pith. sign in

arxiv: 2606.29462 · v1 · pith:XPMBDJGNnew · submitted 2026-06-28 · 💻 cs.CV

MIRROR: Aligning Semantic Relations from Language to Image via Gromov--Wasserstein

Hong-Han Wang , Yuntao Wang , Hu Ding This is my paper

Pith reviewed 2026-06-30 07:26 UTC · model grok-4.3

classification 💻 cs.CV
keywords multimodal large language modelsGromov-Wasserstein alignmentrelational consistencyoptimal transport regularizationvision-language modelsgeometric alignmentcross-modal transfer
0
0 comments X

The pith

Semi-Inverse Gromov-Wasserstein alignment transfers relational priors from language to visual representations with a closed-form solution.

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

Multimodal large language models inherit relational structure from language but often lose it when mapping to images because standard alignment only matches individual token meanings. MIRROR adds a geometric regularization term derived from the Semi-Inverse Gromov-Wasserstein problem to enforce that relations between concepts remain consistent after the language-to-vision crossing. The formulation yields a unique closed-form solution that directly prescribes the target relational structure for visual tokens. This term integrates into existing decoder-only transformers at the layer, head, and token levels with no added parameters or runtime cost. The result improves consistency on relational tasks while leaving general vision-language performance unchanged.

Core claim

We show that this formulation admits a unique closed-form solution that prescribes the ideal visual relational structure implied by language geometry and cross-modal coupling.

What carries the argument

The Semi-Inverse Gromov-Wasserstein (SI-GW) problem, an inverse optimal-transport alignment that matches visual token relations to language-derived relational priors.

If this is right

  • Relational consistency improves on tasks that require applying language-learned relations in visual scenes.
  • General vision-language performance remains unchanged because the regularization preserves individual token semantics.
  • The closed-form solution permits direct application to long token sequences without quadratic scaling issues.
  • Integration requires only targeted placement at layer, head, and token levels and adds no parameters or inference overhead.

Where Pith is reading between the lines

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

  • The same closed-form relational transfer could be tested on other modality pairs that share language backbones, such as audio or 3D.
  • Because the method needs no task-specific tuning, it may reduce the data required for relational reasoning fine-tuning.
  • The geometric view of cross-modal relations opens the possibility of diagnosing which layers already carry strong relational structure before alignment.

Load-bearing premise

Relational priors already present in language representations can be moved into visual representations through SI-GW regularization without distorting the original visual semantics.

What would settle it

An experiment in which models trained with the SI-GW term show no gain on relational consistency benchmarks or exhibit measurable drops on standard vision-language tasks relative to identical baselines without the term.

Figures

Figures reproduced from arXiv: 2606.29462 by Hong-Han Wang, Hu Ding, Yuntao Wang.

Figure 1
Figure 1. Figure 1: Vision can override correct language priors. A GQA [20] example where Qwen2-VL-7B predicts “leafy” from “green and brown” without the image, but salient bare branches bias it toward an in￾correct answer. See Appendix A for more examples. This observation led us to a key in￾sight: effective cross-modal understand￾ing requires aligning not just what concepts mean, but how they relate to each other, that is, … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the MLLM architecture. An image is encoded into visual tokens via a vision encoder and adapter, then concatenated with text tokens and fed into the language model. Av denotes visual self-attention from the vision encoder, C the cross-modal attention from text to visual tokens, and At the causal text self-attention within the language model. concepts that are close (or far) in language space sho… view at source ↗
Figure 3
Figure 3. Figure 3: MIRROR extraction and alignment. At and C are extracted from separate LLM layers, and Av from the final ViT layer, with selection applied at the layer, head, and token levels. MIRROR con￾structs the target geometry Dbv from D t and C, and drives D v toward it via the SI-GW loss. Layer Decoupling for Sta￾ble Geometry Transfer In decoder-only MLLMs, At and C are sub-matrices of the same causal attention map,… view at source ↗
Figure 4
Figure 4. Figure 4: Cross-attention analysis from two perspectives. (a,b) Layer-wise: intermedi￾ate layers achieve balanced information richness and peak concentration. (c) Head￾wise (Layer 16): low-entropy heads yield clearer cross-modal alignment. Design Choice Validation We empirically justify the specific architectural instantiations of each component. Coupling layer selection [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
read the original abstract

Multimodal Large Language Models (MLLMs) inherit rich relational priors from their language backbones, yet often fail when asked to apply these relationships in visual contexts. We trace this failure to a structural blind spot: projection-based alignment trains each visual token to carry the right semantics, but never asks whether the relationships between concepts survive the crossing from language to vision. To address this, we propose MIRROR (Mapping Inter-concept Relations from language to visual Representation via Optimal-transport-based Regularization), a geometric regularization framework that transfers relational priors from language to vision by exploiting the rich relational structure encoded in language representations. Specifically, we derive a surrogate loss from the proposed Semi-Inverse Gromov-Wasserstein (SI-GW) problem, an inverse geometric problem that aligns visual representations with language-derived relational priors. We show that this formulation admits a unique closed-form solution that prescribes the ideal visual relational structure implied by language geometry and cross-modal coupling. The structure of the formulation also enables efficient computation, making it applicable to long token sequences. Applying SI-GW inside decoder-only Transformers requires careful design. We introduce targeted strategies at the layer, head, and token levels to ensure stable extraction without additional parameters or inference cost. MIRROR improves relational consistency while preserving performance on general vision-language tasks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The manuscript proposes MIRROR, a geometric regularization framework for multimodal large language models that transfers relational priors from language to vision using a Semi-Inverse Gromov-Wasserstein (SI-GW) alignment. It derives a surrogate loss from the SI-GW problem and claims this admits a unique closed-form solution prescribing the ideal visual relational structure, with efficient computation suitable for long token sequences; targeted integration strategies at layer/head/token levels in decoder-only Transformers are introduced to enable stable extraction without added parameters or inference cost, yielding improved relational consistency while preserving general vision-language performance.

Significance. If the uniqueness and closed-form solution for SI-GW hold under the stated cross-modal assumptions and the method transfers relations without distorting visual semantics, this would supply an efficient, parameter-free regularization approach grounded in optimal transport that directly targets a structural limitation of projection-based alignments in MLLMs. The emphasis on applicability to long sequences and zero-overhead integration represents a practical strength.

minor comments (3)
  1. The abstract states that SI-GW 'admits a unique closed-form solution' but provides no equation reference or proof outline; adding a brief pointer to the relevant theorem or derivation in the main text would improve accessibility.
  2. Notation for the language-derived relational priors and the cross-modal coupling term should be defined explicitly on first use to avoid ambiguity when the SI-GW objective is introduced.
  3. The description of layer/head/token-level strategies would benefit from a short pseudocode or diagram clarifying how the closed-form solution is applied inside the Transformer without altering inference cost.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their constructive summary and positive assessment of the manuscript, including recognition of the closed-form SI-GW solution, computational efficiency for long sequences, and parameter-free integration. The recommendation for minor revision is appreciated.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The abstract frames the SI-GW problem as derived from optimal transport principles and states that the formulation admits a unique closed-form solution, but provides no equations, proofs, or self-citations that reduce this uniqueness or the surrogate loss to fitted inputs, prior self-work, or definitional equivalence. No load-bearing steps matching the enumerated circularity patterns are identifiable from the given text, and the central claim remains independent of its own outputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Ledger constructed from abstract only; full paper may contain additional parameters or assumptions not visible here.

axioms (1)
  • standard math Standard properties and existence results for Gromov-Wasserstein distances in optimal transport
    The SI-GW problem and its closed-form solution rely on established optimal transport theory.
invented entities (1)
  • Semi-Inverse Gromov-Wasserstein (SI-GW) problem no independent evidence
    purpose: To derive a surrogate loss that aligns visual representations with language-derived relational priors
    New formulation introduced to solve the cross-modal relational alignment task.

pith-pipeline@v0.9.1-grok · 5769 in / 1237 out tokens · 51305 ms · 2026-06-30T07:26:44.510234+00:00 · methodology

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

49 extracted references · 10 canonical work pages · 6 internal anchors

  1. [1]

    GPT-4 Technical Report

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    Advances in neural information processing systems35, 23716– 23736 (2022)

    Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Men- sch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems35, 23716– 23736 (2022)

    2022

  3. [3]

    Advances in neural information processing systems 32(2019)

    Altschuler, J., Bach, F., Rudi, A., Niles-Weed, J.: Massively scalable sinkhorn dis- tances via the nyström method. Advances in neural information processing systems 32(2019)

    2019

  4. [4]

    Advances in neural information processing systems30(2017)

    Altschuler, J., Niles-Weed, J., Rigollet, P.: Near-linear time approximation algo- rithms for optimal transport via sinkhorn iteration. Advances in neural information processing systems30(2017)

    2017

  5. [5]

    Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond

    Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  6. [6]

    Qwen2.5-VL Technical Report

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., et al.: Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 (2025) 16 H.-H. Wang et al

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    SIAM Journal on Scientific Computing37(2), A1111–A1138 (2015)

    Benamou, J.D., Carlier, G., Cuturi, M., Nenna, L., Peyré, G.: Iterative bregman projections for regularized transportation problems. SIAM Journal on Scientific Computing37(2), A1111–A1138 (2015)

    2015

  8. [8]

    Advances in Neural Information Processing Systems37, 69814–69839 (2024)

    Chen, J., Nguyen, B.T., Koh, S., Soh, Y.S.: Semidefinite relaxations of the gromov- wasserstein distance. Advances in Neural Information Processing Systems37, 69814–69839 (2024)

    2024

  9. [9]

    In: Eu- ropean Conference on Computer Vision

    Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. In: Eu- ropean Conference on Computer Vision. pp. 370–387. Springer (2024)

    2024

  10. [10]

    In: Forty-third International Conference on Machine Learning (2026),https://openreview.net/ forum?id=PK0zeaGAJ9

    Cheng, K., Huang, J., Song, J., Zhang, W., Han, B., Ding, H.: Achieving struc- turally robust gromov wasserstein distance via adaptive dual-mask. In: Forty-third International Conference on Machine Learning (2026),https://openreview.net/ forum?id=PK0zeaGAJ9

    2026

  11. [11]

    Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality (March 2023),https://lmsys.org/ blog/2023-03-30-vicuna/

    2023

  12. [12]

    Ad- vances in neural information processing systems26(2013)

    Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Ad- vances in neural information processing systems26(2013)

    2013

  13. [13]

    Advances in Neural Information Processing Systems36(2024)

    Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P.N., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems36(2024)

    2024

  14. [14]

    SIAM Journal on Imaging Sciences7(3), 1853–1882 (2014)

    Ferradans, S., Papadakis, N., Peyré, G., Aujol, J.F.: Regularized discrete optimal transport. SIAM Journal on Imaging Sciences7(3), 1853–1882 (2014)

    2014

  15. [15]

    Journal of Machine Learning Research22(78), 1–8 (2021)

    Flamary, R., Courty, N., Gramfort, A., Alaya, M.Z., Boisbunon, A., Chambon, S., Chapel, L., Corenflos, A., Fatras, K., Fournier, N., et al.: Pot: Python optimal transport. Journal of Machine Learning Research22(78), 1–8 (2021)

    2021

  16. [16]

    In: European Conference on Computer Vision

    Fu, X., Hu, Y., Li, B., Feng, Y., Wang, H., Lin, X., Roth, D., Smith, N.A., Ma, W.C., Krishna, R.: Blink: Multimodal large language models can see but not per- ceive. In: European Conference on Computer Vision. pp. 148–166. Springer (2024)

    2024

  17. [17]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6904–6913 (2017)

    2017

  18. [18]

    Gromov, M., Katz, M., Pansu, P., Semmes, S.: Metric structures for Riemannian and non-Riemannian spaces, vol. 152. Springer (1999)

    1999

  19. [19]

    Han, J., Tong, S., Fan, D., Ren, Y., Sinha, K., Torr, P., Kokkinos, F.: Learning to see before seeing: Demystifying llm visual priors from language pre-training (2025), https://arxiv.org/abs/2509.26625

    work page arXiv 2025

  20. [20]

    In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition

    Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition. pp. 6700–6709 (2019)

    2019

  21. [21]

    In: Proceedings of the 41st International Conference on Machine Learn- ing

    Huh, M., Cheung, B., Wang, T., Isola, P.: Position: the platonic representation hy- pothesis. In: Proceedings of the 41st International Conference on Machine Learn- ing. pp. 20617–20642 (2024)

    2024

  22. [22]

    Mistral 7B

    Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  23. [23]

    In: International conference on machine learning

    Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large language models. In: International conference on machine learning. pp. 19730–19742. PMLR (2023) MIRROR 17

    2023

  24. [24]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10965–10975 (2022)

    2022

  25. [25]

    Li, M., Su, N., Qu, F., Zhong, Z., Chen, Z., Li, Y., Tu, Z., Li, X.: Vista: Enhancing vision-text alignment in mllms via cross-modal mutual information maximization (2025),https://arxiv.org/abs/2505.10917

    work page arXiv 2025

  26. [26]

    In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hal- lucination in large vision-language models. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, Singapore (2023),https://openreview.net/forum? id=xozJw0kZXF

    2023

  27. [27]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Lin, J., Yin, H., Ping, W., Molchanov, P., Shoeybi, M., Han, S.: Vila: On pre- training for visual language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26689–26699 (2024)

    2024

  28. [28]

    In: International Conference on Ma- chine Learning

    Lin, T., Ho, N., Jordan, M.: On efficient optimal transport: An analysis of greedy and accelerated mirror descent algorithms. In: International Conference on Ma- chine Learning. pp. 3982–3991. PMLR (2019)

    2019

  29. [29]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tun- ing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 26296–26306 (2024)

    2024

  30. [30]

    io/blog/2024-01-30-llava-next/

    Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-next: Improved reasoning, ocr, and world knowledge (January 2024),https://llava-vl.github. io/blog/2024-01-30-llava-next/

    2024

  31. [31]

    Advances in neural information processing systems36(2024)

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems36(2024)

    2024

  32. [32]

    In: Second Workshop on Representational Alignment at ICLR 2025 (2025)

    Masry, A., Rodriguez, J.A., Zhang, T., Wang, S., Wang, C., Feizi, A., Suresh, A.K., Puri, A., Jian, X., Noël, P.A., et al.: Alignvlm: Bridging vision and language latent spaces for multimodal understanding. In: Second Workshop on Representational Alignment at ICLR 2025 (2025)

    2025

  33. [33]

    Foundations of computational mathematics11, 417–487 (2011)

    Mémoli, F.: Gromov–wasserstein distances and the metric approach to object matching. Foundations of computational mathematics11, 417–487 (2011)

    2011

  34. [34]

    Michel, P., Levy, O., Neubig, G.: Are sixteen heads really better than one? Ad- vances in neural information processing systems32(2019)

    2019

  35. [35]

    In: International conference on machine learning

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

    2021

  36. [36]

    In: Ku, L.W., Martins, A., Srikumar, V

    Ren, J., Guo, Q., Yan, H., Liu, D., Zhang, Q., Qiu, X., Lin, D.: Identifying semantic induction heads to understand in-context learning. In: Ku, L.W., Martins, A., Srikumar, V. (eds.) Findings of the Association for Computational Linguistics: ACL 2024. pp. 6916–6932. Association for Computational Linguistics, Bangkok, Thailand (Aug 2024).https://doi.org/1...

    work page doi:10.18653/v1/2024.findings-acl.412 2024

  37. [37]

    Journal of Machine Learning Research25(363), 1–52 (2024)

    Rioux, G., Goldfeld, Z., Kato, K.: Entropic gromov-wasserstein distances: Stability and algorithms. Journal of Machine Learning Research25(363), 1–52 (2024)

    2024

  38. [38]

    Probabil- ity Theory and Related Fields70(1), 117–129 (1985)

    Rüschendorf, L.: The wasserstein distance and approximation theorems. Probabil- ity Theory and Related Fields70(1), 117–129 (1985)

    1985

  39. [39]

    Advances in Neural Information Process- ing Systems34, 8766–8779 (2021) 18 H.-H

    Séjourné, T., Vialard, F.X., Peyré, G.: The unbalanced gromov wasserstein dis- tance: Conic formulation and relaxation. Advances in Neural Information Process- ing Systems34, 8766–8779 (2021) 18 H.-H. Wang et al

    2021

  40. [40]

    ACM Trans

    Solomon, J., Peyré, G., Kim, V.G., Sra, S.: Entropic metric alignment for corre- spondence problems. ACM Trans. Graph.35(4), 72:1–72:13 (2016)

    2016

  41. [41]

    In: Forty-third International Conference on Machine Learning (2026),https://openreview.net/forum?id=dxwRakfEFm

    Song, J., Huang, J., Cheng, K., Han, B., Ding, H.: LoBCD-GW: A fast and data- dependentalgorithmforcomputinggromov-wassersteindistancevialocalizedblock coordinate descent. In: Forty-third International Conference on Machine Learning (2026),https://openreview.net/forum?id=dxwRakfEFm

    2026

  42. [42]

    In: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)

    Tao,M.,Huang,Q.,Xu,K.,Chen,L.,Feng,Y.,Zhao,D.:Probingmultimodallarge language models for global and local semantic representations. In: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024). pp. 13050–13056 (2024)

    2024

  43. [43]

    Llama 2: Open Foundation and Fine-Tuned Chat Models

    Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bash- lykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  44. [44]

    Villani, C., et al.: Optimal transport: old and new, vol. 338. Springer (2009)

    2009

  45. [45]

    Viswanathan, K., Gardinazzi, Y., Panerai, G., Cazzaniga, A., Biagetti, M.: The geometry of tokens in internal representations of large language models (2025), https://arxiv.org/abs/2501.10573

    work page arXiv 2025

  46. [46]

    In: ACL 2019-57th Annual Meeting of the Association for Computational Linguistics, Proceedings of the Conference

    Voita, E., Talbot, D., Moiseev, F., Titov, I., Sennrich, R.: Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In: ACL 2019-57th Annual Meeting of the Association for Computational Linguistics, Proceedings of the Conference. pp. 5797–5808 (2020)

    2019

  47. [47]

    x.ai: Grok-1.5 vision preview.https://x.ai/blog/grok-1.5v(2025)

    2025

  48. [48]

    MiniCPM-V: A GPT-4V Level MLLM on Your Phone

    Yao, Y., Yu, T., Zhang, A., Wang, C., Cui, J., Zhu, H., Cai, T., Li, H., Zhao, W., He, Z., et al.: Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  49. [49]

    In: The Twelfth International Conference on Learning Representations (2023)

    Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision- language understanding with advanced large language models. In: The Twelfth International Conference on Learning Representations (2023)

    2023