Pith. sign in

REVIEW 4 major objections 5 minor 82 references

The paper claims that a few-shot diffusion pipeline, guided by geometric crack maps and a hypernetwork, can generate substation-meter defect images realistic enough that a detector trained on real-plus-synthetic data outperforms one trained

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

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

T0 review · deepseek-v4-flash

2026-08-03 10:39 UTC pith:3LPMHIX7

load-bearing objection A plausible few-shot defect-generation recipe on a niche industrial dataset, undercut by an unspecified hypernetwork and inconsistent headline numbers. the 4 major comments →

arxiv 2601.09238 v2 pith:3LPMHIX7 submitted 2026-01-14 cs.CV

Knowledge-Embedded and Hypernetwork-Guided Few-Shot Substation Meter Defect Image Generation Method

Jackie Alex , Justin Petter This is my paper
classification cs.CV
keywords Defect Image GenerationFew-Shot LearningStable DiffusionKnowledge EmbeddingHypernetworkSubstation EquipmentConditionally Controllable SynthesisData Augmentation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims that severe scarcity of cracked substation-meter images can be addressed by generating synthetic defects with a diffusion model, and that the synthetic images are good enough to improve a downstream detector beyond what real data alone achieves. The proposed framework has three parts: a DreamBooth-style fine-tuning step that teaches a pre-trained Stable Diffusion model the visual character of substation meters; a geometric crack model that turns parameters such as location, length, curvature, and branching into pixel-level control maps; and a hypernetwork that adjusts the frozen denoising process in response to those maps. On a dataset of 100 cracked and 529 normal meter images, the authors report an FID of 76.72 and an Inception Score of 2.45, and that adding 40% generated images to a YOLOv8n training set raises mAP50 by 19.1% and precision by 26.9% relative to real data alone. If correct, the method offers a practical recipe for few-shot industrial defect synthesis: a handful of real defect examples plus a controllable generator can supply the training diversity that rare-defect domains lack.

Core claim

The central discovery is that a three-part recipe—domain knowledge embedding, parametric crack modeling, and hypernetwork-guided control—can make a pre-trained text-to-image diffusion model produce industrially usable defect images from very few examples. Knowledge embedding (a DreamBooth-style subject and prior-preservation loss) adapts the backbone to meter-specific structure and texture. The crack feature module fuses an edge map, a defect mask, and a segmentation-derived dial constraint into a control image that specifies where cracks may appear. The conditional control generation mechanism then uses a hypernetwork to modulate the frozen denoiser from that control image at multiple featu

What carries the argument

The central object is the control image produced by the crack feature modeling module: a pixel-level map formed by taking the pixel-wise maximum of an edge map and a crack mask, then constrained by a Segment Anything-derived dial region. This control image carries the spatial specification of the defect. The hypernetwork-guided conditional control module is the mechanism that injects that specification into generation: a trainable branch processes the control image and produces dynamic weights that modulate the frozen Stable Diffusion denoiser at multiple feature layers, letting the model balance fidelity to the control map with visual diversity. The DreamBooth-style knowledge embedding is t

Load-bearing premise

The load-bearing premise is that the hypernetwork is actually implemented and actually modulates the frozen denoiser in response to the control maps—the paper supplies no architecture, parameterization, or equation for it, so if this component does not exist or does not behave as described, the controllability mechanism and the reported downstream gains lose their explanation.

What would settle it

Re-run the pipeline with the hypernetwork branch ablated and test the downstream detector: if mAP50 still rises by ~19%, the hypernetwork is not the cause. More directly, fix the random seed, change only the crack mask (e.g., move the crack to a different dial region), and check whether the generated defect's location and shape track the mask; if outputs ignore the control map, the claimed pixel-level controllability is not real. Also check generalization: compute FID against held-out real defect images rather than the training set—if the model has memorized the 100 training cracks, FID will d

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

If this is right

  • If the numbers hold, industrial inspection systems can train rare-defect detectors with a handful of real defect images plus controllable synthetic ones, cutting annotation cost sharply.
  • The recipe generalizes by design: any defect whose geometry can be parameterized (location, length, curvature, branching) can be turned into a control map and fed to the same pipeline.
  • Because performance improves across SD1.5, SDXL, and FLUX, the gains are not tied to one backbone; future base-model improvements should compound with the method.
  • The 40% mixing ratio being best, with declines at 60–100% synthetic, implies the generator's distribution is useful but narrower than the real distribution; practitioners should tune this ratio rather than assume more is better.

Where Pith is reading between the lines

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

  • An implication the authors leave implicit: the hypernetwork, if implemented as a small weight-predicting network, could be made to predict layer-wise LoRA-style deltas, which would give a per-layer view of where control is actually being applied—something the current paper does not show.
  • A testable extension outside meters: apply the same control-map-plus-hypernetwork recipe to other parameterizable rare defects (weld seams, rail cracks, turbine blades); the paper's claims are demonstrated only for cracks on substation meters, so transferability is an open question.
  • The reported drop in mAP50 when synthetic data exceeds 40% suggests the generator's output distribution is narrower than the real one; an editor's reading is that synthetic data should be treated as a regularizer, not a replacement for real examples.
  • The paper mentions multimodal-LLM judging but gives no systematic results; standardizing that evaluation could connect visual generation quality to detection utility, which is the metric that actually matters for deployment.

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

4 major / 5 minor

Summary. The manuscript proposes a few-shot image generation method for substation meter defects. It combines DreamBooth-style fine-tuning of Stable Diffusion (Section 3.1), geometric crack control maps with SAM-based region constraints (Section 3.2), and a hypernetwork-based conditional control branch (Section 3.3). The authors evaluate generation quality with FID/IS (Table 2) and downstream YOLOv8n detection performance on a new Substation Meter Dataset of 629 images (Table 1). They report best FID 76.72, IS 2.45, and an mAP50 increase from 0.587 to 0.778 when augmenting with 40% generated data.

Significance. The practical goal—using synthetic images to improve industrial defect detectors under severe data scarcity—is important, and the inclusion of a downstream detection evaluation is a strength. If the reported gains were reproducible and attributable to the proposed hypernetwork mechanism, the work would be a useful contribution to few-shot industrial defect synthesis. However, the central mechanism is described only verbally, and the quantitative claims are internally inconsistent. The paper also does not compare downstream detection performance against generated data from the other baselines. At present, the evidence is insufficient to support the stated contributions.

major comments (4)
  1. [§3.3, Eq. (8)] The hypernetwork is not specified. Section 3.3 describes it only verbally as "dynamically adjust[ing] generation parameters" and "combining" control latents with "hypernetwork-generated dynamic weights," but provides no architecture, no parameterization, no weight-modulation equation, and no training details. Eq. (8) is a generic conditional denoising loss that any ControlNet-style conditioning branch would satisfy, so it does not distinguish the proposed mechanism from existing conditional-control methods. The ablation in Table 3 also does not isolate the hypernetwork: the "Conditional Generation" row reports only FID/IS, and no downstream detection result is given for any variant without the hypernetwork. This is load-bearing because the claimed novelty and the reported controllability gains are attributed to this component.
  2. [Abstract vs. §5 and Table 1] The abstract states the method "reduces Frechet Inception Distance (FID) by 32.7%" and "boosts the mAP of a downstream defect detector by 15.3%." The conclusion in §5 reports FID 76.72 and a "19.1% mAP50 increase." Table 3 shows a 33.47% FID reduction relative to the baseline (115.31 → 76.72), while Table 1 shows mAP50 rising from 0.587 to 0.778, an absolute increase of 19.1 percentage points (32.5% relative). Neither the 32.7% FID number nor the 15.3% mAP number appears in any table. The reader cannot tell which figures are the definitive reported results; this inconsistency must be resolved in the abstract, tables, and text.
  3. [Table 1 vs. Table 2] Downstream detection is not evaluated for the generation baselines. Table 2 compares FID/IS of FastGAN, ProjectedGAN, DFMGAN, IDDPM, RDDM, and the proposed method, but Table 1 only reports YOLOv8n results when augmenting with the proposed method's generated images. The claims in §4.4 and §5 that the method "substantially outperforms existing augmentation and generation baselines" for practical detection are therefore unsupported. The same detection protocol should be applied to images generated by all baselines, or the claim should be limited to FID/IS.
  4. [Tables 1–3] No error bars, standard deviations, or multiple-seed runs are reported. The dataset contains only 100 defect images, split 80/20, so the differences in Table 1 (e.g., precision 0.967 vs. 0.698) may be within run-to-run variability. Furthermore, the +40% augmentation ratio appears to be selected post hoc from the five values evaluated; if so, this should be stated and the multiple-comparison issue acknowledged. Confidence intervals or at least results over several seeds are needed for the central claims.
minor comments (5)
  1. [Figures] Section 4.3 references "Fig. 6" and Section 4.5 references "Fig. 2," but the only figure caption present is for Fig. 1, whose caption says "overall comparisons of the proposed method with other SoTA methods," while the text in §4.3 uses it to illustrate Precision-Recall curves. Figures are missing or mislabeled.
  2. [Section 1 and §2.6] The introduction and related work promise MLLM-based evaluation (GPT-4, LLaMA, SEED1.5), but no such evaluation appears in Section 4. Either report the results or remove the promise.
  3. [Eqs. (1)–(3), (8)] Symbols such as ω_t, x̂_θ, x_pr, c_pr, and c+ are not defined. In Eq. (8), it is unclear how c+ is encoded and how control signals are injected at "multiple feature layers." Please define all notation.
  4. [References] The bibliography contains unattributed or irrelevant entries (e.g., [54] on the Afrikaans culture industry, [68] Pride and Prejudice) and many preprints not cited in the text. The reference list should be thoroughly pruned and checked.
  5. [Organization] Section 1 states "Section 4.6 discusses the findings and concludes the paper," but Section 4.6 is a short discussion and Section 5 is the conclusion. Also, the phrase "increases diversity metrics" in the abstract is not matched by any explicit diversity metric beyond IS.

Circularity Check

0 steps flagged

No circularity found; the hypernetwork is underspecified, but that is a reproducibility/correctness issue, not a circular derivation.

full rationale

The paper's derivation chain is empirical rather than definitional. Its three components (DreamBooth-style knowledge embedding, crack-mask control maps, and a conditional loss in Eq. 8) are described as standard building blocks with generic diffusion losses, and the reported improvements are measured against external baselines and independent metrics (FID, IS, mAP50) using a separate downstream detector. Nothing in the paper fits a parameter to the target outcome and then re-presents that fit as a prediction: for example, Table 1 varies the proportion of generated data and reports detector mAP, but the 40% augmentation ratio is selected from that same comparison, which is a soundness/selection issue, not a reduction of the claim to the input. The skeptic's concern that the 'hypernetwork' is never specified (no architecture, no parameterization, no weight-modulation equation, with Eq. 8 being only a generic conditional denoising objective) is a legitimate reproducibility and attribution concern: it means the reported gains cannot be attributed to the proposed mechanism with the evidence given. However, that is a missing-support/correctness critique, not circularity in the sense of a claimed result being equivalent to its inputs by construction. There is also no load-bearing self-citation: the reference list contains no prior work by the authors that is invoked to force the framework, and no uniqueness theorem or ansatz is imported from self-authored sources. Accordingly, the honest finding is no significant circularity, score 0.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced; the modules are standard neural-network components. The SMD dataset is a new dataset but is not released, so it provides no independent evidence. The central claim rests on unstated implementation details and an asserted but unspecified hypernetwork.

free parameters (3)
  • prior-preservation weight lambda = not reported
    Eq. (3) balances subject loss and prior-preservation loss; its value controls generation fidelity vs. diversity and is never specified.
  • generated-data augmentation ratio = 40%
    Table 1 shows mAP50 peaks at +40% generated data and declines at higher ratios; the headline downstream gain depends on this post hoc selected proportion.
  • diffusion/control fine-tuning hyperparameters = not reported
    Learning rate, steps, batch size, hypernetwork width/depth, and edge-detection thresholds are unstated; the method cannot be reproduced without them.
axioms (4)
  • domain assumption Stable Diffusion's pre-trained weights, after DreamBooth-style fine-tuning on ~100 meter images, retain a general generative prior while learning meter identity.
    Section 3.1 assumes a small fine-tuning set is sufficient for domain embedding; overfitting would invalidate the diversity/FID claims.
  • domain assumption SAM reliably segments the meter dial from a prompt on this industrial image domain.
    Section 3.2 Eqs. (5)-(7) use SAM's mask I_r to constrain crack placement; inaccurate masks would misplace defects.
  • ad hoc to paper The hypernetwork described in Section 3.3 exists and functionally modulates the frozen denoiser.
    No architecture, parameterization, or modulating equation is given; Eq. (8) is a standard diffusion L2 loss. The mechanism is asserted, not defined.
  • domain assumption FID/IS and YOLOv8n mAP on the validation split are stable indicators of quality without repeated runs or error bars.
    Tables 1-3 report single numbers; the non-monotonic Table 1 trend suggests instability.

pith-pipeline@v1.3.0-alltime-deepseek · 12908 in / 17531 out tokens · 162806 ms · 2026-08-03T10:39:22.599837+00:00 · methodology

0 comments
read the original abstract

Substation meters play a critical role in monitoring and ensuring the stable operation of power grids, yet their detection of cracks and other physical defects is often hampered by a severe scarcity of annotated samples. To address this few-shot generation challenge, we propose a novel framework that integrates Knowledge Embedding and Hypernetwork-Guided Conditional Control into a Stable Diffusion pipeline, enabling realistic and controllable synthesis of defect images from limited data. First, we bridge the substantial domain gap between natural-image pre-trained models and industrial equipment by fine-tuning a Stable Diffusion backbone using DreamBooth-style knowledge embedding. This process encodes the unique structural and textural priors of substation meters, ensuring generated images retain authentic meter characteristics. Second, we introduce a geometric crack modeling module that parameterizes defect attributes--such as location, length, curvature, and branching pattern--to produce spatially constrained control maps. These maps provide precise, pixel-level guidance during generation. Third, we design a lightweight hypernetwork that dynamically modulates the denoising process of the diffusion model in response to the control maps and high-level defect descriptors, achieving a flexible balance between generation fidelity and controllability. Extensive experiments on a real-world substation meter dataset demonstrate that our method substantially outperforms existing augmentation and generation baselines. It reduces Frechet Inception Distance (FID) by 32.7%, increases diversity metrics, and--most importantly--boosts the mAP of a downstream defect detector by 15.3% when trained on augmented data. The framework offers a practical, high-quality data synthesis solution for industrial inspection systems where defect samples are rare.

Figures

Figures reproduced from arXiv: 2601.09238 by Jackie Alex, Justin Petter.

Figure 1
Figure 1. Figure 1: The overall comparisons of the proposed method with other SoTA [PITH_FULL_IMAGE:figures/full_fig_p003_1.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

82 extracted references · 34 linked inside Pith

  1. [1]

    Setransformer: A hybrid attention-based architecture for robust human activity recognition,

    Y. Liu, X. Qin, Y. Gao, X. Li, and C. Feng, “Setransformer: A hybrid attention-based architecture for robust human activity recognition,”INNO- PRESS: Journal of Emerging Applied AI, vol. 1, no. 1, 2025

    2025

  2. [2]

    Gated multimodal graph learning for personalized recommendation,

    S. Liu, Y. Zhang, X. Li, Y. Liu, C. Feng, and H. Yang, “Gated multimodal graph learning for personalized recommendation,”INNO-PRESS: Journal of Emerging Applied AI, vol. 1, no. 1, 2025

    2025

  3. [3]

    Extraction/conversion of geometric dimensions and tolerances for machining features,

    J. Gao, D. T. Zheng, N. Gindy, and D. Clark, “Extraction/conversion of geometric dimensions and tolerances for machining features,”International Journal of Advanced Manufacturing Technology, vol. 26, no. 4, pp. 405–414, Aug 2005

    2005

  4. [4]

    A datum-based model for practicing geometric di- mensioning and tolerancing,

    W. Sun and Y. Gao, “A datum-based model for practicing geometric di- mensioning and tolerancing,”Journal of Engineering Technology, vol. 35, pp. 38–47, Sep 2018

    2018

  5. [5]

    Tolerance information extraction for mechanical engineer- ing drawings: A digital image processing and deep learning-based model,

    Y. Xuet al., “Tolerance information extraction for mechanical engineer- ing drawings: A digital image processing and deep learning-based model,” CIRP Journal of Manufacturing Science and Technology, vol. 50, pp. 55–64, Jun 2024

    2024

  6. [6]

    Inte- gration of deep learning for automatic recognition of 2D engineering draw- ings,

    Y.-H. Lin, Y.-H. Ting, Y.-C. Huang, K.-L. Cheng, and W.-R. Jong, “Inte- gration of deep learning for automatic recognition of 2D engineering draw- ings,”Machines, vol. 11, no. 8, Aug 2023

    2023

  7. [7]

    A systematic review of machine learn- ing applications in infectious disease prediction, diagnosis, and outbreak forecasting,

    Y. Wang, J. Zhong, and R. Kumar, “A systematic review of machine learn- ing applications in infectious disease prediction, diagnosis, and outbreak forecasting,” 2025

    2025

  8. [8]

    Medical graph rag: Towards safe medical large language model via graph retrieval-augmented generation,

    J. Wu, J. Zhu, Y. Qi, J. Chen, M. Xu, F. Menolascina, and V. Grau, “Medical graph rag: Towards safe medical large language model via graph retrieval-augmented generation,” 2024. [Online]. Available: https://arxiv.org/abs/2408.04187

    Pith/arXiv arXiv 2024

  9. [9]

    Enhancing thyroid disease prediction using ma- chine learning: A comparative study of ensemble models and class balanc- ing techniques,

    J. Zhong and Y. Wang, “Enhancing thyroid disease prediction using ma- chine learning: A comparative study of ensemble models and class balanc- ing techniques,” 2025

    2025

  10. [10]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,”Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 10 684–10 695, 2022

    2022

  11. [11]

    From local to global: A graph rag approach to query-focused summarization,

    D. Edge, H. Trinh, N. Cheng, J. Bradley, A. Chao, A. Mody, S. Truitt, and J. Larson, “From local to global: A graph rag approach to query-focused summarization,”arXiv preprint arXiv:2404.16130, 2024. 16

    Pith/arXiv arXiv 2024

  12. [12]

    Lightrag: Simple and fast retrieval-augmented generation,

    Z. Guo, L. Xia, Y. Yu, T. Ao, and C. Huang, “Lightrag: Simple and fast retrieval-augmented generation,” 2024. [Online]. Available: https://arxiv.org/abs/2410.05779

    Pith/arXiv arXiv 2024

  13. [13]

    Communitykg-rag: Leveraging community structures in knowledge graphs for advanced retrieval-augmented generation in fact-checking,

    R.-C. Chang and J. Zhang, “Communitykg-rag: Leveraging community structures in knowledge graphs for advanced retrieval-augmented generation in fact-checking,” 2024. [Online]. Available: https://arxiv.org/ abs/2408.08535

    Pith/arXiv arXiv 2024

  14. [14]

    Tabpedia: Towards comprehensive visual table under- standing with concept synergy,

    W. Zhao, H. Feng, Q. Liu, J. Tang, B. Wu, L. Liao, S. Wei, Y. Ye, H. Liu, W. Zhouet al., “Tabpedia: Towards comprehensive visual table under- standing with concept synergy,”Advances in Neural Information Process- ing Systems, vol. 37, pp. 7185–7212, 2025

    2025

  15. [15]

    Doc- pedia: Unleashing the power of large multimodal model in the frequency domain for versatile document understanding,

    H. Feng, Q. Liu, H. Liu, J. Tang, W. Zhou, H. Li, and C. Huang, “Doc- pedia: Unleashing the power of large multimodal model in the frequency domain for versatile document understanding,”Science China Information Sciences, vol. 67, no. 12, pp. 1–14, 2024

    2024

  16. [16]

    Dolphin: Document image parsing via heterogeneous anchor prompting,

    H. Feng, S. Wei, X. Fei, W. Shi, Y. Han, L. Liao, J. Lu, B. Wu, Q. Liu, C. Linet al., “Dolphin: Document image parsing via heterogeneous anchor prompting,”arXiv preprint arXiv:2505.14059, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    A bounding box is worth one token: Interleaving layout and text in a large language model for document understanding,

    J. Lu, H. Yu, Y. Wang, Y. Ye, J. Tang, Z. Yang, B. Wu, Q. Liu, H. Feng, H. Wanget al., “A bounding box is worth one token: Interleaving layout and text in a large language model for document understanding,”arXiv preprint arXiv:2407.01976, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Textsquare: Scaling up text-centric visual instruction tun- ing,

    J. Tang, C. Lin, Z. Zhao, S. Wei, B. Wu, Q. Liu, H. Feng, Y. Li, S. Wang, L. Liaoet al., “Textsquare: Scaling up text-centric visual instruction tun- ing,”arXiv preprint arXiv:2404.12803, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Mtvqa: Benchmarking multilingual text-centric visual question answering,

    J. Tang, Q. Liu, Y. Ye, J. Lu, S. Wei, C. Lin, W. Li, M. F. F. B. Mahmood, H. Feng, Z. Zhaoet al., “Mtvqa: Benchmarking multilingual text-centric visual question answering,”arXiv preprint arXiv:2405.11985, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Pargo: Bridging vision-language with partial and global views,

    A.-L. Wang, B. Shan, W. Shi, K.-Y. Lin, X. Fei, G. Tang, L. Liao, J. Tang, C. Huang, and W.-S. Zheng, “Pargo: Bridging vision-language with partial and global views,” vol. 39, no. 7, pp. 7491–7499, 2025

    2025

  21. [21]

    Mctbench: Multimodal cognition towards text-rich visual scenes benchmark,

    B. Shan, X. Fei, W. Shi, A.-L. Wang, G. Tang, L. Liao, J. Tang, X. Bai, and C. Huang, “Mctbench: Multimodal cognition towards text-rich visual scenes benchmark,”arXiv preprint arXiv:2410.11538, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Judging llm-as-a-judge with mt-bench and chat- bot arena,

    L. Zheng, W.-L. Chiang, Y. Sheng, S. Zhuang, Z. Wu, Y. Zhuang, Z. Lin, Z. Li, D. Li, E. Xinget al., “Judging llm-as-a-judge with mt-bench and chat- bot arena,”Advances in Neural Information Processing Systems, vol. 36, pp. 46 595–46 623, 2023. 17

    2023

  23. [23]

    Is chatgpt a good nlg evaluator? a preliminary study,

    J. Wang, Y. Liang, F. Meng, Z. Sun, H. Shi, Z. Li, J. Xu, J. Qu, and J. Zhou, “Is chatgpt a good nlg evaluator? a preliminary study,”arXiv preprint arXiv:2303.04048, 2023

    Pith/arXiv arXiv 2023

  24. [24]

    Gpt-4 techni- cal report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “Gpt-4 techni- cal report,”arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    The llama 3 herd of models,

    A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Yang, A. Fanet al., “The llama 3 herd of models,”arXiv preprint arXiv:2407.21783, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    Seed1. 5-vl technical report,

    D. Guo, F. Wu, F. Zhu, F. Leng, G. Shi, H. Chen, H. Fan, J. Wang, J. Jiang, J. Wanget al., “Seed1. 5-vl technical report,”arXiv preprint arXiv:2505.07062, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    Self-preference bias in llm-as-a- judge,

    K. Wataoka, T. Takahashi, and R. Ri, “Self-preference bias in llm-as-a- judge,”arXiv preprint arXiv:2410.21819, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020

    2020

  29. [29]

    Attentive eraser: Unleash- ing diffusion model’s object removal potential via self-attention redirection guidance,

    W. Sun, X.-M. Dong, B. Cui, and J. Tang, “Attentive eraser: Unleash- ing diffusion model’s object removal potential via self-attention redirection guidance,” vol. 39, no. 19, pp. 20 734–20 742, 2025

    2025

  30. [30]

    Cme-cad: Heterogeneous collaborative multi-expert reinforce- ment learning for cad code generation,

    K. Niu, H. Yu, Z. Chen, Z. Yao, W. Jia, X. Ge, J. Tang, B. Cui, B. Li, and X. Xue, “Cme-cad: Heterogeneous collaborative multi-expert reinforce- ment learning for cad code generation,”arXiv preprint arXiv:2512.23333, 2025

    arXiv 2025

  31. [31]

    Mindev: Multi-modal integrated diffusion framework for video reconstruction from eeg signals,

    S. Huang, Y. Wang, H. Luo, H. Jing, C. Qin, and J. Tang, “Mindev: Multi-modal integrated diffusion framework for video reconstruction from eeg signals,” pp. 3350–3359, 2025

    2025

  32. [32]

    Meml-grpo: Heterogeneous multi-expert mutual learning for rlvr advancement,

    W. Jia, J. Lu, H. Yu, S. Wang, G. Tang, A.-L. Wang, W. Yin, D. Yang, Y. Nie, B. Shanet al., “Meml-grpo: Heterogeneous multi-expert mutual learning for rlvr advancement,”arXiv preprint arXiv:2508.09670, 2025

    arXiv 2025

  33. [33]

    Fine-tuning vision-language model for automated engineering drawing in- formation extraction,

    M. T. Khan, L. Chen, Y. H. Ng, W. Feng, N. Y. J. Tan, and S. K. Moon, “Fine-tuning vision-language model for automated engineering drawing in- formation extraction,” 2024, preprint

    2024

  34. [34]

    Few could be better than all: Feature sampling and grouping for scene text de- tection,

    J. Tang, W. Zhang, H. Liu, M. Yang, B. Jiang, G. Hu, and X. Bai, “Few could be better than all: Feature sampling and grouping for scene text de- tection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 4563–4572. 18

    2022

  35. [35]

    Optimal boxes: boosting end-to-end scene text recognition by adjusting annotated bound- ing boxes via reinforcement learning,

    J. Tang, W. Qian, L. Song, X. Dong, L. Li, and X. Bai, “Optimal boxes: boosting end-to-end scene text recognition by adjusting annotated bound- ing boxes via reinforcement learning,” inEuropean Conference on Com- puter Vision. Springer, 2022, pp. 233–248

    2022

  36. [36]

    Spts v2: single-point scene text spotting,

    Y. Liu, J. Zhang, D. Peng, M. Huang, X. Wang, J. Tang, C. Huang, D. Lin, C. Shen, X. Baiet al., “Spts v2: single-point scene text spotting,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    2023

  37. [37]

    You can even annotate text with voice: Transcription-only-supervised text spotting,

    J. Tang, S. Qiao, B. Cui, Y. Ma, S. Zhang, and D. Kanoulas, “You can even annotate text with voice: Transcription-only-supervised text spotting,” inProceedings of the 30th ACM International Conference on Multimedia, ser. MM ’22. New York, NY, USA: Association for Computing Machinery, 2022, pp. 4154–4163. [Online]. Available: https://doi.org/10.1145/350316...

    arXiv 2022

  38. [38]

    Unidoc: A universal large multimodal model for simultaneous text detection, recog- nition, spotting and understanding,

    H. Feng, Z. Wang, J. Tang, J. Lu, W. Zhou, H. Li, and C. Huang, “Unidoc: A universal large multimodal model for simultaneous text detection, recog- nition, spotting and understanding,”arXiv preprint arXiv:2308.11592, 2023

    Pith/arXiv arXiv 2023

  39. [39]

    Wilddoc: How far are we from achieving compre- hensive and robust document understanding in the wild?

    A.-L. Wang, J. Tang, L. Lei, H. Feng, Q. Liu, X. Fei, J. Lu, H. Wang, W. Liu, H. Liuet al., “Wilddoc: How far are we from achieving compre- hensive and robust document understanding in the wild?”arXiv preprint arXiv:2505.11015, 2025

    Pith/arXiv arXiv 2025

  40. [40]

    Paddleocr: A versatile ocr toolkit with 80+ languages recog- nition,

    P. Authors, “Paddleocr: A versatile ocr toolkit with 80+ languages recog- nition,” 2023

    2023

  41. [41]

    tesseract-ocr/tesseract,

    tesseract-ocr, “tesseract-ocr/tesseract,” https://github.com/tesseract-ocr/ tesseract, 2024, accessed: 2024-09-27

    2024

  42. [42]

    Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression,

    H. Jiang, Q. Wu, X. Luo, D. Li, C.-Y. Lin, Y. Yang, and L. Qiu, “Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression,” 2024. [Online]. Available: https://arxiv.org/abs/2310.06839

    Pith/arXiv arXiv 2024

  43. [43]

    Introducing contextual retrieval,

    Anthropic, “Introducing contextual retrieval,” https://www.anthropic. com/news/contextual-retrieval, 2024, accessed: 2024-11-02

    2024

  44. [44]

    Late chunk- ing: contextual chunk embeddings using long-context embedding models,

    M. G¨ unther, I. Mohr, D. J. Williams, B. Wang, and H. Xiao, “Late chunk- ing: contextual chunk embeddings using long-context embedding models,” arXiv preprint arXiv:2409.04701, 2024

    Pith/arXiv arXiv 2024

  45. [45]

    Resolving evi- dence sparsity: Agentic context engineering for long-document understand- ing,

    K. Liu, Z. Chen, M. Li, J. Tang, D. Yang, and L. Zhang, “Resolving evi- dence sparsity: Agentic context engineering for long-document understand- ing,”arXiv preprint arXiv:2511.22850, 2025. 19

    arXiv 2025

  46. [46]

    Llms for low resource languages in multilingual, multimodal and dialectal settings,

    F. Alam, S. A. Chowdhury, S. Boughorbel, and M. Hasanain, “Llms for low resource languages in multilingual, multimodal and dialectal settings,” inConference of the European Chapter of the Association for Computational Linguistics, 2024. [Online]. Available: https://api.semanticscholar.org/CorpusID:268417133

    2024

  47. [47]

    Translation and fusion improves zero-shot cross-lingual information extraction,

    Y. Chen, V. Shah, and A. Ritter, “Translation and fusion improves zero-shot cross-lingual information extraction,” 2024. [Online]. Available: https://arxiv.org/abs/2305.13582

    Pith/arXiv arXiv 2024

  48. [48]

    Ocrbench v2: An improved benchmark for evaluating large multimodal models on visual text localization and reasoning,

    L. Fu, B. Yang, Z. Kuang, J. Song, Y. Li, L. Zhu, Q. Luo, X. Wang, H. Lu, M. Huanget al., “Ocrbench v2: An improved benchmark for evaluating large multimodal models on visual text localization and reasoning,”arXiv preprint arXiv:2501.00321, 2024

    Pith/arXiv arXiv 2024

  49. [49]

    Character recognition competition for street view shop signs,

    J. Tang, W. Du, B. Wang, W. Zhou, S. Mei, T. Xue, X. Xu, and H. Zhang, “Character recognition competition for street view shop signs,”National Science Review, vol. 10, no. 6, p. nwad141, 2023

    2023

  50. [50]

    Benchmarking vision-language models on chinese ancient documents: From ocr to knowledge reasoning,

    H. Yu, Y. Wu, F. Shi, L. Liao, J. Lu, X. Ge, H. Wang, M. Zhuo, X. Wu, X. Feiet al., “Benchmarking vision-language models on chinese ancient documents: From ocr to knowledge reasoning,”arXiv preprint arXiv:2509.09731, 2025

    Pith/arXiv arXiv 2025

  51. [51]

    Blind image quality assessment via vision-language correspondence: A multitask learning per- spective,

    W. Zhang, G. Zhai, Y. Wei, X. Yang, and K. Ma, “Blind image quality assessment via vision-language correspondence: A multitask learning per- spective,” pp. 14 071–14 081, 2023

    2023

  52. [52]

    Jack and the beanstalk: Towards question answering in plant biology

    G. Jhajj and Y. Nomura, “Jack and the beanstalk: Towards question answering in plant biology.” [Online]. Available: https: //api.semanticscholar.org/CorpusID:274567831

  53. [53]

    Enhancing knowledge graph construction using large language models,

    M. Trajanoska, R. Stojanov, and D. Trajanov, “Enhancing knowledge graph construction using large language models,” 2023. [Online]. Available: https://arxiv.org/abs/2305.04676

    Pith/arXiv arXiv 2023

  54. [54]

    Afrikaans, inc.: the afrikaans culture industry after apartheid,

    A. S. Steyn, “Afrikaans, inc.: the afrikaans culture industry after apartheid,”Social Dynamics, vol. 42, pp. 481–503, 2016. [Online]. Available: https://api.semanticscholar.org/CorpusID:152269054

    2016

  55. [55]

    AutoCAD mechanical 2022 help — about balloons (autocad mechanical toolset) — autodesk,

    “AutoCAD mechanical 2022 help — about balloons (autocad mechanical toolset) — autodesk,” https://help.autodesk.com/view/AMECH PP/ 2022/ENU/?guid=GUID-F12F0EA0-0810-42EE-A3FE-327041AF AEEE, 2022, accessed: 2024-09-27

    2022

  56. [56]

    Leading image & video data annotation platform CV AT,

    “Leading image & video data annotation platform CV AT,” https://www. cvat.ai, 2024, accessed: 2025-03-22. 20

    2024

  57. [57]

    Advancing sequential numerical prediction in autoregressive models,

    X. Fei, J. Lu, Q. Sun, H. Feng, Y. Wang, W. Shi, A.-L. Wang, J. Tang, and C. Huang, “Advancing sequential numerical prediction in autoregressive models,”arXiv preprint arXiv:2505.13077, 2025

    Pith/arXiv arXiv 2025

  58. [58]

    Retrieval-augmented generation for large language models: A survey,

    Y. Gao, Y. Xiong, X. Gao, K. Jia, J. Pan, Y. Bi, Y. Dai, J. Sun, and H. Wang, “Retrieval-augmented generation for large language models: A survey,”arXiv preprint arXiv:2312.10997, 2023

    Pith/arXiv arXiv 2023

  59. [59]

    How good are llms at relation extraction under low-resource scenario? comprehensive evaluation,

    D. Jinensibieke, M. Maimaiti, W. Xiao, Y. Zheng, and X. Wang, “How good are llms at relation extraction under low-resource scenario? comprehensive evaluation,” 2024. [Online]. Available: https://arxiv.org/abs/2406.11162

    Pith/arXiv arXiv 2024

  60. [60]

    Retrieval- augmented generation for knowledge-intensive nlp tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. K¨ uttler, M. Lewis, W.-t. Yih, T. Rockt¨ aschelet al., “Retrieval- augmented generation for knowledge-intensive nlp tasks,”Advances in Neu- ral Information Processing Systems, vol. 33, pp. 9459–9474, 2020

    2020

  61. [61]

    Long-context llms struggle with long in-context learning,

    T. Li, G. Zhang, Q. D. Do, X. Yue, and W. Chen, “Long-context llms struggle with long in-context learning,” 2024. [Online]. Available: https://arxiv.org/abs/2404.02060

    Pith/arXiv arXiv 2024

  62. [62]

    Audio and multiscale visual cues driven cross-modal transformer for idling vehicle detection,

    X. Li, R. Whitaker, and T. Tasdizen, “Audio and multiscale visual cues driven cross-modal transformer for idling vehicle detection,”arXiv preprint arXiv:2504.16102, 2025

    arXiv 2025

  63. [63]

    Joint audio-visual idling vehicle detection with streamlined input dependencies,

    X. Li, R. Mohammed, T. Mangin, S. Saha, K. Kelly, R. Whitaker, and T. Tasdizen, “Joint audio-visual idling vehicle detection with streamlined input dependencies,” inProceedings of the Winter Conference on Applica- tions of Computer Vision, 2025, pp. 885–894

    2025

  64. [64]

    Real-time idling vehicles detection using combined audio-visual deep learning,

    X. Li, T. Mangin, S. Saha, R. Mohammed, E. Blanchard, D. Tang, H. Poppe, O. Choi, K. Kelly, and R. Whitaker, “Real-time idling vehicles detection using combined audio-visual deep learning,” inEmerging Cutting- Edge Developments in Intelligent Traffic and Transportation Systems. IOS Press, 2024, pp. 142–158

    2024

  65. [65]

    Prolonged reasoning is not all you need: Certainty- based adaptive routing for efficient llm/mllm reasoning,

    J. Lu, H. Yu, S. Xu, S. Ran, G. Tang, S. Wang, B. Shan, T. Fu, H. Feng, J. Tanget al., “Prolonged reasoning is not all you need: Certainty- based adaptive routing for efficient llm/mllm reasoning,”arXiv preprint arXiv:2505.15154, 2025

    Pith/arXiv arXiv 2025

  66. [66]

    Data management and SPC software,

    “Data management and SPC software,” https://measurlink.com/, ac- cessed: 2024-09-27

    2024

  67. [67]

    torchvision.transforms torchvision master documentation,

    “torchvision.transforms torchvision master documentation,” https:// pytorch.org/vision/0.9/transforms.html, accessed: 2025-03-22

    2025

  68. [68]

    Austen,Pride and Prejudice

    J. Austen,Pride and Prejudice. Urbana, Illinois: Project Gutenberg,

  69. [69]

    You only look once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” May 2016, preprint

    2016

  70. [70]

    Hermes 3 technical report,

    R. Teknium, J. Quesnelle, and C. Guang, “Hermes 3 technical report,”

  71. [71]

    Rag based question-answering for contextual response prediction system,

    S. Veturi, S. Vaichal, R. L. Jagadheesh, N. I. Tripto, and N. Yan, “Rag based question-answering for contextual response prediction system,”

  72. [72]

    Vision as lora,

    H. Wang, Y. Ye, B. Li, Y. Nie, J. Lu, J. Tang, Y. Wang, and C. Huang, “Vision as lora,”arXiv preprint arXiv:2503.20680, 2025

    Pith/arXiv arXiv 2025

  73. [73]

    Im- proving text embeddings with large language models,

    L. Wang, N. Yang, X. Huang, L. Yang, R. Majumder, and F. Wei, “Im- proving text embeddings with large language models,”arXiv preprint arXiv:2401.00368, 2023

    Pith/arXiv arXiv 2023

  74. [74]

    Available: https://arxiv.org/abs/2409.03708

    [Online]. Available: https://arxiv.org/abs/2409.03708

    Pith/arXiv arXiv

  75. [75]

    Evaluating supervised learning models for fraud detection: A comparative study of classical and deep architectures on imbalanced transaction data,

    C. Wang, C. Nie, and Y. Liu, “Evaluating supervised learning models for fraud detection: A comparative study of classical and deep architectures on imbalanced transaction data,”arXiv preprint arXiv:2505.22521, 2025

    arXiv 2025

  76. [76]

    Fine-grained heartbeat waveform mon- itoring with rfid: A latent diffusion model,

    Y. Wang, T. Zhao, and X. Wang, “Fine-grained heartbeat waveform mon- itoring with rfid: A latent diffusion model,” pp. 86–91, 2025

    2025

  77. [77]

    Enhancing code llms with reinforcement learning in code generation,

    J. Wang, Z. Zhang, Y. He, Y. Song, T. Shi, Y. Li, H. Xu, K. Wu, G. Qian, Q. Chenet al., “Enhancing code llms with reinforcement learning in code generation,”arXiv preprint arXiv:2412.20367, 2024

    Pith/arXiv arXiv 2024

  78. [78]

    Multi-modal in-context learning makes an ego-evolving scene text recognizer,

    Z. Zhao, J. Tang, C. Lin, B. Wu, C. Huang, H. Liu, X. Tan, Z. Zhang, and Y. Xie, “Multi-modal in-context learning makes an ego-evolving scene text recognizer,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 567–15 576

    2024

  79. [79]

    Ragas: Automated evaluation of retrieval augmented generation,

    S. Es, J. James, L. Espinosa-Anke, and S. Schockaert, “Ragas: Automated evaluation of retrieval augmented generation,” 2023. [Online]. Available: https://arxiv.org/abs/2309.15217 22

    Pith/arXiv arXiv 2023

  80. [80]

    Harmonizing visual text comprehension and gen- eration,

    Z. Zhao, J. Tang, B. Wu, C. Lin, S. Wei, H. Liu, X. Tan, Z. Zhang, C. Huang, and Y. Xie, “Harmonizing visual text comprehension and gen- eration,”arXiv preprint arXiv:2407.16364, 2024

    Pith/arXiv arXiv 2024

Showing first 80 references.