Pith. sign in

REVIEW 5 major objections 5 minor 33 references

VisuCraft: Enhancing Large Vision-Language Models for Complex Visual-Guided Creative Content Generation via Structured Information Extraction

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read VisuCraft shows that a structured visual-information extractor plus a dynamic prompt composer lets a frozen vision-language model generate more creative, visually grounded, instruction-following stories and poems.

desk verdict Plausible prompt-engineering idea, but the evaluation is too thin to support the claims: unreleased benchmark, undefined metrics, vague baselines, and a pipeline that may never show the image to the LVLM. read the letter →

arxiv 2508.02890 v1 pith:ZX2WRCFV submitted 2025-08-04 cs.CV cs.CL

Rongxin Jiang , Robert Long , Chenghao Gu , Mingrui Yan This is my paper
classification cs.CVcs.CL
keywords largevision-languagemodelscreativecontentgenerationstructuredinformationextractiondynamicpromptvisualgroundinginstructionadherencestorypoetrycomposition
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

The paper tries to establish that a large vision-language model (LVLM) does not need to be retrained to generate long, creative, visually grounded text: a lightweight front-end that extracts structured visual information and rewrites the user's instruction into an optimized prompt is enough. VisuCraft consists of a multimodal structured information extractor (E) and a dynamic prompt generation module (G), which together convert an input image and a user instruction into a single prompt for a frozen LVLM such as LLaVA or InstructBLIP. On a self-built benchmark of 500,000 image-instruction pairs, the paper reports consistent gains over two baselines, with mean scores of 0.822 for story generation and 0.810 for poetry, and the largest improvements in creativity and instruction adherence. If the claim holds, it offers a cheap, modular path to better creative AI output from existing models.

What carries the argument

The load-bearing machinery is the two-module front-end wrapped around a fixed LVLM. E, the multimodal structured information extractor, turns the raw image into a rich structured representation (typically JSON) of fine-grained visual attributes; G, the dynamic prompt generation module, fuses that representation with the user's instruction through integration, prioritization, and contextualization to produce the final prompt. The chain $T = M(G(E(I), U))$ is what carries the argument: every reported improvement is attributed to changing the prompt fed to an unchanged model. The granularity analysis makes the mechanism concrete by showing that moving from object lists to objects-plus-attributes to full structured information raises the StoryGen mean from 0.769 to 0.800 to 0.822.

What would settle it

An independent replication that releases the ImageStoryGen-500K subset, the VisuGen metric implementation, and the exact prompts, and then re-runs LVLM-Base, LVLM-Enhanced, and VisuCraft with the same underlying LVLM, would settle the claim: if the StoryGen mean gain of 0.822 over 0.781 does not reproduce (or shrinks to noise under random seeds), the central claim collapses.

Watch

Extended reading notes

Core claim

VisuCraft's central claim is that the quality bottleneck in visual-guided creative text generation lies less in the generative model itself than in how visual information and instructions are presented to it. The framework computes $V = E(I)$ — a structured, JSON-like description of the image capturing object poses, materials, lighting, spatial relations, and even emotional atmosphere — then computes $P = G(V, U)$ to integrate, prioritize, and contextualize that description with the user's instruction, and finally generates $T = M(P)$ with the base LVLM left unchanged. The reported experiments show this chain improving all three VisuGen metrics over both an unenhanced and an enhanced LVLM baseline across story generation, poetry composition, and advertising copy, with the largest margins in creativity and instruction adherence. The paper also argues, through ablations and a granularity analysis, that both components matter: replacing E with a generic visual encoder or a plain object list, or replacing G with simple concatenation, degrades the mean score.

Load-bearing premise

The whole comparison rests on trusting that the self-built 500,000-pair benchmark, the VisuGen scoring, and the two baseline models are honest and fair, since the paper does not release the dataset, the metric code, the prompts, or the model weights.

Editorial extensions

If this is right

  • A user of an existing LVLM can expect better visual grounding, creativity, and instruction adherence by adopting a structured-extraction-plus-prompt-composition layer, without any weight updates to the text generator.
  • The same frozen model is claimed to improve across story generation, poetry composition, and advertising copy, so the benefit is not tied to a single creative genre.
  • The ablation results imply that simply adding more visual text to a prompt is not enough: the information must be structured and the prompt must be dynamically composed to realize the gains.
  • The granularity trend predicts that investing in richer fine-grained annotations (attributes, relations, lighting, atmosphere) pays off in generation quality, not just in perception tasks.
  • Human ratings on a 200-sample subset align with the automatic metrics, so the reported gains are not purely an artifact of the VisuGen scoring scheme.

Reading between the lines

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

  • If the claim holds, prompt-side restructuring could serve as a cheap substitute for fine-tuning in creative tasks, and the same extractor-and-composer front-end could be adapted to other modalities such as video or audio. This is an extension the paper only gestures at in its future-work section.
  • The modular design implies a low-cost domain-adaptation path: retrain only the small extractor E and prompt module G on a new visual domain while keeping the base LVLM frozen; the paper does not test this explicitly.
  • A direct stress test of the model-agnosticism claim would be to swap the base LVLM for a different architecture or size and verify that the same gain profile appears; the paper reports results using architectures similar to LLaVA or InstructBLIP only.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. VisuCraft proposes a modular enhancement for large vision-language models (LVLMs) in which a structured information extractor E converts an input image into a JSON-like representation, a dynamic prompt generation module G combines that representation with the user instruction into an optimized text prompt, and a pre-trained LVLM M generates the final long-form creative text. The authors claim consistent improvements over two baselines (LVLM-Base and LVLM-Enhanced) on a self-constructed benchmark ImageStoryGen-500K, using custom VisuGen metrics (Visual Grounding, Creativity, Instruction Adherence), plus ablations, granularity analysis, human evaluation, and qualitative examples. The central empirical claim is that prompt-side restructuring alone, without end-to-end training, improves visual grounding, creativity, and instruction adherence for story generation, poetry, and advertising copy.

Significance. If the empirical results were trustworthy, VisuCraft would present a lightweight, model-agnostic way to boost LVLMs' creative text generation, which is practically appealing because it avoids retraining the large model. The modular design is clearly described, and the idea of enriching prompts with structured visual attributes is reasonable. However, the manuscript as written provides no verifiable evidence for its claims: the evaluation dataset is unreleased, the VisuGen metrics have no operational definitions, the baselines are only vaguely specified, the human evaluation lacks inter-annotator reliability, and the formal pipeline in Eq. (4) appears to pass only text to the generator, undermining the 'vision-language' claim. The significance is therefore conditional on a complete re-validation; in its current form, the contribution cannot be assessed.

major comments (5)
  1. [Section IV-A (Metrics)] The VisuGen metrics (VG., C., IA.) are defined only in one-sentence prose in Section IV-A, with no scoring protocol, rubric, computational formula, or validation against human judgments. Since Tables I–III report only these metrics, the quantitative results cannot be reproduced or interpreted. The authors must either provide an exact operational definition (e.g., using established metrics like CLIPScore for visual grounding, self-BLEU or distinct-n for creativity, and instruction-overlap metrics for adherence) or release the detailed rubric, annotation instructions, and raw scores used to compute VG., C., and IA.
  2. [Section IV-A (Baselines)] LVLM-Base and LVLM-Enhanced are not concretely specified: the paper only says they use 'architectures similar to LLaVA or InstructBLIP' and that LVLM-Enhanced 'might incorporate' fine-tuning or better prompting. No model checkpoints, prompt templates, decoding parameters, or training data are given. The reported margins (e.g., 0.822 vs. 0.811 StoryGen Mean in Table I) could easily arise from under-prompted or poorly tuned baselines. A fair comparison requires exact baseline configurations, ideally using official released checkpoints with standardized prompts, and reporting multiple random seeds with variance.
  3. [Section III-C, Eq. (4)] The generation formula T = M(G(E(I), U)) passes only the text prompt P into M; the image I appears only inside E and is not an input tensor to the generator. If M is a pre-trained LVLM as claimed, this is internally inconsistent, because the model receives no visual tokens. If M is actually a text-only LLM, then the comparison to image-grounded LVLM baselines is invalid, since the baselines would be doing a fundamentally different task. The authors must either modify the pipeline so that M also receives the image (e.g., T = M(P, I) or by feeding visual features alongside P) or explicitly reframe the contribution as text-based prompt augmentation for LLMs, with appropriate baselines.
  4. [Section IV-A (Dataset)] The evaluation benchmark ImageStoryGen-500K is self-constructed and not released. The paper gives no details on how the 500,000 images were sourced, how instructions were written, how image-instruction pairs were filtered, or how the test split was formed. Similarly, the extractor E is said to be trained on 'subsets' of ImageNet, COCO, and OpenImages with 'augmented' scene-graph, attribute, and sentiment annotations, but the exact subsets and annotation generation process are unspecified. Without a public dataset or detailed construction protocol, the experimental results are not independently verifiable.
  5. [Section IV-E (Human Evaluation)] The human evaluation uses only 5 raters and 200 samples per model, with no inter-annotator agreement measure (e.g., Cohen's kappa), no statistical significance test, and no explanation of how the subset was sampled. Table IV's differences, such as 4.25 vs. 3.80 for Overall Quality in StoryGen, are presented without confidence intervals, so the apparent human preference cannot be distinguished from noise. More raters, a defined sampling procedure, and appropriate statistics are needed before these results can support the paper's conclusions.
minor comments (5)
  1. [References] The citation [7] is used for both LLaVA and InstructBLIP in the Introduction, but the reference list entry [7] is Video-LLaVA; the paper should cite the actual LLaVA and InstructBLIP papers. Several other references (e.g., [1], [5], [11], [12]) are given only as 'CoRR' preprints without volume or page numbers, which is acceptable for preprints but should be labeled consistently.
  2. [Table I] Table I lists AdCopyGen only for VisuCraft and not for the two baselines, making it impossible to compare the advertising-copy task; either include baseline results for this task or state clearly why it was omitted from the comparison.
  3. [Abstract] There is a typo in the abstract: 'LLaV A' should be 'LLaVA'.
  4. [Section IV-D] The text says Level 1 granularity (Mean 0.769) is 'comparable to or even slightly below' LVLM-Base StoryGen (Mean 0.781); given the 0.012 gap, it would be clearer to state that it is slightly below, as the numbers show.
  5. [Section V] The conclusion repeatedly uses strong terms like 'unequivocally demonstrate' and 'significant improvements' without any statistical tests; the wording should be tempered to match the evidence actually presented.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is an empirical prompting pipeline and no equation or metric is defined in terms of the claimed outcome.

full rationale

VisuCraft's pipeline is explicitly compositional (T = M(G(E(I), U))), and the paper makes empirical claims about this pipeline rather than deriving a prediction from its own assumptions. The extractor E and prompt generator G are trained components whose behavior is reported through quantitative tables, but the paper does not define the VisuGen metrics in terms of E or G, does not fit a parameter and then rename it as a prediction, and does not rely on any load-bearing self-citation or imported uniqueness theorem to force its conclusion. The absence of released code, dataset, metric operationalization, and precise baseline configurations is a serious reproducibility and validity concern: the reported differences could in principle reflect metric bias, prompt mismatch, or decoding choices. That concern, however, is about the strength of the evidence for the empirical claim, not about the claim being equivalent to its inputs by construction. No equation in the paper collapses a 'prediction' into the fitted or design choices of the method, so under the stated criteria no significant circularity is found.

Assumptions & free parameters 3 free parameters · 3 assumptions · 3 invented entities

VisuCraft is an engineering framework whose claimed benefit rests entirely on private components: a self-built dataset, self-designed metrics, unspecified extractor training, and unnamed baselines. There are no fitted parameters in the mathematical sense because there is no mathematical derivation; however, the experimental setup contains many hand-chosen configurations that act as free parameters. The invented entities (benchmark, metrics, baselines) are not independently verifiable, which raises the circularity burden.

free parameters (3)
  • Granularity levels of the extractor output = Level 1, Level 2, Level 3
    The paper selects three granularity levels for E's output and reports that Level 3 works best in Section IV-D; these are hand-designed choices, not derived from theory, and they directly determine the reported scores.
  • Heuristic rules or learned policies in the prompt generation module = Not specified
    Section III-B states that prioritization 'employs heuristic rules or learned policies', but no rules, policy, or training details are given. These unspecified mechanisms are the core of the claimed improvement.
  • Training subsets of ImageNet, COCO, and OpenImages = Subsets, not quantified
    The extractor is 'trained or fine-tuned using extensive image datasets' with augmented fine-grained annotations, but the paper never specifies which subsets, which annotation sources, or how the augmentation was done, so the exact training configuration is a hidden free parameter.
assumptions (3)
  • domain assumption Structured textual/JSON visual representations are more useful for creative generation than raw embeddings or captions.
    This is the central assumption of the method, stated in Section III-A; the experiments are designed to confirm it but cannot, because they are not independently reproducible.
  • domain assumption The VisuGen metrics measure visual grounding, creativity, and instruction adherence.
    The metrics are defined only in prose in Section IV-A with no rubric, no model-based or human scoring implementation, and no validation against external benchmarks.
  • domain assumption LLaVA and InstructBLIP are faithful representatives of current LVLMs and are implemented fairly in baselines.
    The baselines are described only as 'architectures similar to LLaVA or InstructBLIP' in Section IV-A, with no exact versions, prompting strategy, or hyperparameters.
invented entities (3)
  • ImageStoryGen-500K benchmark
    purpose: Evaluation dataset of 500,000 image-instruction pairs used to compare VisuCraft against baselines.
    The dataset is self-constructed and not released; no samples, annotation protocol, or quality checks are shown, so it cannot be independently inspected or used by others.
  • VisuGen Metrics (Visual Grounding, Creativity, Instruction Adherence)
    purpose: Custom quantitative metrics used to measure the framework's improvements.
    The metrics are named and described in prose only; their implementation, calibration, and agreement with human judgment are not provided.
  • LVLM-Base and LVLM-Enhanced baselines
    purpose: Comparison systems intended to represent standard and enhanced vision-language models.
    The baselines are unnamed and unreleased, making it impossible to verify that VisuCraft is compared against strong, relevant systems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VisuCraft: Enhancing Large Vision-Language Models for Complex Visual-Guided Creative Content Generation via Structured Information Extraction." pith.science (2026). https://pith.science/paper/ZX2WRCFV

@misc{pith2026250802890,
  author       = {Pith},
  title        = {Pith review of: VisuCraft: Enhancing Large Vision-Language Models for Complex Visual-Guided Creative Content Generation via Structured Information Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZX2WRCFV}},
  note         = {Machine review of arXiv:2508.02890}
}
read the original abstract

This paper introduces VisuCraft, a novel framework designed to significantly enhance the capabilities of Large Vision-Language Models (LVLMs) in complex visual-guided creative content generation. Existing LVLMs often exhibit limitations in maintaining high visual fidelity, genuine creativity, and precise adherence to nuanced user instructions when generating long-form texts. VisuCraft addresses these challenges by integrating a multimodal structured information extractor (E) and a dynamic prompt generation module (G). The extractor distills fine-grained visual attributes from input images into a rich, structured representation, which the dynamic prompt module then combines with user instructions to create highly optimized prompts for underlying LVLMs (e.g., LLaVA, InstructBLIP). Evaluated on the self-constructed ImageStoryGen-500K dataset using VisuGen Metrics (Visual Grounding, Creativity, and Instruction Adherence), VisuCraft consistently outperforms baseline LVLMs across tasks like story generation and poetry composition. Our results demonstrate remarkable improvements, particularly in creativity and instruction adherence, validating VisuCraft's effectiveness in producing imaginative, visually grounded, and user-aligned long-form creative text. This work unlocks new potential for LVLMs in sophisticated creative AI applications.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 23 canonical work pages

  1. [1]

    A survey on large language model (LLM) security and privacy: The good, the bad, and the ugly,

    Y . Yao, J. Duan, K. Xu, Y . Cai, E. Sun, and Y . Zhang, “A survey on large language model (LLM) security and privacy: The good, the bad, and the ugly,” CoRR, 2023

    2023

  2. [2]

    Visual in-context learning for large vision-language models,

    Y . Zhou, X. Li, Q. Wang, and J. Shen, “Visual in-context learning for large vision-language models,” in Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024 . Association for Computational Linguis- tics, 2024, pp. 15 890–15 902

    2024

  3. [3]

    Rethinking visual dependency in long-context reasoning for large vision-language models,

    Y . Zhou, Z. Rao, J. Wan, and J. Shen, “Rethinking visual dependency in long-context reasoning for large vision-language models,” arXiv preprint arXiv:2410.19732, 2024

    arXiv 2024

  4. [4]

    Lvlm-ehub: A comprehensive evaluation bench- mark for large vision-language models,

    P. Xu, W. Shao, K. Zhang, P. Gao, S. Liu, M. Lei, F. Meng, S. Huang, Y . Qiao, and P. Luo, “Lvlm-ehub: A comprehensive evaluation bench- mark for large vision-language models,” IEEE Trans. Pattern Anal. Mach. Intell., pp. 1877–1893, 2025

    2025

  5. [5]

    COVID-19 detection in chest x-ray images using swin-transformer and transformer in transformer,

    J. Jiang and S. Lin, “COVID-19 detection in chest x-ray images using swin-transformer and transformer in transformer,” CoRR, 2021

    work page 2021

  6. [6]

    Vislinginstruct: Elevating zero-shot learning in multi-modal lan- guage models with autonomous instruction optimization,

    D. Zhu, X. Tang, W. Han, J. Lu, Y . Zhao, G. Xing, J. Wang, and D. Yin, “Vislinginstruct: Elevating zero-shot learning in multi-modal lan- guage models with autonomous instruction optimization,” arXiv preprint arXiv:2402.07398, 2024

    arXiv 2024

  7. [7]

    Video-llava: Learning united visual representation by alignment before projection,

    B. Lin, Y . Ye, B. Zhu, J. Cui, M. Ning, P. Jin, and L. Yuan, “Video-llava: Learning united visual representation by alignment before projection,” in Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, November 12-16,

    work page 2024

  8. [8]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA. IEEE Computer Society, 2009, pp. 248–255

    2009

Show all 33 references
  1. [9]

    Microsoft COCO: common objects in context,

    T. Lin, M. Maire, S. J. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft COCO: common objects in context,” in Computer Vision - ECCV 2014 - 13th European Confer- ence, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V . Springer...

    2014

  2. [10]

    The open images dataset V4: unified image classification, object detection, and visual relationship detection at scale,

    A. Kuznetsova, H. Rom, N. Alldrin, J. R. R. Uijlings, I. Krasin, J. Pont- Tuset, S. Kamali, S. Popov, M. Malloci, T. Duerig, and V . Ferrari, “The open images dataset V4: unified image classification, object detection, and visual relationship detection at scale,” CoRR, 2018

    2018

  3. [11]

    A comprehensive survey and guide to multimodal large language models in vision-language tasks,

    C. X. Liang, P. Tian, C. H. Yin, Y . Yua, W. An-Hou, L. Ming, T. Wang, Z. Bi, and M. Liu, “A comprehensive survey and guide to multimodal large language models in vision-language tasks,” CoRR, 2024

    2024

  4. [12]

    Foundational models defining a new era in vision: A survey and outlook,

    M. Awais, M. Naseer, S. H. Khan, R. M. Anwer, H. Cholakkal, M. Shah, M. Yang, and F. S. Khan, “Foundational models defining a new era in vision: A survey and outlook,” CoRR, 2023

    2023

  5. [13]

    Lmfusion: Adapting pretrained language models for multimodal generation,

    W. Shi, X. Han, C. Zhou, W. Liang, X. V . Lin, L. Zettlemoyer, and L. Yu, “Lmfusion: Adapting pretrained language models for multimodal generation,” CoRR, 2024

    2024

  6. [14]

    Multi-modal generation via cross-modal in-context learn- ing,

    A. Kumar, M. Naseer, S. Narayan, R. M. Anwer, S. Khan, and H. Cholakkal, “Multi-modal generation via cross-modal in-context learn- ing,” CoRR, 2024

    2024

  7. [15]

    Improving cross-modal alignment for text- guided image inpainting,

    Y . Zhou and G. Long, “Improving cross-modal alignment for text- guided image inpainting,” in Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics , 2023, pp. 3445–3456

    2023

  8. [16]

    Fine-grained distillation for long document retrieval,

    Y . Zhou, T. Shen, X. Geng, C. Tao, J. Shen, G. Long, C. Xu, and D. Jiang, “Fine-grained distillation for long document retrieval,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 17, 2024, pp. 19 732–19 740

    2024

  9. [17]

    Towards robust ranker for text retrieval,

    Y . Zhou, T. Shen, X. Geng, C. Tao, C. Xu, G. Long, B. Jiao, and D. Jiang, “Towards robust ranker for text retrieval,” in Findings of the Association for Computational Linguistics: ACL 2023 , 2023, pp. 5387– 5401

    2023

  10. [18]

    Sda: simple discrete augmentation for contrastive sentence representation learning,

    D. Zhu, Z. Mao, J. Lu, R. Zhao, and F. Tan, “Sda: simple discrete augmentation for contrastive sentence representation learning,” arXiv preprint arXiv:2210.03963, 2022

    2022 arXiv

  11. [19]

    Improving zero-shot cross-lingual transfer for multilingual question answering over knowledge graph,

    Y . Zhou, X. Geng, T. Shen, W. Zhang, and D. Jiang, “Improving zero-shot cross-lingual transfer for multilingual question answering over knowledge graph,” in Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Huma...

    2021

  12. [20]

    Divide-then-aggregate: An efficient tool learning method via parallel tool invocation,

    D. Zhu, W. Shi, Z. Shi, Z. Ren, S. Wang, L. Yan, and D. Yin, “Divide-then-aggregate: An efficient tool learning method via parallel tool invocation,” arXiv preprint arXiv:2501.12432 , 2025

    2025 arXiv

  13. [21]

    Grounding multimodal large language models to the world,

    Z. Peng, W. Wang, L. Dong, Y . Hao, S. Huang, S. Ma, Q. Ye, and F. Wei, “Grounding multimodal large language models to the world,” in The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net, 2024

    2024

  14. [22]

    Minigpt-4: Enhancing vision-language understanding with advanced large language models,

    D. Zhu, J. Chen, X. Shen, X. Li, and M. Elhoseiny, “Minigpt-4: Enhancing vision-language understanding with advanced large language models,” in The Twelfth International Conference on Learning Represen- tations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024

    2024

  15. [23]

    Personalizing multimodal large language models for image captioning: An experimental analysis,

    D. Bucciarelli, N. Moratelli, M. Cornia, L. Baraldi, and R. Cucchiara, “Personalizing multimodal large language models for image captioning: An experimental analysis,” in Computer Vision - ECCV 2024 Workshops - Milan, Italy, September 29-October 4, 2024, Proceedings, Part XXII...

    2024

  16. [24]

    Prompting AI art: An investigation into the creative skill of prompt engineering,

    J. Oppenlaender, R. Linder, and J. M. Silvennoinen, “Prompting AI art: An investigation into the creative skill of prompt engineering,” CoRR, 2023

    2023

  17. [25]

    Crafting effective prompts: A guideline for successful image generation,

    N. W. Ahmad and S. Ruslan, “Crafting effective prompts: A guideline for successful image generation,” in 14th International Conference on System Engineering and Technology, ICSET 2024, Bandung, Indonesia, October 2-3, 2024 . IEEE, 2024, pp. 84–89

    2024

  18. [26]

    Prompt engineering for chatgpt: a quick guide to techniques, tips, and best practices,

    S. Ekin, “Prompt engineering for chatgpt: a quick guide to techniques, tips, and best practices,” Authorea Preprints, 2023

    2023

  19. [27]

    Automatic generation of multimedia teaching materials based on generative AI: taking tang poetry as an example,

    X. Chen and D. Wu, “Automatic generation of multimedia teaching materials based on generative AI: taking tang poetry as an example,” IEEE Trans. Learn. Technol., pp. 1353–1366, 2024

    2024

  20. [28]

    STAR: boosting low-resource information extraction by structure-to- text data generation with large language models,

    M. D. Ma, X. Wang, P. Kung, P. J. Brantingham, N. Peng, and W. Wang, “STAR: boosting low-resource information extraction by structure-to- text data generation with large language models,” in Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conf...

    2024

  21. [29]

    Knowgpt: Knowledge graph based prompting for large language models,

    Q. Zhang, J. Dong, H. Chen, D. Zha, Z. Yu, and X. Huang, “Knowgpt: Knowledge graph based prompting for large language models,” in Advances in Neural Information Processing Systems 38: Annual Con- ference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, B...

    2024

  22. [30]

    Prompt-based length controlled generation with multiple control types,

    R. Jie, X. Meng, L. Shang, X. Jiang, and Q. Liu, “Prompt-based length controlled generation with multiple control types,” in Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024 . Association for Computa...

    2024

  23. [31]

    Weak to strong generalization for large language models with multi-capabilities,

    Y . Zhou, J. Shen, and Y . Cheng, “Weak to strong generalization for large language models with multi-capabilities,” in The Thirteenth International Conference on Learning Representations , 2025

    2025

  24. [32]

    Prompting creativity: Tiered approach to copyright protection for ai-generated content in the digital age,

    W. Jon, “Prompting creativity: Tiered approach to copyright protection for ai-generated content in the digital age,” Media and Communication , 2025. 9

    2025

  25. [2024]

    5971–5984

    Association for Computational Linguistics, 2024, pp. 5971–5984

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.