Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs

We introduce Cambrian-1, a family of multimodal LLMs (MLLMs) designed with a vision-centric approach. While stronger language models can enhance multimodal capabilities, the design choices for vision components are often insufficiently explored and disconnected from visual representation learning research. This gap hinders accurate sensory grounding in real-world scenarios. Our study uses LLMs and visual instruction tuning as an interface to evaluate various visual representations, offering new insights into different models and architectures -- self-supervised, strongly supervised, or combinations thereof -- based on experiments with over 20 vision encoders. We critically examine existing MLLM benchmarks, address the difficulties involved in consolidating and interpreting results from various tasks, and introduce a new vision-centric benchmark, CV-Bench. To further improve visual grounding, we propose the Spatial Vision Aggregator (SVA), a dynamic and spatially-aware connector that integrates high-resolution vision features with LLMs while reducing the number of tokens. Additionally, we discuss the curation of high-quality visual instruction-tuning data from publicly available sources, emphasizing the importance of data source balancing and distribution ratio. Collectively, Cambrian-1 not only achieves state-of-the-art performance but also serves as a comprehensive, open cookbook for instruction-tuned MLLMs. We provide model weights, code, supporting tools, datasets, and detailed instruction-tuning and evaluation recipes. We hope our release will inspire and accelerate advancements in multimodal systems and visual representation learning.

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer ED7N6/10 · confidence 5/52024-06-17

Summary

This paper presents extensive studies on existing MLLM benchmarks addressing the difficulties involved in consolidating and interpreting results from various tasks for MLLM designs. Moreover, the authors also propose Spatial Vision Aggregator (SVA), a dynamic and spatially-aware connector to fuse vision features with LLMs while reducing vision tokens. In addition, authors also collect high quality visual instruction-tuning data. The proposed model, Cambrian-1, achieves the state-of-the-art performance on multiple MLLM benchmarks.

Strengths

1, Overall writing is good and easy to follow. The motivation is clear. 2, The goal of this work is interesting. A good study for the effects on Large Language Model, Visual Encoder, Multimodal Connector, Data Curation Pipeline, Instruction in the MLLM system. 3, Experiments are extensive, and several findings are useful when designing MLLM models. 4, A new dataset Cambrian-7M is proposed, which may benefit the MLLM fields for further research. 5, The performance looks good compared with LLaVA-Next. If open-sourced, this model can definitely benefit for the community. 6, Open-sourced code and model.

Weaknesses

1, The technical novelty is limited, combining multiple vision experts into MLLM is not new. Moreover, fusing visual tokens dynamically is also not new in dynamic network design. [-] Sphinx: the joint mixing of weights, tasks, and visual embeddings for multi-modal large language models, arixv-2023. [-] Mova: Adapting mixture of vision experts to multimodal context, arxiv-2024. I think there are more close related works than [1][2]. The authors should cite all these works as respect. 2, I also have one question, What if you do not use the dynamic token but only using proposed tuning datasets, compared with LLaVA-1.6. 3, Several findings are well known and also verified from previous works. For example, Finding-6 is common knowledge using high resolution visual encoder. [-] Vary: scaling up the vision vocabulary for large vision-language models 4, Several figures are useless and common knowledge for MLLM community. For example, Fig.1 and Fig.2 can be merged into one figure. 5, The ablation studies for SVA are not enough. For example, which tokens are more important in which datasets? This needs further analysis. Moreover, the effect of increasing instruction tuning data size is not well explored. Given these, I rate this work as weak accept.

Questions

See the weakness part.

Rating

6

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

None

Reviewer K2Un7/10 · confidence 5/52024-07-13

Summary

This paper introduces a multimodal large language models (MLLMs), named Cambrian-1, designed with a vision-centric approach. In current MLLM researches, the choices of visual encoder are not sufficiently explored. This study utilizes MLLM performance as a visual representation evaluator, showing different characteristics over differently trained vision encoders and revealing that various widely-used MLLM benchmarks are disconnected from visual understanding capability but connected to language capability. Furthermore, this study proposes spatial vision aggregator (SVA) to effectively connect vision and language models with spatial inductive bias. Additionally, curation of high-quality visual instruction-tuning dataset and its distribution balancing are discussed. As a result, Cambrian-1 achieves state-of-the-art performances and provides an open cookbook for MLLMs.

Strengths

- This paper is notably well-written and easy to follow. - Section 3.1 shows the limitations of MLLM benchmarks. The finding that several existing benchmarks like MMMU, which were considered important benchmarks in the MLLM field, do not properly evaluate multimodal capabilities is very interesting. - This study releases model weights, code, tools, datasets, and detailed recipes, which is a great contribution to this field.

Weaknesses

- There exists a previous work about vision-language connectors with spatial inductive bias [1]. The comparison or at least discussion between the proposed SVA and C/D-Abstractor [1] is essential but lacks. - There are many overlapping findings with existing studies. For example, language-supervised models are effective [2], high-res encoders are beneficial [3], increasing data size and spatial inductive bias are advantageous for connectors [1], and so on. I believe that re-examining these aspects and analyzing them in different settings has its own contribution due to the empirical nature of this field. Nevertheless, it is difficult to attribute high value to the overlapping findings. - Findings 7 (the second Findings 6 in the paper, seems to be a typo) is not consistent with the results. The finding claims that performance improves with the vision encoder ensemble, but Table 11 does not seem to support this. For example, SigLIP+DINOv2 performs worse than sole SigLIP. [1] Cha, Junbum, et al. "Honeybee: Locality-enhanced projector for multimodal llm." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024. [2] Chen, Xi, et al. "Pali-3 vision language models: Smaller, faster, stronger." arXiv preprint arXiv:2310.09199 (2023). [3] Liu, Haotian, et al. "Improved baselines with visual instruction tuning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.

Questions

In Table 11, why are there two entries for "SigLIP+DINOv2+ConvNext" with different numbers?

Rating

7

Confidence

5

Soundness

3

Presentation

4

Contribution

3

Limitations

This study construct a new dataset based on web search, but it does not appear to address any privacy issues. It would be better to address this issue.

Reviewer K2Un2024-08-12

Thank you for your clear responses. As most my concerns are addressed, I will maintain the original rating of accept.

Reviewer xvQn8/10 · confidence 4/52024-07-13

Summary

The paper conducts a comprehensive study of multimodal LLMs from a vision-centric perspective. Different from the lines of previous literature which aim to propose new architectures/algorithms for multimodal LLMs, this paper carefully splits the design space of visual parts of multimodal LLMs into several individual parts and diagnoses each with controlled experiments. This leads to several innovative conclusions about the visual aspects of multimodal LLMs, including the validity of standard benchmarks, the choices of visual encoders, etc.

Strengths

Generally, the paper is a pleasure to read. - The paper is well-motivated. MLLMs are differentiated from pure LLMs by the visual components. Thus it is quite natural to study the visual aspects of MLLMs. - The paper draws rich connections from the visual representation learning literature, which I see as an original perspective, putting the paper into the appropriate position and delivering more valuable information to a broader community. - The controlled analysis is precise and rigorous with carefully designed experiments. Particularly, the experiments start with an examination of existing benchmarks, which is a prerequisite of all following experiments. - The conclusions are insightful and can be valuable for the future development of multimodal LLMs.

Weaknesses

- The experiments only consider one particular LLaVa-like formulation of MLLMs built upon a pretrained LLMs, while Sota MLLMs like GPT-4o and Reka are more likely to have completely different training diagrams and architectures, e.g., treating images and texts equally and training a native multimodal LLMs from scratch, or training with interleaved visual and text contents instead of fixed image-first formulations. The value of the paper is thus limited. - A minor point: There is no analysis of why the findings hold.

Questions

- What's your opinion of "native" multimodal LLMs? Do you think the findings in the paper will in a way transfer to more advanced models?

Rating

8

Confidence

4

Soundness

4

Presentation

4

Contribution

4

Limitations

The authors have discussed the limitations quite adequately.

Reviewer BwuE7/10 · confidence 3/52024-07-30

Summary

The paper explores Multimodal Large Language Models (MLLMs) and constructs the Cambrian-1 series models. This approach builds a series of advanced MLLMs through five key pillars, achieving exceptional performance in vision-centric tasks and diverse benchmark tests. By exploring different visual encoders, the method designs a novel spatial-aware connector, SVA, to reduce the number of annotations and enhance the integration of visual and language models. The method also curates high-quality visual instruction fine-tuning data from public sources, emphasizing the importance of balanced data distribution and discussing various instruction fine-tuning strategies. Additionally, this paper critically analyzes and supplements existing MLLM benchmarks, introducing the vision-centric benchmark CV-Bench to more accurately evaluate the models' visual-centric capabilities. This approach achieves top performance across diverse benchmarks and excels in visual-centric tasks.

Strengths

The paper aims to bridge the gap in visual understanding by exploring Multimodal Large Language Models (MLLMs) from a vision-centric perspective. By investigating various visual encoders, this method introduces an innovative spatial-aware connector, SVA, which minimizes the need for annotations and improves the synergy between visual and language models. Additionally, the paper offers a critical assessment and enhancement of current MLLM benchmarks, presenting a new vision-centric benchmark, CV-Bench, to more precisely measure the models' visual-centric abilities. Cambrian-1 achieves top performance across diverse benchmarks and excels in visual-centric tasks. The paper is well written and the experiment is well solid.

Weaknesses

No Weaknesses. See questions

Questions

(1) As the number of cross-attention layers (D) and distinct groups of learnable queries (G) increases, the performance does not show continuous improvement (lines 256-258). It is worth exploring whether performance saturation occurs with the increase of (D) and (G). (2) Instruction tuning data collected from open web maybe raise the potential data leakage. And provide a statistical analysis of the data categories. (3) Table 2 suggests that increasing the value t does not continuously enhance performance. The proportion of data varies across different tasks. Additionally explore the data scaling law. (4) From Cambrian 10M to 7M, whether higher data quality results in better model performance. (5) Statistical analysis of the response length and the difficulty, diversity distribution of the instruction data. (6) Compare the performence of current MLLMs like LLaVA and BLIP-2 using the same data and other models (i.e., Minicpm v2.5) with Cambrian-1. (7) Evaluate model performance on high-resolution images or tasks with extreme aspect ratios, (i.e., V*Bench). (8) Determine whether training the visual encoder in all tasks outperforms freezing it, and compare the convergence speed of end-to-end training versus two-stage training. (9) Table 11 indicates that integrating more vision encoders does not necessarily lead to higher performance improvements, as seen with models like MMB, VStar, and MMEP. (10) Provide detailed information of the parameter counts, training duration, and training hyperparameters for different model backbones. (11) The paper improves performance across various tasks by integrating most of the current vision encoders. Could a unified visual encoder be used instead? (12) Some related work needs to be included and discussed. Zhu D, Chen J, Shen X, et al. Minigpt-4: Enhancing vision-language understanding with advanced large language models[J]. arXiv preprint arXiv:2304.10592, 2023. Wang W, Chen Z, Chen X, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks[J]. Advances in Neural Information Processing Systems, 2024, 36. Chen G, Liu X, Wang G, et al. Tem-adapter: Adapting image-text pretraining for video question answer[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023: 13945-13955. Huang X, Wang J, et al. Segment and Caption Anything[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024 Liu Y, Zhang C, et al. Universal Segmentation at Arbitrary Granularity with Language Instruction[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024. Gao P, Han J, Zhang R, et al. Llama-adapter v2: Parameter-efficient visual instruction model[J]. arXiv preprint arXiv:2304.15010, 2023.

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

The authors have discussed the limitations of this paper, and this paper has no direct negative societal impact.

Reviewer ED7N2024-08-12

Rebuttal comments

Thanks for your reply. The rebuttal solves my concern. I keep original rating as weak accept.

Authorsrebuttal2024-08-13

Author Response to Ethics Reviewer VuFm

We thank Reviewer K2Un for requesting an ethics review regarding "data privacy, copyright, and consent" for our submission, and further thank Ethics Reviewer VuFm for their valuable perspective and recommendations regarding such ethical considerations. We appreciate the importance of the points you raise and will address them in detail in our revised manuscript. --- > **Q1: Data Privacy, Security, Copyright, and Consent** We will expand our section explaining our data collection process, focusing on how we ensure compliance with data protection regulations, maintain privacy, respect copyright, and address consent issues. Our datasets come from two sources: existing datasets that we modify and/or use, and new data scraped from the web via our data engine. 1. **Existing datasets:** We only use datasets with licenses that explicitly permit usage and repackaging for research purposes. We currently list all datasets used in Appx. E1, but will add information on their licenses, copyright information, and how we maintain compliance with each dataset’s terms. 2. **Data engine:** Our web agent collects data solely from Wikipedia, which is licensed under CC BY-SA (https://en.wikipedia.org/wiki/Wikipedia:Copyrights). We will add details on how we adhere to this license. 3. **Consent considerations:** - Existing datasets: We will review and report on any consent procedures used by the datasets we incorporate, acknowledging any limitations or concerns. - Wikipedia data: We will mention the implicit consent given by Wikipedia contributors under the site's terms of service. > **Q2: Bias and Fairness** We appreciate the importance of addressing potential biases related to demographic variables. While comprehensive mitigation of bias in MLLMs is not the objective of our work—and, in our opinion, is challenging enough to warrant a standalone work in itself—we recognize its significance. We will include a section including: 1. An analysis of demographic variables in our Cambrian-10M dataset, highlighting any observed patterns or imbalances. 2. A discussion of how training a model on data with inherent biases perpetuates them, as well as challenges and potential unintended consequences of artificially modifying data distributions to address said biases. 3. Acknowledgment that the thorough mitigation of bias in large-scale MLLMs is a complex issue worthy of dedicated research, and we hope our analysis will spur future work focused on this critical issue. > **Q3: Transparency and Interpretability of the Model’s decision-making processes** We appreciate the reviewer's suggestion around model interpretability. However, it's important to note that **understanding the decision-making processes of LLMs / MLLMs is a challenging research problem in itself, with only very preliminary work in this direction.** Recent work [1][2][3] has made initial progress toward the mechanistic interpretability of LLMs, but such methods are not yet applicable to open-source models or MLLMs. While full interpretability of a MLLM's decision-making process is beyond the current state-of-the-art, we can provide some insights into the workings of our SVA module. Specifically, as described in our rebuttal to Reviewer ED7N, we will include an analysis of the attention scores across different vision encoders in our SVA module for various types of images and tasks. This analysis will shed light on how different vision models contribute to the overall visual understanding in different scenarios. > **Q4: Potential Misuse and Mitigation Strategies** We appreciate the reviewer's concern about potential misuse of our technology. However, we must clarify a fundamental misunderstanding: **MLLMs like Cambrian-1 are not capable of generating deepfakes.** They take text + images as input and produce text as output. They cannot generate, manipulate, or synthesize images or videos, as deepfakes entail. That said, we acknowledge that MLLMs can potentially be misused in other ways, and we take these concerns seriously. We will add discussion of potential risks and mitigation strategies, including: 1. **Misinformation:** MLLMs could be used to generate misleading text descriptions of images or to craft false narratives about visual content. To mitigate this, we will provide guidelines for responsible use of our model. 2. **Hallucination:** Like LLMs, MLLMs might generate outputs not grounded in facts or input data. We will clearly communicate this limitation to users. --- Thank you again for your valuable feedback on the ethical considerations of our work. Please let us know if you have any other questions or suggestions. ### References - [1] Bricken et al. Towards monosemanticity: Decomposing language models with dictionary learning. 2023. - [2] Templeton et al. Scaling monosemanticity: Extracting interpretable features from claude 3 sonnet. 2024. - [3] Lieberum et al. Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2. 2024.

Reviewer BwuE2024-08-13

Thank you for the rebuttal. My concerns are addressed by the authors. Therefore, I will keep my score.

Reviewer xvQn2024-08-14

Thanks for your reply! I appreciate your perspective on the native multimodal LLMs.

Program Chairsdecision2024-09-25

Decision

Accept (oral)

© 2026 NYSGPT2525 LLC