Pith. sign in

REVIEW 4 major objections 6 minor 35 references

Improving Large Vision-Language Models' Understanding for Flow Field Data

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read FieldLVLM shows that a general-purpose vision-language model, fine-tuned on compressed flow fields and structured captions generated by specialist models, can classify flow type, compute Reynolds numbers, and identify vortices with high…

desk verdict A plausible engineering pipeline for adapting LVLMs to flow fields, but the evaluation is circular and the baselines are vacuous—the quantitative claims do not hold up. read the letter →

arxiv 2507.18311 v3 pith:YCPN6OO3 submitted 2025-07-24 cs.CV

Xiaomei Zhang , Hanyu Zheng , Xiangyu Zhu , Jinghuan Wei , Junhong Zou , Zhen Lei , Zhaoxiang Zhang This is my paper · ORCID
classification cs.CV
keywords largevision-languagemodelsfielddataVQGANcompressionReynoldsnumberestimationvortexidentificationmultimodalscientificflowunderstandingLoRAfine-tuning
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 show that a large vision-language model can be made to understand scientific field data, specifically fluid velocity and pressure fields, rather than only natural images and text. The proposed framework, FieldLVLM, first converts raw fields into images, compresses them into a small token budget, and fine-tunes the model on captions produced by a pipeline of specialist machine-learning models. The reported results are high: 100% flow-type classification, 99.79% Reynolds-number estimation, 97.23% vortex identification, and 85.41% general field-data analysis on benchmark datasets, while comparison vision-language models score zero on the same tasks. If these numbers hold, the work would offer a recipe for pointing general-purpose multimodal models at scientific data without building task-specific architectures from scratch.

What carries the argument

The load-bearing machinery is the two-stage input pipeline. Raw 256-by-256 fields of horizontal velocity, vertical velocity, and pressure are normalized and stacked as RGB channels, then encoded by a pretrained VQGAN into 256 tokens, a 99.6% reduction from the 65,536-token raw representation. In parallel, a handful of representative physical values are selected from the field and fed alongside the tokens, and the specialist-generated textual description is itself converted into an image input. These compressed tokens, key values, and the semantic image are what the fine-tuned language decoder attends to, so the model never sees the raw field as text and stays within the context window.

What would settle it

Take a held-out set of flow fields, generate captions with the specialist models but deliberately corrupt or reverse the stated Reynolds numbers and vortex parameters in the caption text, retrain under the same scheme, and check whether the model still reports the original correct values; if it reports the corrupted values instead, the claimed physical understanding collapses.

Watch

Extended reading notes

Core claim

The central discovery claimed is that a standard open vision-language model, Qwen2.5-VL-7B, can be adapted to fluid-flow field data by two coordinated interventions: a field-aware language generation strategy that uses specialized models to label flow type, Reynolds number, and vortex structure and then rewrites those labels into consistent text, and a data-compression strategy that maps velocity and pressure fields to an RGB image, compresses it with VQGAN into 256 discrete tokens, and augments the input with a small number of representative key values. With LoRA fine-tuning, the resulting model answers structured questions about flow fields with the accuracies listed above. The paper interprets this as evidence that the bottleneck for vision-language models in science is not the model itself but the format and availability of training data: once field data is compressed and paired with structured physical language, the model can reason about it.

Load-bearing premise

The paper's numbers stand on the assumption that the evaluation labels are independent of the teacher models that generated the training captions, so the reported accuracy reflects genuine physical understanding rather than the model repeating labels it was trained on.

Editorial extensions

If this is right

  • If the reported accuracy holds, a 256-token compressed representation preserves enough physical information for a vision-language model to compute Reynolds numbers and locate vortices, so token compression is not the limiting factor.
  • Field-aware language generation provides a scalable way to build multimodal scientific datasets: specialist models supply precise labels and a language model turns them into consistent captions, reducing the need for manual annotation.
  • The large gap between base fine-tuning and the version with key-value guidance (53.94% to 100% on field-data analysis) implies that explicitly feeding selected informative values is the dominant driver of semantic understanding.
  • The zero scores of comparison models suggest that, in their default form, general vision-language models cannot even process such long-sequence scientific inputs, making the compression component a prerequisite rather than an enhancement.

Reading between the lines

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

  • Because the training captions were produced by the same specialist models that define the evaluation labels, the high scores may partly reflect the student learning to echo the teacher rather than deriving physics from the field; a version trained on captions whose labels are withheld or corrupted would separate these possibilities.
  • The same pipeline could generalize to other scientific field data, such as weather maps, medical images, or material microstructures, wherever a specialist model can generate structured captions; that extension is not tested in the paper.
  • A testable extension is to scramble or drop the key-value channel while keeping the VQGAN tokens; if accuracy collapses, the model is reading the provided numbers rather than the field itself.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes FieldLVLM, a framework for fine-tuning the Qwen2.5-VL-7B large vision-language model on scientific flow field data (velocity and pressure fields). The method combines a 'field-aware language generation strategy,' in which special-purpose machine-learning models produce textual descriptions of flow type, Reynolds number, vortex parameters, and maximum velocity, with a 'data-compressed multimodal model tuning' stage that maps fields to RGB images, compresses them with VQGAN into discrete tokens, selects 'key values,' and fine-tunes the LVLM with LoRA. The authors report very high accuracies on four tasks (flow categorization 100%, Reynolds number estimation 99.79%, vortex identification 97.23%, field data analysis 85.41%) and claim that FieldLVLM significantly outperforms DeepSeek-VL-7B, LLaVA-v1.6, and Llama-3.2-Vision, all of which score 0/NA. Ablation studies on vortex identification and field data analysis show gains from compression and key-data selection, with the latter jumping to 100%.

Significance. If the claims were sound, the paper would present a useful recipe for adapting LVLMs to scientific field data, including a scalable caption-generation pipeline and a token-compression scheme that addresses input-length limits. The authors also introduce a benchmark from extended FlowBench and CFDBench datasets, which could be a community resource. However, the central empirical claims are not supported: the baseline comparison is vacuous because all comparison models score 0/NA, the evaluation is circular with respect to the teacher-generated training captions, and the reported results show suspicious patterns (e.g., 100% accuracy with key-data selection) without error bars or held-out generalization checks. As presented, the paper does not establish that FieldLVLM understands physical content beyond reproducing teacher outputs.

major comments (4)
  1. [§IV.C, Table I] The central comparative claim is unsupported because all three baselines (DeepSeek-VL-7B, LLaVA-v1.6-7B, Llama-3.2-11B-Vision) score 0/NA on every task. The paper attributes this to 'architectural incompatibilities with scientific data structures,' yet Tables IV and V show these same models producing text responses on the same types of input. If the baselines cannot process the inputs, the evaluation protocol itself is at fault and the scores are meaningless; if they can process the inputs, the 0/NA scores are unexplained. Either way, no valid conclusion about 'significantly outperforming existing methods' can be drawn from this comparison.
  2. [§III.A and §IV.B] The evaluation is circular. The training captions are generated by special-purpose models that output exactly the quantities used as evaluation targets: flow classification, Reynolds number, vortex parameters, and maximum velocity (Section III.A). The benchmarks are 'extended versions' of FlowBench and CFDBench, the same sources used for training (Section IV.A), and no train/test split is described. The paper never establishes that the ground-truth labels are independent of the teacher-generated captions, so the reported high accuracies may reflect memorization of teacher outputs rather than physical understanding of the field data. This is the weakest assumption of the paper, and it is not addressed.
  3. [Table III] The ablation study reports that adding key-data selection raises Field Data Analysis accuracy from 53.94% to 100% (Table III). This is a red flag: a single design choice producing perfect accuracy on a benchmark of over 70,000 samples, without error bars, statistical significance, or any discussion of variance, strongly suggests that the task is trivially solvable by the selected heuristics or that there is leakage between training and evaluation. The paper should provide a detailed analysis of why 100% is achieved and on what subset.
  4. [§IV.B and Tables I–III] The evaluation protocol is under-specified. The paper defines accuracy for each task but does not state how many samples per task, how the 70,000+ samples are partitioned into training, validation, and test sets, or whether any test sample overlaps with training data. It also does not define what '0/NA' means operationally (e.g., failure to produce parseable output, out-of-token-limit errors, or incorrect answers). Without this information, the reported numbers cannot be interpreted or reproduced.
minor comments (6)
  1. [§III.B.1] The sentence 'Therefore, based on? The principle is to select a key piece of data for focused analysis to enhance the effectiveness of the overall data analysis.' is garbled and contains a dangling question mark; it should be rewritten.
  2. [References [29]–[31]] Reference [29] is cited as DeepSeek-VL-7B, but the reference list entry is for DeepSeek-Coder-V2, which is a different model; please correct the citation or the reference. Similarly, [30] and [31] are not clearly tied to the classification and detection models used in Section III.A.
  3. [Tables IV and V] Both tables include the placeholder phrase 'Visual input example, Extreme Ironing:' followed by a source path. 'Extreme Ironing' appears to be an artifact and should be removed or replaced with a proper description of the input image.
  4. [Fig. 1] Figure 1 is referenced in the text but the caption and the figure content do not appear in the provided manuscript; all figures should be checked for inclusion and correct placement.
  5. [§IV.B] The metric 'Accuracy (%)' is not fully defined: it is unclear whether it is a per-sample exact match, a tolerance-based match, or an aggregate over a test set, and what the denominator in '0/NA' represents. Please define the metrics precisely.
  6. [Tables II and III] The ablation tables report different 'Base Fine-tuning' numbers for the same baseline model on different tasks (82.28% vs 53.94%), which is expected for different tasks, but the caption should clarify that these numbers are task-specific; also, the 0/NA baseline appears in both tables and warrants explanation.

Circularity Check

2 steps flagged · score 8.0 of 10

Evaluation targets are the same quantities that the teacher models inject into the training text, and the benchmark is built from the same extended datasets; the reported accuracies therefore reduce to reproducing teacher labels rather than establishing physical understanding.

  1. fitted input called prediction [Section I, Introduction (pp. 1-2); echoed in Section III.A]
    "Specifically, given field data, the field-aware language generation strategy uses domain-special models to generate flow classification, Reynolds number, and vortex detection. Then, the results and original field data are sent into a large language model to produce a consistent field language representation for downstream learning and model tuning."

    The training text is generated by teacher models that output exactly the quantities later scored: Section IV.B defines Flow Categorization, Reynolds Number Estimation, Vortex Identification, and Field Data Analysis. Fine-tuning the LVLM to imitate these teacher descriptions is a fit to the evaluation targets; the subsequent 'predictions' of those same quantities are not independent measurements of physical understanding. Since the benchmark (Section IV.A) is built from the same extended FlowBench/CFDBench samples that carry the teacher captions, high accuracy can be explained by reproducing teacher outputs rather than by reasoning from the field images.

  2. self definitional [Section IV.A, Dataset Statistics (p. 5); Section IV.B, Evaluation Benchmark and Metrics (p. 5)]
    "Our experiments are based on extended versions of FlowBench [34] and CFDBench [35], where each sample is paired with structured language descriptions generated by our field-aware strategy."

    The benchmark tasks (Section IV.B) are evaluated on 'over 70,000 samples from the aforementioned datasets'—the same extended versions used for training. No held-out test set, no sample-level disjointness, and no independence of labels from the teacher-generated captions is stated; the only split, a 10% validation split, is described in Implementation Details and is not an independent benchmark. Thus the evaluation is defined over the same data whose captions contain the answers, making the reported accuracies forced by training distribution rather than a genuine test of generalization.

full rationale

The central claimed contribution—that FieldLVLM learns physical understanding from velocity-pressure fields—rests on an evaluation whose labels coincide with the quantities injected into the training text. Section III.A describes a teacher pipeline that outputs flow classification, Reynolds number, and vortex detection; Section IV.B scores exactly those capabilities; Section IV.A builds the benchmark from the same extended FlowBench/CFDBench samples that carry the teacher-generated captions. With no stated held-out test set, the reported accuracies can be achieved by imitating the teacher text rather than by inferring physics from the images. The comparison baselines are vacuous (0/NA due to 'architectural incompatibilities'), so they do not provide independent calibration. The +Key data ablation (Table III) is additionally suspect because 'representative key values' are extracted from the original data and may include the Umax that Field Data Analysis measures; the paper does not disclose these values, so this specific reduction is not proven. I therefore count the teacher-caption/evaluation-target overlap and the shared benchmark as the load-bearing circular steps; the remainder of the pipeline (VQGAN compression, LoRA tuning) is not itself circular.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the accuracy of teacher-generated labels, the representational sufficiency of the compressed image, and the fairness of the baseline comparison. None of these is independently validated.

free parameters (4)
  • Reynolds number error tolerance = 10%
    Hand-chosen threshold used to judge a prediction correct; changes accuracy numbers.
  • Vortex center deviation tolerance = 25% of flow domain size
    Hand-chosen threshold for vortex identification success.
  • VQGAN token budget = 256 tokens
    Compression target chosen to fit Qwen2.5-VL context; not derived from data.
  • Key data selection method = not specified
    The representative key values are selected by an unspecified criterion; in Table III this selection alone raises accuracy from 53.94% to 100%, suggesting it may encode the answer.
assumptions (3)
  • domain assumption Flow fields can be represented as RGB images and compressed by VQGAN without losing the physical information needed to answer the benchmark questions.
    Invoked in Section III.B.1; no validation that the 256-token latent preserves the relevant physics.
  • domain assumption The specialist teacher models provide accurate labels for flow type, Reynolds number, and vortices, and those labels match the evaluation ground truth.
    The entire training set depends on these labels; the paper gives no accuracy figures for the teachers or evidence that their outputs are independent of the evaluation annotations.
  • domain assumption The baselines' 0/NA scores are due to architectural incompatibility rather than a failure to provide parseable inputs or to parse outputs.
    The paper asserts incompatibility (Section IV.C) without describing the exact input/output protocol; this assumption underlies the claim of superiority.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Large Vision-Language Models' Understanding for Flow Field Data." pith.science (2026). https://pith.science/paper/YCPN6OO3

@misc{pith2026250718311,
  author       = {Pith},
  title        = {Pith review of: Improving Large Vision-Language Models' Understanding for Flow Field Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YCPN6OO3}},
  note         = {Machine review of arXiv:2507.18311}
}
read the original abstract

Large Vision-Language Models (LVLMs) have shown impressive capabilities across a range of tasks that integrate visual and textual understanding, such as image captioning and visual question answering. These models are trained on large-scale image and video datasets paired with text, enabling them to bridge visual perception and natural language processing. However, their application to scientific domains, especially in interpreting complex field data commonly used in the natural sciences, remains underexplored. In this work, we introduce FieldLVLM, a novel framework designed to improve large vision-language models' understanding of field data. FieldLVLM consists of two main components: a field-aware language generation strategy and a data-compressed multimodal model tuning. The field-aware language generation strategy leverages a special-purpose machine learning pipeline to extract key physical features from field data, such as flow classification, Reynolds number, and vortex patterns. This information is then converted into structured textual descriptions that serve as a dataset. The data-compressed multimodal model tuning focuses on LVLMs with these generated datasets, using a data compression strategy to reduce the complexity of field inputs and retain only the most informative values. This ensures compatibility with the models language decoder and guides its learning more effectively. Experimental results on newly proposed benchmark datasets demonstrate that FieldLVLM significantly outperforms existing methods in tasks involving scientific field data. Our findings suggest that this approach opens up new possibilities for applying large vision-language models to scientific research, helping bridge the gap between large models and domain-specific discovery.

Figures

Figures reproduced from arXiv: 2507.18311 by the authors.

Figure 1
Figure 1. (a) Quantitative comparison of vision-language responses in field data across different methods. (b) The performance of key evaluation metrics, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The pipeline of field-aware language data generation strategy integrating special-purpose machine learning models for field classification, Reynolds [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustrative examples of generated field language representations, [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Input-output architecture of the data-compressed multimodal model featuring VQGAN-based token compression, key value selection, and image [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Q&A analysis on vortex shedding dynamics and pressure distribution in flow past a bluff body highlighting k [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Q&A session on velocity distribution and vorticity analysis for dam break flow field characterization demonstrating shear layer dynamics and critical [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 12 canonical work pages

  1. [1]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning . PmLR, 2021, pp. 8748–8763

    2021

  2. [2]

    Florence: A new foundation model for computer vision,

    L. Yuan, D. Chen, Y .-L. Chen, N. Codella, X. Dai, J. Gao, H. Hu, X. Huang, B. Li, C. Li et al. , “Florence: A new foundation model for computer vision,” arXiv preprint arXiv:2111.11432 , 2021

    arXiv 2021

  3. [3]

    Grounded language-image pre-training,

    L. H. Li, P. Zhang, H. Zhang, J. Yang, C. Li, Y . Zhong, L. Wang, L. Yuan, L. Zhang, J.-N. Hwang et al. , “Grounded language-image pre-training,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 10 965–10 975

    work page 2022

  4. [4]

    Regionclip: Region-based language-image pretraining,

    Y . Zhong, J. Yang, P. Zhang, C. Li, N. Codella, L. H. Li, L. Zhou, X. Dai, L. Yuan, Y . Liet al., “Regionclip: Region-based language-image pretraining,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 16 793–16 803

    work page 2022

  5. [5]

    Generalized decoding for pixel, image, and language,

    X. Zou, Z.-Y . Dou, J. Yang, Z. Gan, L. Li, C. Li, X. Dai, H. Behl, J. Wang, L. Yuan et al. , “Generalized decoding for pixel, image, and language,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2023, pp. 15 116–15 127

    work page 2023

  6. [6]

    Language-driven semantic segmentation,

    B. Li, K. Q. Weinberger, S. Belongie, V . Koltun, and R. Ranftl, “Language-driven semantic segmentation,” arXiv preprint arXiv:2201.03546, 2022

    arXiv 2022

  7. [7]

    Qwen2. 5-vl technical report,

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang et al., “Qwen2. 5-vl technical report,” arXiv preprint arXiv:2502.13923, 2025

    arXiv 2025

  8. [8]

    Taming transformers for high- resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming transformers for high- resolution image synthesis,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 12 873–12 883

    2021

Show all 35 references
  1. [9]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” in Pro- ceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technolo...

    2019

  2. [10]

    Exploring the limits of transfer learning with a unified text-to-text transformer,

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y . Zhou, W. Li, and P. J. Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer,” Journal of machine learning research, vol. 21, no. 140, pp. 1–67, 2020

    2020

  3. [11]

    Improving language understanding by generative pre-training,

    A. Radford, K. Narasimhan, T. Salimans, I. Sutskever et al., “Improving language understanding by generative pre-training,” 2018

    2018

  4. [12]

    Language mod- els are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al., “Language mod- els are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

    1901

  5. [13]

    Palm: Scal- ing language modeling with pathways,

    A. Chowdhery, S. Narang, J. Devlin, M. Bosma, G. Mishra, A. Roberts, P. Barham, H. W. Chung, C. Sutton, S. Gehrmann et al., “Palm: Scal- ing language modeling with pathways,” Journal of Machine Learning Research, vol. 24, no. 240, pp. 1–113, 2023

    2023

  6. [14]

    Training language models to follow instructions with human feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray et al. , “Training language models to follow instructions with human feedback,” Advances in neural information processing systems , vol. 35, pp. 27 730–27 744, 2022

    2022

  7. [15]

    Role of chat gpt in public health,

    S. S. Biswas, “Role of chat gpt in public health,” Annals of biomedical engineering, vol. 51, no. 5, pp. 868–869, 2023

    2023

  8. [16]

    Scaling up visual and vision-language representation learning with noisy text supervision,

    C. Jia, Y . Yang, Y . Xia, Y .-T. Chen, Z. Parekh, H. Pham, Q. Le, Y .-H. Sung, Z. Li, and T. Duerig, “Scaling up visual and vision-language representation learning with noisy text supervision,” in International conference on machine learning . PMLR, 2021, pp. 4904–4916

    2021

  9. [17]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” Advances in neural information processing systems, vol. 36, pp. 34 892– 34 916, 2023

    2023

  10. [18]

    Instructblip: Towards general-purpose vision- language models with instruction tuning,

    W. Dai, J. Li, D. Li, A. M. H. Tiong, J. Zhao, W. Wang, B. Li, P. Fung, and S. Hoi, “Instructblip: Towards general-purpose vision- language models with instruction tuning,” 2023. 10

    2023

  11. [19]

    The art of artificial intelligence: Themes and case studies of knowledge engineering,

    E. A. Feigenbaum et al., “The art of artificial intelligence: Themes and case studies of knowledge engineering,” 1977

    1977

  12. [20]

    Large language models for scientific discovery in molecular property prediction,

    Y . Zheng, H. Y . Koh, J. Ju, A. T. Nguyen, L. T. May, G. I. Webb, and S. Pan, “Large language models for scientific discovery in molecular property prediction,” Nature Machine Intelligence , pp. 1–11, 2025

    2025

  13. [21]

    Unsupervised word embeddings capture latent knowledge from materials science literature,

    V . Tshitoyan, J. Dagdelen, L. Weston, A. Dunn, Z. Rong, O. Kononova, K. A. Persson, G. Ceder, and A. Jain, “Unsupervised word embeddings capture latent knowledge from materials science literature,” Nature, vol. 571, no. 7763, pp. 95–98, 2019

    2019

  14. [22]

    Highly accurate protein structure prediction with alphafold,

    J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger, K. Tunyasuvunakool, R. Bates, A. ˇZ´ıdek, A. Potapenko et al., “Highly accurate protein structure prediction with alphafold,” nature, vol. 596, no. 7873, pp. 583–589, 2021

    2021

  15. [23]

    Improving clip training with language rewrites,

    L. Fan, D. Krishnan, P. Isola, D. Katabi, and Y . Tian, “Improving clip training with language rewrites,” Advances in Neural Information Processing Systems, vol. 36, pp. 35 544–35 575, 2023

    2023

  16. [24]

    Datacomp: In search of the next generation of multimodal datasets,

    S. Y . Gadre, G. Ilharco, A. Fang, J. Hayase, G. Smyrnis, T. Nguyen, R. Marten, M. Wortsman, D. Ghosh, J. Zhang et al. , “Datacomp: In search of the next generation of multimodal datasets,” Advances in Neural Information Processing Systems , vol. 36, pp. 27 092–27 112, 2023

    2023

  17. [25]

    From scarcity to efficiency: Improving clip training via visual-enriched captions,

    Z. Lai, H. Zhang, W. Wu, H. Bai, A. Timofeev, X. Du, Z. Gan, J. Shan, C.-N. Chuah, Y . Yang et al. , “From scarcity to efficiency: Improving clip training via visual-enriched captions,” 2023

    2023

  18. [26]

    Improv- ing multimodal datasets with image captioning,

    T. Nguyen, S. Y . Gadre, G. Ilharco, S. Oh, and L. Schmidt, “Improv- ing multimodal datasets with image captioning,” Advances in Neural Information Processing Systems , vol. 36, pp. 22 047–22 069, 2023

    2023

  19. [27]

    Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts,

    S. Changpinyo, P. Sharma, N. Ding, and R. Soricut, “Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 3558–3568

    2021

  20. [28]

    Laion- 5b: An open large-scale dataset for training next generation image-text models,

    C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortsmanet al., “Laion- 5b: An open large-scale dataset for training next generation image-text models,” Advances in neural information processing systems , vol. 35, pp....

    2022

  21. [29]

    Deepseek-coder-v2: Breaking the barrier of closed-source models in code intelligence,

    Q. Zhu, D. Guo, Z. Shao, D. Yang, P. Wang, R. Xu, Y . Wu, Y . Li, H. Gao, S. Ma et al., “Deepseek-coder-v2: Breaking the barrier of closed-source models in code intelligence,” arXiv preprint arXiv:2406.11931 , 2024

    2024 arXiv

  22. [30]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

    2016

  23. [31]

    Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection,

    S. Zhang, C. Chi, Y . Yao, Z. Lei, and S. Z. Li, “Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 9759–9768

    2020

  24. [32]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen et al., “Lora: Low-rank adaptation of large language models.” ICLR, vol. 1, no. 2, p. 3, 2022

    2022

  25. [33]

    [Online]

    Meta, “Llama,” 2024. [Online]. Available: https://www.llama.com/

    2024

  26. [34]

    Flowbench: A large scale benchmark for flow simulation over complex geometries,

    R. Tali, A. Rabeh, C.-H. Yang, M. Shadkhah, S. Karki, A. Upadhyaya, S. Dhakshinamoorthy, M. Saadati, S. Sarkar, A. Krishnamurthy et al. , “Flowbench: A large scale benchmark for flow simulation over complex geometries,” arXiv preprint arXiv:2409.18032 , 2024

    2024 arXiv

  27. [35]

    Cfdbench: A large-scale bench- mark for machine learning methods in fluid dynamics,

    Y . Luo, Y . Chen, and Z. Zhang, “Cfdbench: A large-scale bench- mark for machine learning methods in fluid dynamics,” arXiv preprint arXiv:2310.05963, 2023

    2023 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.