Elamur Research
☰
home / papers / 2024 / Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs

Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs

Written for the Research library (claude-opus-5-5) from the paper text. Original — arxiv.org/pdf/2401.06209

Summary

Multimodal LLMs owe much of their recent progress to strong language backbones, but their visual side usually relies on a single CLIP encoder. This paper argues that this encoder is a weak link. MLLMs, GPT-4V included, get very simple visual questions wrong and often make up justifications for the wrong answers. To expose this, the authors mine "CLIP-blind pairs": images that CLIP embeds almost identically, while the self-supervised DINOv2 model treats them as clearly different. Human annotators then write easy questions about each pair's difference, which forms the MMVP benchmark. Most tested open and proprietary MLLMs score below chance, and even GPT-4V trails humans by a wide margin. The authors sort the failures into nine visual patterns, such as orientation, counting and viewpoint. Bigger CLIP models and datasets fix few of these patterns, and CLIP's weak spots predict where MLLMs fail. As a first remedy, the authors propose Mixture-of-Features, which feeds DINOv2 features into the MLLM. The work shows that visual representation learning, not only language scaling, currently limits multimodal systems.

Key points

  • How pairs are selected: Pairs come from ImageNet and LAION-Aesthetics. A pair qualifies when its CLIP-ViT-L-14 cosine similarity is above 0.95 and its DINOv2-ViT-L-14 similarity is below 0.6.
  • How MMVP is scored: A model earns a point for a pair only if it answers the questions for both images correctly.
  • Models evaluated: LLaVA-1.5, InstructBLIP, Mini-GPT4, GPT-4V, Gemini and Bard. Most fall under random-guess level, and GPT-4V's gap to humans exceeds 50%.
  • Effect of scaling: A companion analysis of CLIP variants, MMVP-VLM, finds that 7 of the 9 visual patterns remain unsolved by any large-scale CLIP-based model. Scaling data or model size alone is therefore not enough.
  • Mixture-of-Features variants:
    • Additive-MoF blends CLIP and DINOv2 features linearly. It improves grounding but weakens instruction following.
    • Interleaved-MoF alternates visual tokens from both encoders. It improves grounding while keeping instruction following intact.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Other

Cited by

115 papers of our library cite it (of the 140673 whose bibliographies we read); 8 of them have a page here, 107 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2401.06209 · arXiv 2401.06209

@misc{tong2024eyes,
  title = {{Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs}},
  author = {Tong, Shengbang and Liu, Zhuang and Zhai, Yuexiang and Ma, Yi and LeCun, Yann and Xie, Saining},
  year = {2024},
  doi = {10.48550/arxiv.2401.06209},
  eprint = {2401.06209},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2401.06209}
}

References

10 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeruIPS, 2022. find our card
  2. Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. find our card
  3. Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In CVPR, 2023. find
  4. Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. 2022. find
  5. Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023. find our card
  6. Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. find our card
  7. Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructblip: Towards general-purpose vision-language models with instruction tuning, 2023. find our card
  8. Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICML, 2021. find
  9. Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023. find
  10. Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. [47] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Kra… find our card