Pith. sign in

REVIEW 4 major objections 5 minor 28 references

Can Visual Encoder Learn to See Arrows?

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Removing positional and textual biases from diagram training lets a CLIP image encoder learn explicit edge representations, lifting edge-direction probing accuracy from 0.509 to 0.860.

desk verdict Useful empirical result: contrastive finetuning on synthetic diagrams makes CLIP encode arrow direction, but the causal debiasing claim is overreaching without a biased-synthetic control. read the letter →

arxiv 2505.19944 v1 pith:RIKMTE7T submitted 2025-05-26 cs.CV cs.AIcs.CLcs.LG

Naoyuki Terashita , Yusuke Tozaki , Hideaki Omote , Congkha Nguyen , Ryosuke Nakamoto , Yuta Koreeda , Hiroaki Ozaki This is my paper · ORCID
classification cs.CVcs.AIcs.CLcs.LG
keywords diagramunderstandingedgerecognitionarrowsCLIPcontrastivelearningpositionalbiastextualcaptioning
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 image encoders in vision-language models fail to recognize arrows and lines in diagrams because they lean on textual and positional shortcuts, and that removing those shortcuts during training is enough for the encoder to learn real edge representations. To test this, the authors generate a synthetic dataset of directed graphs with arbitrary letter node labels and randomized force-directed layouts, so no edge can be inferred from text or position, and fine-tune CLIP on diagram–caption pairs with contrastive learning. The fine-tuned ViT-L/14 model jumps from near-chance (0.509) to 0.860 on edge-direction probing, raises image-retrieval mean average precision from 0.131 to 0.996, and reaches 0.966 F1 on diagram captioning, beating zero-shot GPT-4o. If true, this points toward a practical path for improving diagram comprehension in domain-specific applications.

What carries the argument

The controlling mechanism is the debiased diagram dataset: randomly generated directed graphs over two to eight alphabet-labeled nodes, with edges sampled independently, rendered by force-directed placement from random initial layouts, and captioned in Mermaid syntax such as A --> B. Randomizing node positions breaks positional bias, while arbitrary letter labels break textual bias, so edge presence and direction are only visible in the drawn lines. Contrastive learning on image–caption pairs then aligns the CLIP image embedding with the graph-structure text, forcing the encoder to represent the edges themselves rather than shortcut correlations.

What would settle it

Train the same CLIP encoder with the same contrastive and captioning setup on a deliberately biased variant of the generator by fixing node positions in a grid and using meaningful node words; if edge-direction probing accuracy still reaches about 0.86 and captioning F1 still approaches 0.97, the debiasing explanation fails.

Watch

Extended reading notes

Core claim

The central claim is that VLM failures on edges come from an over-reliance on textual and positional biases, and that training on data free of both biases lets an image encoder acquire edge representations. The authors demonstrate this with a synthetic diagram–caption dataset in which graph edges are independent of node labels and layouts, then contrastively fine-tune CLIP on it. They report large gains over the pretrained encoder on all three evaluations, including edge-direction probing rising from 0.509 to 0.860 for ViT-L/14 and diagram-captioning F1 reaching 0.966, above zero-shot GPT-4o and LLaVA-Mistral. The conclusion drawn is that removing these biases fosters accurate edge recognition in VLMs.

Load-bearing premise

The load-bearing assumption is that the measured gains come from removing positional and textual biases rather than from simply adapting the encoder to synthetic diagrams, since the paper never trains a biased control dataset.

Editorial extensions

If this is right

  • A vision encoder can acquire explicit edge features from contrastively matched diagram–caption pairs, without any edge annotations or graph-supervision labels.
  • Diagram understanding is bottlenecked by the image encoder rather than the decoder, since freezing the decoder and swapping in the debiased encoder lifts captioning F1 from 0.668 to 0.966.
  • The learned edge representations are layout-invariant, achieving image-retrieval MAP of 0.996 when the same graph appears with a different node arrangement.
  • A small fine-tuned CLIP encoder can beat zero-shot GPT-4o (F1 0.966 vs 0.500) on diagram captioning when textual and positional cues are absent.
  • Mermaid-format diagram captioning provides an automatable, edge-level evaluation of diagram understanding rather than a coarse image-level score.

Reading between the lines

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

  • We infer that the causal role of debiasing, as opposed to domain adaptation to synthetic diagrams, is not yet isolated, because a biased control with fixed layouts or meaningful labels is missing from the experiments.
  • A testable extension is to train the same encoder on a deliberately biased variant of the generator; if edge-probing and captioning gains persist, the explanation would shift from debiasing to contrastive training itself.
  • We would expect the method's gains to shrink on real diagrams, where occlusion, arrow style, and label placement introduce visual variation the random-layout generator does not model.
  • If the debiasing account holds, diagram-captioning benchmarks for large VLMs should be redesigned so that positional and textual shortcuts cannot answer the questions by themselves.
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

4 major / 5 minor

Summary. This paper hypothesizes that VLMs fail to recognize diagram edges because of an over-reliance on textual and positional biases. To test this, the authors generate a synthetic dataset of 100k diagram–caption pairs in which node labels are arbitrary letters, directed edges are generated independently with fixed probability, and node positions are obtained from force-directed layout with random initial layouts. They then contrastively finetune CLIP ViT-B/32 and ViT-L/14 and evaluate the image encoders via linear probing of node/edge existence and edge direction, image retrieval of graphs with different layouts, and a proposed diagram-captioning task using a GPT-2 decoder. The finetuned models substantially improve over pretrained CLIP, and the finetuned ViT-L/14 achieves 0.966 captioning F1, exceeding zero-shot GPT-4o. The paper concludes that removing textual and positional biases enables VLMs to learn edge representations.

Significance. The strengths of the paper are the controlled synthetic generation, the held-out test split, and the use of three complementary evaluation tasks that make the improvement in edge-direction probing concrete and falsifiable: the near-chance-to-0.860 jump for ViT-L/14 is a clear demonstration that the image encoder can acquire arrow-direction features. If the causal interpretation were established, the paper would offer a practical training recipe for diagram understanding in VLMs. However, the significance as stated depends on the claim that debiasing, rather than domain adaptation to synthetic diagram style or residual layout statistics, is the cause of the gains; the current experimental design does not separate these factors. The evidence is sufficient for the weaker claim that contrastive training on this synthetic distribution teaches edge features, and the paper should either add the missing controls or revise the conclusion accordingly.

major comments (4)
  1. [Sec. 3.1, 3.2; Tables 1 and 2] The causal attribution 'removing textual and positional biases fosters accurate edge recognition' (abstract and conclusion) is not supported by the comparisons reported. The experiments compare pretrained CLIP against CLIP finetuned on the debiased synthetic dataset, so any improvement could be due to domain adaptation to synthetic diagrams (arrowheads, line thickness, Mermaid rendering conventions) rather than to the removal of biases. A necessary control is to finetune the same CLIP encoders, with the same training budget and data size, on a synthetic dataset that intentionally retains textual or positional biases (e.g., fixed hierarchical layouts or semantically meaningful labels). Without such a biased-synthetic control, the load-bearing causal claim remains untested.
  2. [Sec. 3.1] The claim that the dataset eliminates positional bias is questionable because node positions are generated by force-directed placement [8] with random initial layout. In force-directed algorithms, the equilibrium configuration is a deterministic function of the graph structure: connected nodes are attracted to each other and high-degree nodes tend to move toward the center. Therefore the geometric arrangement itself carries information about edge existence, and the high retrieval and captioning scores could partly reflect matching these layout statistics rather than reading arrows. I request a quantitative check of residual positional information, for example by training a linear probe that predicts edge existence from node coordinates alone, or by comparing against a control dataset with node positions sampled independently of the graph. The edge-direction probing result (0.509 to 0.860) is more direct evidence for arrow-feature learning, but the paper generalizes beyond this in Sec. 4.2, 4.3, and the conclusion.
  3. [Abstract / Conclusion] The conclusion states that eliminating textual and positional biases 'fosters accurate edge recognition', and the abstract says this is 'confirmed'. Given the missing biased control and the residual layout correlations identified above, this statement is over-strong. I recommend reframing the central claim as: contrastive training on synthetic diagrams with randomized labels and randomized initial layouts improves edge-feature learning in CLIP encoders, with debiasing as a plausible mechanism. The manuscript should also discuss what evidence would distinguish the debiasing mechanism from general domain adaptation.
  4. [Table 2, Sec. 4.3] The comparison between the finetuned encoder+GPT-2 captioner and zero-shot GPT-4o/LLaVA is confounded by supervised training on the synthetic distribution. The GPT-2 decoder is trained on the same 100k synthetic captions, while GPT-4o is zero-shot, so the F1 gap may partly reflect in-distribution decoding. The internal comparison with pretrained CLIP+GPT-2 is more informative, since it holds the decoder and data distribution fixed. Please report captioning results on out-of-distribution or human-drawn diagrams; the paper mentions non-isomorphic graphs only in a sentence, without quantitative results.
minor comments (5)
  1. [Introduction] The text says 'shown in Fig. 2' for the GPT-4o demonstration, but those examples appear in Figure 1, while Figure 2 is the method overview; the same cross-reference issue occurs in Sec. 4.3 where 'consistent with our findings in Fig. 2' should refer to Figure 1.
  2. [Tables 1 and 2] The tables report point estimates without variances or the number of independent runs; given randomness in graph generation, layout initialization, and training, please provide means and standard deviations over multiple seeds or at least confidence intervals.
  3. [Sec. 3.1] The generation protocol is underspecified: it is not stated how the number of nodes is sampled, what the fixed edge probability is, or whether the test split uses the same distribution; these details are needed for reproducibility and for assessing the difficulty of the probing tasks.
  4. [Sec. 4.1] The edge-direction probing protocol should be spelled out more precisely: samples with no edge in either direction are skipped, and balanced undersampling is used, but the exact construction of positive and negative directed-edge samples should be described so that the reported chance level of 0.5 is unambiguous.
  5. [Sec. 4.3] The paper states that the models were tested on diagrams whose graphs are non-isomorphic to any training sample, but gives no quantitative result; please include these numbers or a supplementary table.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's evaluation targets are held out and no parameter is fitted to them; the causal caveat is an experimental-design concern, not a definitional reduction.

full rationale

The paper's derivation chain is: generate a synthetic diagram-caption dataset, finetune a CLIP image encoder with contrastive learning, then measure edge-related features with linear probing, image retrieval, and diagram captioning. None of these evaluation targets is used to fit the encoder: probing trains only a logistic regression on frozen features, retrieval queries are newly generated graphs matched against the test set, and captioning freezes the image encoder while training a GPT-2 decoder and reporting F1 on a held-out split. No load-bearing self-citation or imported uniqueness theorem appears; all cited prior work is external. The skeptical concern, namely that the absence of a biased synthetic control and the force-directed layout's dependence on graph edges weaken the causal claim that debiasing causes the gains, is a validity or confounding issue rather than circularity. The results do not reduce to their inputs by construction, so the appropriate circularity score is 0.

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

The main free parameters are unstated experimental settings rather than fitted constants, and the causal claim rests on assumptions about what the synthetic generation removes and about representativeness of the synthetic distribution.

free parameters (3)
  • edge_generation_probability = not reported
    Directed edges are generated independently with a fixed but unreported probability; this controls graph density and, combined with the largest-component filter, the distribution of graphs seen in training and evaluation.
  • contrastive_training_hyperparameters = not reported
    Learning rate, batch size, temperature, number of steps, and optimizer settings are omitted; only 'sufficiently large number of training steps' is stated, so the exact training recipe is not reproducible.
  • diagram_captioning_hyperparameters = not reported
    GPT-2 decoder training details such as epochs, learning rate, and cross-attention configuration are not given; checkpoint selection is by validation loss on 1% of the training set.
assumptions (4)
  • domain assumption Pretrained CLIP image encoders are a representative foundation for VLM visual encoders, so gains transfer to VLMs.
    The paper motivates finetuning CLIP because it is used in many VLMs (Sec. 3.2), but does not test the finetuned encoder in an actual VLM beyond a GPT-2 decoder.
  • domain assumption Linear probing classification accuracy measures whether the feature representation contains the relevant information.
    Sec. 4.1 uses logistic regression on frozen features; this is a standard but not guaranteed measure of representational content.
  • domain assumption Force-directed placement with random initial layouts produces diagrams whose visual layout carries no information about the graph structure.
    Sec. 3.1 relies on this to remove positional bias; no quantitative check of the correlation between layout and edges is provided.
  • domain assumption The synthetic distribution of small directed graphs with 2 to 8 alphabet-labeled nodes is adequate to draw conclusions about diagram recognition in general.
    Training and evaluation use the same generator; generalization to real-world diagrams is asserted only in a single unquantified sentence in Sec. 4.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Visual Encoder Learn to See Arrows?." pith.science (2026). https://pith.science/paper/RIKMTE7T

@misc{pith2026250519944,
  author       = {Pith},
  title        = {Pith review of: Can Visual Encoder Learn to See Arrows?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RIKMTE7T}},
  note         = {Machine review of arXiv:2505.19944}
}
read the original abstract

The diagram is a visual representation of a relationship illustrated with edges (lines or arrows), which is widely used in industrial and scientific communication. Although recognizing diagrams is essential for vision language models (VLMs) to comprehend domain-specific knowledge, recent studies reveal that many VLMs fail to identify edges in images. We hypothesize that these failures stem from an over-reliance on textual and positional biases, preventing VLMs from learning explicit edge features. Based on this idea, we empirically investigate whether the image encoder in VLMs can learn edge representation through training on a diagram dataset in which edges are biased neither by textual nor positional information. To this end, we conduct contrastive learning on an artificially generated diagram--caption dataset to train an image encoder and evaluate its diagram-related features on three tasks: probing, image retrieval, and captioning. Our results show that the finetuned model outperforms pretrained CLIP in all tasks and surpasses zero-shot GPT-4o and LLaVA-Mistral in the captioning task. These findings confirm that eliminating textual and positional biases fosters accurate edge recognition in VLMs, offering a promising path for advancing diagram understanding.

Figures

Figures reproduced from arXiv: 2505.19944 by the authors.

Figure 1
Figure 1. Examples of diagram captioning by GPT-4o [ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of our approach: (a) training a CLIP model with diagram–caption pairs that eliminate positional and textual biases, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Examples of query images (top row) and the top re [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 17 canonical work pages

  1. [8]

    Thomas M. J. Fruchterman and Edward M. Reingold. Graph drawing by force-directed placement.Software: Practice and Experience, 21(11):1129–1164, 1991. 2

    work page 1991

  2. [1]

    Understanding inter- mediate layers using linear classifier probes.arXiv preprint arXiv:1610.01644, 2016

    Guillaume Alain and Yoshua Bengio. Understanding inter- mediate layers using linear classifier probes.arXiv preprint arXiv:1610.01644, 2016. 3

    arXiv 2016

  3. [2]

    OpenFlamingo: An open-source framework for training large autoregressive vision-language models.arXiv preprint arXiv:2308.01390,

    Adham Awadalla, Iris Gao, Jonathan Gardner, Jack Hessel, Younes Hanafy, Wenzheng Zhu, Karan Marathe, Yacine Bit- ton, Samir Gadre, Shixiang Sagawa, et al. OpenFlamingo: An open-source framework for training large autoregressive vision-language models.arXiv preprint arXiv:2308.01390,

    arXiv

  4. [3]

    Neural codes for image retrieval

    Artem Babenko, Anton Slesarev, Alexandr Chigorin, and Victor Lempitsky. Neural codes for image retrieval. InEur. Conf. Comput. Vis., pages 584–599. Springer, 2014. 3

    work page 2014

  5. [4]

    GeoQA: A geomet- ric question answering benchmark towards multimodal nu- merical reasoning.arXiv preprint arXiv:2105.14517, 2021

    Jiaqi Chen, Jianheng Tang, Jinghui Qin, Xiaodan Liang, Lingbo Liu, Eric P Xing, and Liang Lin. GeoQA: A geomet- ric question answering benchmark towards multimodal nu- merical reasoning.arXiv preprint arXiv:2105.14517, 2021. 2

    arXiv 2021

  6. [5]

    Are we on the right way for evaluating large vision-language models? InAdv

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models? InAdv. Neural Inform. Process. Syst., 2024. 2

    work page 2024

  7. [6]

    Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme, Sebastian Goodman, Xiao Wang, Yi Tay, Siamak Shakeri, Mostafa De- hghani, Daniel Salz, Mario Lu ˇci´c, Michael Tschannen, Ar- sha Nagrani, Hexiang Hu, Mandar Joshi, Bo Pang, Ceslee Montgomery, Paulina Pietrzyk, Marvin Ritter, Alexander J. Piergiovanni,...

    arXiv 2023

  8. [7]

    What you can cram into a single $&!#* vector: Probing sentence embeddings for lin- guistic properties

    Alexis Conneau, German Kruszewski, Guillaume Lample, Lo¨ıc Barrault, and Marco Baroni. What you can cram into a single $&!#* vector: Probing sentence embeddings for lin- guistic properties. InAnnu. Meet. Assoc. Comput. Linguist., pages 2126–2136, Melbourne, Australia, 2018. Association for Computational Linguistics. 3

    work page 2018

Show all 28 references
  1. [9]

    HF Canonical Model Maintainers.gpt2(revision 909a290), URL:https://huggingface.co/gpt2, 2022. 4

    2022

  2. [10]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3

    2023 arXiv

  3. [11]

    Visual instruction tuning.arXiv preprint arXiv:2304.08485,

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.arXiv preprint arXiv:2304.08485,

    arXiv

  4. [12]

    LLaV A-NeXT: Im- proved reasoning, OCR, and world knowledge.https: / / llava - vl

    Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaV A-NeXT: Im- proved reasoning, OCR, and world knowledge.https: / / llava - vl . github . io / blog / 2024 - 01 - 30 - llava-next/, 2024. 4

    2024

  5. [13]

    MathVista: Evaluating mathemat- ical reasoning of foundation models in visual contexts

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. MathVista: Evaluating mathemat- ical reasoning of foundation models in visual contexts. In Int. Conf. Learn. Represent., 2024. 2

    2024

  6. [14]

    CLIP ViT-B/32.https://huggingface.co/ openai/clip-vit-base-patch32, 2021

    OpenAI. CLIP ViT-B/32.https://huggingface.co/ openai/clip-vit-base-patch32, 2021. Hugging Face Model Hub. Accessed: 2025-04-12. 3

    2021

  7. [15]

    CLIP ViT-L/14-336.https://huggingface

    OpenAI. CLIP ViT-L/14-336.https://huggingface. co / openai / clip - vit - large - patch14 - 336,

  8. [16]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    OpenAI. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023. 2

    2023 arXiv

  9. [17]

    Hello, GPT-4o, 2023

    OpenAI. Hello, GPT-4o, 2023. URL: https://openai.com/index/hello-gpt-4o/, Accessed: 2025-02-14. 1, 2, 4

    2023

  10. [18]

    Language models are unsu- pervised multitask learners.OpenAI blog, 1(8):9, 2019

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsu- pervised multitask learners.OpenAI blog, 1(8):9, 2019. 4

    2019

  11. [19]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, p...

    2021

  12. [20]

    Vision language models are blind

    Pooyan Rahmanzadehgervi, Logan Bolton, Moham- mad Reza Taesiri, and Anh Totti Nguyen. Vision language models are blind. InACCV, pages 18–34, 2024. 1, 2

    2024

  13. [21]

    CNN features off-the-shelf: an astound- ing baseline for recognition

    Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. CNN features off-the-shelf: an astound- ing baseline for recognition. InIEEE Conf. Comput. Vis. Pattern Recog. Worksh., pages 806–813, 2014. 3

    2014

  14. [22]

    FlowVQA: Mapping multimodal logic in visual question an- swering with flowcharts

    Shubhankar Singh, Purvi Chaurasia, Yerram Varun, Pran- shu Pandya, Vatsal Gupta, Vivek Gupta, and Dan Roth. FlowVQA: Mapping multimodal logic in visual question an- swering with flowcharts. InAnnu. Meet. Assoc. Comput. Lin- guist., pages 1330–1350, 2024. 1, 2

    2024

  15. [23]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of con- text.arXiv preprint arXiv:2403.05530, 2024

    Team Gemini, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of con- text.arXiv preprint arXiv:2403.05530, 2024. 1, 2 5

    2024 arXiv

  16. [24]

    How well do vision models encode diagram attributes? Inthe ACL 2024 Student Research Workshop, 2024

    Haruto Yoshida, Keito Kudo, Yoichi Aoki, Ryota Tanaka, Itsumi Saito, Keisuke Sakaguchi, and Kentaro Inui. How well do vision models encode diagram attributes? Inthe ACL 2024 Student Research Workshop, 2024. 1

    2024

  17. [25]

    MMMU: A massive multi-discipline multimodal understanding and reasoning benchmark for ex- pert AGI

    Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. MMMU: A massive multi-discipline multimodal understanding and reasoning benchmark for ex- pert AGI. InIEEE Conf. Comput. Vis. Pattern Recog., pa...

    2024

  18. [26]

    Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? In European Conference on Computer Vision, pages 169–186

    Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Yu Qiao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? In European Conference on Computer Vision, pages 169–186. Sprin...

    2024

  19. [27]

    MiniGPT-4: Enhancing vision- language understanding with advanced large language mod- els.arXiv preprint arXiv:2304.10592, 2023

    Deyao Zhu, Jiaming Chen, Xiaoqian Shen, Xiatian Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision- language understanding with advanced large language mod- els.arXiv preprint arXiv:2304.10592, 2023. 3 6

    2023 arXiv

  20. [2021]

    Accessed: 2025-04-12

    Hugging Face Model Hub. Accessed: 2025-04-12. 3

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.