5. [5]

    Metric learning-based fault diagnosis and anomaly detection for industrial data with intraclass variance,

    K. Huang, S. Wu, B. Sun, C. Yang, and W. Gui, “Metric learning-based fault diagnosis and anomaly detection for industrial data with intraclass variance,”IEEE Transactions on Neural Networks and Learning Sys- tems, vol. 35, no. 1, pp. 547–558, 2024

    2024

  6. [6]

    Intelligent small sample defect detection of water walls in power plants using novel deep learning integrating deep convolutional gan,

    Z. Geng, C. Shi, and Y . Han, “Intelligent small sample defect detection of water walls in power plants using novel deep learning integrating deep convolutional gan,”IEEE Transactions on Industrial Informatics, vol. 19, no. 6, pp. 7489–7497, 2023

    2023

  7. [7]

    Thermal defect diagnosis of power equipments with spectral residual transformation,

    Z. Huang and S. Wu, “Thermal defect diagnosis of power equipments with spectral residual transformation,” in2023 8th International Con- ference on Computational Intelligence and Applications (ICCIA), 2023, pp. 226–230

    2023

  8. [8]

    Partial discharge detection in an oil-filled power transformer using fiber bragg grating sensors: A review,

    S. N. Meitei, K. Borah, and S. Chatterjee, “Partial discharge detection in an oil-filled power transformer using fiber bragg grating sensors: A review,”IEEE Sensors Journal, vol. 21, no. 9, pp. 10 304–10 316, 2021

    2021

  9. [9]

    A novel analytical methodology for estimating high-frequency lumped model inductances and series capacitance of transformer winding: an indirect measurement procedure,

    M. S. Chaouche, F. Didi, A. Amara, H. Houassine, M. F. Mohd Yousof, A. F. Tazay, A. Flah, M. K. Metwaly, R. N. Ghaly, and S. S. Ghoneim, “A novel analytical methodology for estimating high-frequency lumped model inductances and series capacitance of transformer winding: an indirect measurement procedure,”Electric Power Systems Research, vol. 255, p. 112722, 2026

    2026

  10. [10]

    An expert system for transformer fault diagnosis using dissolved gas analysis,

    C. Lin, J.-M. Ling, and C.-L. Huang, “An expert system for transformer fault diagnosis using dissolved gas analysis,”IEEE Transactions on Power Delivery, vol. 8, no. 1, pp. 231–238, 1993

    1993

  11. [11]

    Neural net and expert system diagnose transformer faults,

    Z. Wang, Y . Liu, and P. Griffin, “Neural net and expert system diagnose transformer faults,”IEEE Computer Applications in Power, vol. 13, no. 1, pp. 50–55, 2000

    2000

  12. [12]

    A knowledge-augmented multistage reason- ing approach for wind turbine fault cause analysis,

    Y . Hu, P. Wen, and Y . Dai, “A knowledge-augmented multistage reason- ing approach for wind turbine fault cause analysis,”IEEE Transactions on Industrial Informatics, pp. 1–12, 2026

    2026

  13. [13]

    A review and progress of insulation fault diagnosis for cable using partial discharge approach,

    G. Wu, T. Zhang, B. Cao, K. Liu, K. Chen, and G. Gao, “A review and progress of insulation fault diagnosis for cable using partial discharge approach,”IEEE Transactions on Dielectrics and Electrical Insulation, vol. 32, no. 3, pp. 1639–1652, 2025

    2025

  14. [14]

    A review of ai-based partial discharge analysis studies in medium voltage switchgear,

    T. Aydin and A. Ozdemir, “A review of ai-based partial discharge analysis studies in medium voltage switchgear,”2025 60th International Universities Power Engineering Conference (UPEC), pp. 1–6, 2025

    2025

  15. [15]

    A review of power equipment defect detection based on deep learning,

    J. Wang, Z. Cheng, and F. Meng, “A review of power equipment defect detection based on deep learning,” in2023 IEEE 12th International Con- ference on Communication Systems and Network Technologies (CSNT), 2023, pp. 120–125

    2023

  16. [16]

    Structural defect detection technology of transmission line damper based on uav image,

    X. Huang, Y . Wu, Y . Zhang, and B. Li, “Structural defect detection technology of transmission line damper based on uav image,”IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1–14, 2023

    2023

  17. [17]

    Coma-ikg: Llm- driven multiagent framework for automated construction of industrial knowledge graph,

    J. Zhang, H. Wang, Z. Jia, J. Dong, and L. Ren, “Coma-ikg: Llm- driven multiagent framework for automated construction of industrial knowledge graph,”IEEE Transactions on Industrial Informatics, pp. 1– 12, 2026

    2026

  18. [18]

    Design of intelligent detection method for electricity transmission line equipment defect based on data mining algorithm,

    X. Yin, Y . Zuo, and G. Fu, “Design of intelligent detection method for electricity transmission line equipment defect based on data mining algorithm,”International Journal of Thermofluids, vol. 24, p. 100814, 2024

    2024

  19. [19]

    Cacs-yolo: A lightweight model for insulator defect detection based on improved yolov8m,

    Z. Cao, K. Chen, J. Chen, Z. Chen, and M. Zhang, “Cacs-yolo: A lightweight model for insulator defect detection based on improved yolov8m,”IEEE Transactions on Instrumentation and Measurement, vol. 73, pp. 1–10, 2024

    2024

  20. [20]

    A weakly supervised contrastive learning pretraining method for visual defect detection of transmission lines,

    M. He, L. Qin, Y . Wang, X. Deng, Q. Liu, Y . Zhang, and K. Liu, “A weakly supervised contrastive learning pretraining method for visual defect detection of transmission lines,”IEEE Transactions on Instru- mentation and Measurement, vol. 74, pp. 1–15, 2025

    2025

  21. [21]

    Few- shot power equipment defect classification based on image augmentation and reconstruction,

    L. Di, J. Zhao, F. Wei, X. Zhang, X. Zhu, Y . Yang, and M. Sun, “Few- shot power equipment defect classification based on image augmentation and reconstruction,” in2025 International Conference on Image and Video Processing (ICIVP), 2025, pp. 99–103

    2025

  22. [22]

    Human-like one-shot fault diagnosis via large language model,

    H. Han, X. Gao, H. Han, H. Gao, and Y . Qi, “Human-like one-shot fault diagnosis via large language model,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, pp. 1–14, 2025

    2025

  23. [23]

    Solar photovoltaic assessment with large lan- guage model,

    M. Guo and Y . Weng, “Solar photovoltaic assessment with large lan- guage model,”Applied Energy, vol. 402, p. 126835, 2025

    2025

  24. [24]

    Iad-gpt: Advancing visual knowledge in multimodal large language model for industrial anomaly detection,

    Z. Li, Z. Yu, Q. Ye, W. Xie, W. Zhuo, and L. Shen, “Iad-gpt: Advancing visual knowledge in multimodal large language model for industrial anomaly detection,”IEEE Transactions on Instrumentation and Measurement, vol. 74, pp. 1–12, 2025

    2025

  25. [25]

    T2mfdf: An llm-enhanced multimodal fault diagnosis framework inte- grating time-series and textual data,

    J. Zhou, Y . Guo, Z. Yang, J. Yang, Z. An, K. Li, and S. McLoone, “T2mfdf: An llm-enhanced multimodal fault diagnosis framework inte- grating time-series and textual data,”IEEE Transactions on Instrumen- tation and Measurement, vol. 74, pp. 1–11, 2025

    2025

  26. [26]

    Cmkr-pbdm: A transmission line pin-missing bolts detection method based on cross-media and knowledge reasoning,

    W. Zhao, Y . Ding, L. Zhang, B. Liu, C. Yang, Z. Zhao, Z. Zhao, Y . Zhai, and M. Xu, “Cmkr-pbdm: A transmission line pin-missing bolts detection method based on cross-media and knowledge reasoning,”IEEE Transactions on Power Delivery, vol. 40, no. 2, pp. 1030–1039, 2025

    2025

  27. [27]

    Vlf- detr: Integrating vision-language and high-frequency features for trans- mission line defect detection,

    K. Zhang, J. Yang, J. Wang, Z. Zheng, X. Sheng, and N. Zhang, “Vlf- detr: Integrating vision-language and high-frequency features for trans- mission line defect detection,”IEEE Transactions on Instrumentation and Measurement, vol. 74, pp. 1–15, 2025

    2025

  28. [28]

    Graph computing based electric power equipment defect grading with multi-scale mechanism,

    J. Fei, M. Zhenyuan, X. Jiannan, T. Yuanpeng, D. Minghui, and T. Jie, “Graph computing based electric power equipment defect grading with multi-scale mechanism,” in2022 4th International Conference on Smart Power & Internet Energy Systems (SPIES), 2022, pp. 2344–2349

    2022

  29. [29]

    Knowledge- augmented zero-shot method for power equipment defect grading with chain-of-thought llms,

    J. Du, B. Li, Z. Chen, L. Shen, P. Liu, and Z. Ran, “Knowledge- augmented zero-shot method for power equipment defect grading with chain-of-thought llms,”Electronics, vol. 14, no. 15, 2025

    2025

  30. [30]

    Enhancing power equipment defect identification through multi-label classification methods,

    W. Zheng, Y . Yang, F. Zhang, W. Lv, Y . Li, and S. Li, “Enhancing power equipment defect identification through multi-label classification methods,”Scientific Reports, vol. 14, 2024

    2024

  31. [31]

    True few-shot learning with language models,

    E. Perez, D. Kiela, and K. Cho, “True few-shot learning with language models,” ser. NIPS ’21. Red Hook, NY , USA: Curran Associates Inc., 2021

    2021

  32. [32]

    A survey on in-context learning,

    Q. Dong, L. Li, D. Dai, C. Zheng, J. Ma, R. Li, H. Xia, J. Xu, Z. Wu, B. Chang, X. Sun, L. Li, and Z. Sui, “A survey on in-context learning,” inProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. Miami, Florida, USA: Association for Computational Linguistics, Nov. 2024, pp. 1107–1128

    2024

  33. [33]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. H. Chi, Q. V . Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” inProceedings of the 36th International Conference on Neural Information Processing Systems, ser. NIPS ’22. Red Hook, NY , USA: Curran Associates Inc., 2022

    2022

  34. [34]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, yelong shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” inInternational Conference on Learning Representations, 2022

    2022

  35. [35]

    Qwen3-vl-embedding and qwen3- vl-reranker: A unified framework for state-of-the-art multimodal retrieval and ranking,

    M. Li, Y . Zhang, D. Long, K. Chen, S. Song, S. Bai, Z. Yang, P. Xie, A. Yang, D. Liu, J. Zhou, and J. Lin, “Qwen3-vl-embedding and qwen3- vl-reranker: A unified framework for state-of-the-art multimodal retrieval and ranking,” 2026

    2026

This paper was first reviewed by grok-4.5 on July 13, 2026.