List Items One by One: A New Data Source and Learning Paradigm for Multimodal LLMs

Set-of-Mark (SoM) Prompting unleashes the visual grounding capability of GPT-4V, by enabling the model to associate visual objects with tags inserted on the image. These tags, marked with alphanumerics, can be indexed via text tokens for easy reference. Despite the extraordinary performance from GPT-4V, we observe that other Multimodal Large Language Models (MLLMs) struggle to understand these visual tags. To promote the learning of SoM prompting for open-source models, we propose a new learning paradigm: "list items one by one," which asks the model to enumerate and describe all visual tags placed on the image following the alphanumeric orders of tags. By integrating our curated dataset with other visual instruction tuning datasets, we are able to equip existing MLLMs with the SoM prompting ability. Furthermore, we evaluate our finetuned SoM models on five MLLM benchmarks. We find that this new dataset, even in a relatively small size (10k-30k images with tags), significantly enhances visual reasoning capabilities and reduces hallucinations for MLLMs. Perhaps surprisingly, these improvements persist even when the visual tags are omitted from input images during inference. This suggests the potential of "list items one by one" as a new paradigm for training MLLMs, which strengthens the object-text alignment through the use of visual tags in the training stage. Finally, we conduct analyses by probing trained models to understand the working mechanism of SoM. Our code and data are available at \url{https://github.com/zzxslp/SoM-LLaVA}.

Paper

References (46)

Scroll for more · 34 remaining

Similar papers

Reviewer gxtp7/10 · confidence 4/52024-05-09

Summary

The paper introduces a novel dataset and learning paradigm aimed at enhancing Multimodal Large Language Models (MLLMs) by teaching them to list items in images one by one. This method is designed to improve the association of visual objects with corresponding textual descriptions through a process called Set-of-Mark (SoM) Prompting. By tagging images with numeric identifiers and prompting models to list these tags, the study demonstrates significant improvements in visual reasoning and a reduction in hallucinations across several MLLM benchmarks.

Rating

7

Confidence

4

Ethics flag

1

Reasons to accept

* Clear Motivation: The paper clearly articulates the motivation to enhance Multimodal Large Language Models (MLLMs) with the ability to list items, which could improve how these models handle visual objects in combination with textual descriptions. * Innovative Approach: Introducing a novel dataset and a "list items one by one" learning paradigm is innovative. This method allows models to better integrate and process visual and textual information. * Strong Results: The paper reports substantial benefits in terms of reduced hallucinations and improved visual reasoning across several MLLM benchmarks.

Reasons to reject

The research primarily focuses on t specific aspect of multimodal learning, which might be seen as a limited application area. Expanding this methodology to foundational models could make the work more impactful.

Questions to authors

n/a

Reviewer v9mL5/10 · confidence 4/52024-05-11

Summary

This paper studies the Set-of-Mark capability of existing VLMs other than GPT4-V, and proposes a valid solution to construct training data for getting this capability.

Rating

5

Confidence

4

Ethics flag

1

Reasons to accept

1. Interesting problem of studying the Set-of-Mark capability of the VLM, the study from the paper indicates all VLMs can get similar capability if the model is trained specifically. 2. Promising but not the major contribution of the paper, the authors show that adding the training data designed for learning SoM capabilities can improve the VLM performance in general.

Reasons to reject

1. The contribution of the work is limited. It is yet another paper which constructs a training data, train the model, and get the capability. 2. The paper is not ready to publish yet. Many details about the approach are missing, I tried to read the section 4 multiple times but still cannot get the full picture of the construction process. More illustrations of the approach and the constructed examples will be necessary for reader to understand the proposed approach. Maybe move the examples from Appendix 1.6 into the main content. Instead, section 3 should go to Appendix, conjunction with moving the primary investigation results / conclusion goes to the introduction. 3. One key motivation is that the Set-of-Mark capability is important for visual referring prompting. As pointed in the related work, there is another important type of the referring technology which put user inputs (point, bounding boxes, sketch) directly into the text input. A comparison of Set-of-Mask approach and direct referring approach will make the paper more motivated. Also, the author should discuss and cite more papers here, there are many recent work on this direction.

Questions to authors

In section 3.2 it is unclear what is the binary classifier, and other details about the classifier.

Reviewer L6Ef7/10 · confidence 5/52024-05-11

Summary

This paper presents a new data source to better learn multimodal LLMs by collecting listing and QA data with set-of-mark (SoM) prompting from GPT4V. A new learning paradigm is also designed to enhance the visual understanding ability of MLLMs. Results show that the new dataset can improve the performance of MLLMs on various benchmarks and enable SoM prompting ability of MLLMs.

Rating

7

Confidence

5

Ethics flag

1

Reasons to accept

- Overall I think this is an interesting paper with some new insights and sufficient experimental study. The proposed method to collect SoM prompting data to improve MLLMs is well-motivated and interesting. Enabling the SoM prompting ability of MLLMs may be an effective method to close the gap between open-sourced MLLMs and GPT4V. - Extensive experiments are conducted to show the effectiveness of the proposed method.

Reasons to reject

- The improvements look marginal on some of the benchmarks. In the abstract, it is mentioned that the method can " reduce hallucinations in Multimodal LLMs" but the improvements on POPE are not very significant. Can you provide an analysis of this phenomenon? - How to ensure the quality of the dataset? GPT4V may produce wrong results. Would this affect the final performance of the models trained with the dataset?

Reviewer nfpB8/10 · confidence 4/52024-05-14

Summary

This paper presents a simple and compelling idea for enriching the spatial and fine-grained reasoning of multimodal large language models (MLLMs): rather than solely condition language generation on an image, the key idea is to augment an image with a “set of marks” (alphanumeric tags overlaid on the image) that are associated with key locations or reference points. As an example, given an image of a crowd at a concert, a plausible set of marks might include a different tag on each of the members of the crowd, a mark on the stage, and a mark on each of the musicians and their corresponding instruments. Note that this idea of “marking” is an alternative input scheme that type checks with existing MLLMs that take in arbitrary images and generate text; at inference time, one can still provide untagged images, or they can manually tag an image with marks, and have the MLLM generate outputs referring to individual marked objects. The bulk of the paper 1) discusses the failures of existing MLLMs such as LLaVa-1.5 when it comes to reasoning over “marked images”, 2) curating a small “marked” dataset of just 10-30K examples of images with labeled outputs just consisting of “List items one by one: <tag 1>: <description> ...”, and 3) evaluating MLLMs trained with this additional data on a spectrum of evaluation tasks. In all, the proposed “marking” scheme is simple, scalable, and leads to MLLMs that outperform non-marked variants **in the absence of marking data during evaluation**, especially on tasks requiring fine-grained reasoning. The paper also presents case studies verifying that MLLMs trained with marking can integrate such marks for complex reasoning over full scenes, beyond the capabilities in the training data.

Rating

8

Confidence

4

Ethics flag

1

Reasons to accept

This is a simple and scalable method for enriching arbitrary MLLMs to perform better at fine-grained reasoning tasks, and provides a new “input modality” for controllable interaction with MLLMs for more complex tasks. The work does a great job in identifying the limitations of existing MLLMs and MLLM datasets, and present a scalable approach for generating such “marked” data from existing image datasets (one that scales far beyond the data generated for this work). The evaluations show that MLLMs trained with “marked” data not only can use such “marks” for complex reasoning, but also show that just training with this marked data enriches the abilities of MLLMs to perform fine-grained and spatial reasoning in the **absence of marks during evaluation** as well.

Reasons to reject

While the existing evaluations are compelling, it would make this work even stronger to explicitly evaluate the trained MLLMs on evaluations that provide calibrated/absolute metrics on spatial and compositional reasoning (e.g., GQA question answering accuracy, RefCOCO bounding box prediction, etc.). I recommend that the authors look into strengthening the existing evaluations for the final version of the paper.

Questions to authors

Is there a reason that the results don’t include a 7B parameter model? Table 2 only shows a 13B variant of the mark-augmented models. Do we need GPT-4V to produce the descriptions for each of the tagged objects? If we’re using MSCOCO images to seed our image dataset anyway, can we just use the object labels from that dataset (and possibly other datasets like Visual Genome that have individual region/object captions)?

Reviewer YAe16/10 · confidence 4/52024-05-14

Summary

The paper proposes a new dataset for Set-of-Mark (SoM) prompting for MLLMs. SoM allows MLLMs to associate objects with text by assigning numbers to different objects. Adding the proposed dataset to existing visual instruction tuning datasets improves the performance on five visual reasoning tasks. The SoM dataset is created by modifying an existing dataset (MS-COCO) that has human annotations corresponding to bounding boxes, and captions. An existing model (SAM) is used to segment the images prior to adding numbers on top of each segment then use GPT4-v to generate the textual descriptions. The authors fine-tune the pertained LLaVA-1.5 model using LLAVA-1.5 fine-tuning data in addition to the newly created dataset and show improvements on five visual reasoning tasks.

Rating

6

Confidence

4

Ethics flag

1

Reasons to accept

- Interesting approach & well though-out analysis on using SoM data to improve the visual reasoning performance of MLLMs. - Fine-tuning an MLLM using the proposed dataset improves the performance on five visual reasoning tasks.

Reasons to reject

- The proposed approach and dataset are only applied to one MLLM. It would be good to try other MLLMs to see if the results generalize. - It would be good to evaluate the in-context learning (without fine-tuning) performance of using exemplars from the new dataset.

Reviewer HXE78/10 · confidence 3/52024-05-23

Summary

Set-of-Mark (SoM) prompting (placing numbered tags on images before feeding them to the models) has been shown to improve visual reasoning on GPT-4V. This paper creates a dataset helping MLLMs to acquire the SoM visual prompting ability, and shows that MLLMs other than GPT-4V (such as LLaVA-1.5) can also benefit from this technique. The SoM models are evaluated on 5 benchmarks and show improvement even without the tags. The authors also investigated the mechanism of SoM.

Rating

8

Confidence

3

Ethics flag

1

Reasons to accept

Interesting work, with three-fold contributions: probing SoM's mechanism, proposing an approach adding the SoM capability for other MLLMs with a dedicated dataset. The work's insight can be beneficial for the research community. The paper is clearly written, well structured, and easy to follow. The appendix also contains useful details. The authors stated that the dataset and models will be released.

Reasons to reject

The approach's effectiveness is only demonstrated with LLaVA-1.5. Probing other types or sizes of VLMs can be useful to better understand its applicability. This adds additional overhead and needs to retrain the instruction tuning stage. It seems unclear if fine-tuning an already instruction-tuned model using the SoM data would work.

Questions to authors

In Figure 3 (a), increasing the listing data from 75k to 100k seems to lead to a slightly lower accuracy for both the 7B and 13B models. Do the authors have an idea of why? Adding tags may obstruct very small objects. Can the technique still be helpful in this scenario? Can the authors indicate how long the training was for both models? The LaTeX style seems different from the official version. Typo: - Paragraph 1: "can implicitly aligning [...]" -> "can implicitly align [...]" - 4.3, "GPT-4V is asked to generate mulit-turn" -> "multi-turn" Minor: - Figure 2: it's a bit hard to read the tags behind the attention maps, they can be emphasized. Also, some red patches seem to be going a bit off the image.

Reviewer nfpB2024-06-04

Post-Rebuttal Response

Thank you for answering my questions and taking the time to add the additional 7B results; I thought this paper was very strong to begin with, and that opinion has not changed with the rebuttal - I will definitely be pushing for acceptance at this time! (Would definitely be great to run out the additional benchmarks for the final version though; just as a way to make the paper even stronger).

Authorsrebuttal2024-06-04

Author Response

Thank you so much for the appreciation and support of our work! Here are the results on GQA ( it takes longer than expected to add GQA results with some bug fixes). We will continue exploring additional benchmarks to show the effectiveness of our method. | Model | GQA | |-------------------|------| | LLaVA-1.5-7B | 62.0 | | SoM-LLaVA-1.5-7B | 62.7 | | LLaVA-1.5-13B | 63.3 | | SoM-LLaVA-1.5-13B | 63.8 |

Reviewer HXE72024-06-05

Reponse to the authors

Thanks to the authors for the details and additional results on other open-sourced base models, showing consistent improvement. I believe that adding the details and results provided can make the paper clearer. I increased my score as the answers dealt with my main concerns.

Authorsrebuttal2024-06-05

Author response

Thank you so much for the appreciation and support of our work, and the prompt response!

Reviewer L6Ef2024-06-06

Thanks for your response. After reading other reviews as well as the rebuttal, I think the proposed SoM dataset and the results presented in the paper would be beneficial for future research. Therefore, I would upgrade my rating to "accept".

Authorsrebuttal2024-06-06

Author response

Thank you so much for the support of our work!

Authorsrebuttal2024-06-06

Author follow-up

Dear reviewer, Thank you again for reviewing our paper and providing useful feedback! Since the author-reviewer discussion is coming to an end, please let us know if there are any unresolved concerns. Moreover, we would be happy to hear your thoughts on our replies, especially the additional results on llava-1.5-7b and xgen-mm, and if you could consider increasing the score given our responses above. Thanks, Authors

Authorsrebuttal2024-06-06

Author follow-up

Dear reviewer, Thank you again for reviewing our paper and providing useful feedback! Since the author-reviewer discussion is coming to an end, please let us know if there are any unresolved concerns. Moreover, we would be happy to hear your thoughts on our replies, and if you could consider increasing the score after reading other reviews and rebuttals. Thanks, Authors

Program Chairsdecision2024-07-10

Decision

Accept

© 2026 NYSGPT2525 LLC