VisionLLM: Large Language Model is also an Open-Ended Decoder for Vision-Centric Tasks

Large language models (LLMs) have notably accelerated progress towards artificial general intelligence (AGI), with their impressive zero-shot capacity for user-tailored tasks, endowing them with immense potential across a range of applications. However, in the field of computer vision, despite the availability of numerous powerful vision foundation models (VFMs), they are still restricted to tasks in a pre-defined form, struggling to match the open-ended task capabilities of LLMs. In this work, we present an LLM-based framework for vision-centric tasks, termed VisionLLM. This framework provides a unified perspective for vision and language tasks by treating images as a foreign language and aligning vision-centric tasks with language tasks that can be flexibly defined and managed using language instructions. An LLM-based decoder can then make appropriate predictions based on these instructions for open-ended tasks. Extensive experiments show that the proposed VisionLLM can achieve different levels of task customization through language instructions, from fine-grained object-level to coarse-grained task-level customization, all with good results. It's noteworthy that, with a generalist LLM-based framework, our model can achieve over 60\% mAP on COCO, on par with detection-specific models. We hope this model can set a new baseline for generalist vision and language models. The demo shall be released based on https://github.com/OpenGVLab/InternGPT. The code shall be released at https://github.com/OpenGVLab/VisionLLM.

Paper

References (81)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer eUTR7/10 · confidence 4/52023-06-27

Summary

The paper presents a new vision-language model where tasks are specified via a language interface instead of being “hard-coded” into the architecture. The model relies on a pretrained decoder model to simply output the answer as text. E.g., for the object detection task the model would output a set of coordinates directly. The model architecture is rather complicated – a backbone encodes the input image into a multi-scale feature map. The language information is encoded by a Bert model and injected into the multi-scale feature map via cross-attent. Then a DETR model produces a set of visual “tokens” from the feature map. These tokens are then fed to the decoder which outputs the answer in natural text. The paper presents strong results on object detection, grounding and captioning.

Strengths

* The paper provides very strong empirical results. Especially the object detection scores (60ap) are close to SOTA despite this being essentially zero-shot. * The paper is relatively easy to train -- it relies on LORA and pre-trained models. This and the open-sourced code should make the method available to other researchers. * Decoding the output in natural text is a good research direction, it is much more flexible than visual prompt tuning and predefined formats.

Weaknesses

* The method is rather complicated. There are many components and exactly how they interact is not clear from the paper. As a reader, I would probably not be able to implement this from the paper, and I don’t really know what the crucial components are. I have many questions below which I hope the authors can answer to improve the presentation. E.g. the output-format-as-decoding method is not clearly described. * There are also VL tasks (e.g. VQA) which I think should be added to the main paper. Currently only detection/segmentation, grounding and captioning are available.

Questions

1. Could you provide pseudo-code for your implementations? E.g. you say that “The language features are then injected into each scale of visual features through cross-attention”. Knowing exactly how the cross-attention is implemented would be good. 2. What is output-format-as-decoding? The paper says that you “feed the tokens of structural output format as queries to the decoder”, can you explain in more details what this means? 3. The paper says “except a few LoRA parameters” – can you specify how many? 4. Do you know how much object specific knowledge is available in the pretrained DETR model? Is it possible to do an ablation here? I assume the DETR model has been pretrained on e.g. COCO detection, so it’s maybe not really a zero-shot task for the model? 5. Why is two-stage training needed? Are there ablations showing the effects of this? 6. Are the resnet parameters pretrained? The papers just says “we initialize the model with the pre-trained weights of D-DETR, BERT, and Alpaca-7B”.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

n/a

Reviewer 87vj5/10 · confidence 5/52023-06-27

Summary

This work presents a LLM-based framework VisionLLM for vision-centric tasks. VisionLLM treats images as a foreign language and aligns vision-centric tasks with language tasks using language instructions. Extensive experiments show that VisionLLM deliver comparable performance with task-specific models over different vision-centric tasks.

Strengths

The authors demonstrate the feasibility of using large language models as visual decoders in which vision information is treated as a foreign language. With this idea as the key mind, authors introduce a series of methods to aligned vision tasks in a matched format with LLMs, including designing language instruction, design the decoding process and adding additional vocabulary tokens. Conducted experiments demonstrate the feasibility of such pix2seq modelling can be scaled up as a generalist model. Glad to see that some possible limitations of this technical route are discussed for analyzing the gaps in experimental results.

Weaknesses

This work follows the idea of pix2seq to build a generalist model and the authors believe that it is natural to do so, without providing a sufficient explanation of the motivation behind it. Despite the popularity of LLMs recently, we should not blindly apply them to other modalities without in-depth consideration except we can clarify the benefits and demonstrate them accordingly. Authors claim that the reasoning abilitis and world knowledge of LLMs help vision tasks, but only numerical results can be viewed as kind of evidence for this claim. It's not clear and convincing for me from two aspects: - Whether the reasoning abilities and world knowledge of LLMs truly help? and how? - How about scaling up task-specific vision models when compared to VisionLLM? In other words, how can you demonstrate the performance superiority comes from using LLMs as vision decoders, instead of model scaling up? There are a lot of teachnical designs introduced to model vision-centric tasks in a language-matched format. The reasonableness of some of these are hard to be guaranteed, which is subject to further discussion. One of the strangest designs in VisionLLM is the addition of extra tokens in the vocabulary to represent the position values and categories. This approach appears to go against the initial purpose of leveraging LLMs' world knowledge to gain an advantage in visual tasks. Besides,these additionally added tokens may cause ambiguities with those original ones in the vocabulary. I am looking forward to authors' understandings or/and some necessary evidences towards these. The effevetiveness of VisionLLM is sorely evaluated on ResNet-50 and Intern-H backbones. The general applicability is subject to further evaluation.

Questions

Some questions have been listed in the weakness part. Other than that, more questions about the details are as follows: 1. How do you handle the <image> placeholders in the actual input of VisionLLM? Do you replace them with actual image tokens in the latent space? 2. For instance segmentation, the value of N varies for different objtects. How do you detetermine this value? 3. How are "tasks defined by instructions" parsed into formatted queries in "output-format-as-query" decoding? By pre-defined templates for different task categories?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

This paper tries to catch the takeoff of LLMs by building a LLM-based generalist model for vision tasks. But it cannot provide clear and convincing motivation statement, and cannot dive into the rationale of the proposed technical designs. There are also some unclear method statements as listed in my questions. Overall, the idea behind is straightforward. But I think the entire community should be even more careful for these straightforward ideas, and think more in-depthly about whether they are truly reasonable and whether they can deliver real contributions.

Reviewer TifL6/10 · confidence 4/52023-06-27

Summary

This paper introduces VisionLLM, an instruction-following agent that can perform various vision-only (classification/detection/segmentation) and vision-language (captioning/VQA) tasks. The proposed model connects a pre-trained visual backbone with a language decoder Alpaca with a language-aware image-tokenizer. To unify vision-only and vision-language tasks, VisionLLM adopts different language instruction formats. Furthermore, it proposes an “output-format-as-query” framework for efficient parallel decoding for vision-centric tasks.

Strengths

The paper presents an impressive effort in developing a large decoder for vision-only and vision-language tasks, using state-of-the-art multimodal foundational models while treating images as foreign language. The technical details are comprehensive and sound. Extensive experiments demonstrate the effectiveness of the proposed system compared to state of the art Pix2Seq approaches. Ablation studies show insights on how the system perform in single-/multi-task scenarios and different image tokenizing schemes.

Weaknesses

Even though the proposed system is technically sound, it is still quite complicated. It is unclear what its major advantages are compared to previous Pix2Seq methods and task-specific models. I have some doubts about the system design:
 - Is the pre-trained instruction-following LLM (Alpaca) crucial in your system design? Can the system perform as well with a (non-instruction-following) LLaMA? Do more advanced instruction-following agents bring advantages over naive LLMs such as T5? - One of the major disadvantages of using a large-scale pre-trained LLM (Alpaca) is the increased training and inference costs. Could you compare the efficiency of the proposed system to prior task-specific models/generalist models that do not use such large-scale pre-trained decoder? - Why not adopt a ViT-B visual encoder to facilitate comparison to other generalist models such as Uni-Perceiver and Pix2Seq (as in Table 1)? The paper is overall well-written. I found one typo: L166: placeholdersok?

Questions

I have some questions about technical details: - For detection/segmentation tasks, I am confused how the proposed “output-format-as-query” approach (L227-239) works with variable numbers of objects per image. How many “<cls> <x1> <y1> <x2> <y2>” did you send to decoder during training and inference? - How does the “output-format-as-query” approach avoid parallel decoding for image captioning (L232)? As for as I understand, the decoder still adopts casual attention masking and therefore it seems token-by-token generation is necessary. - The proposed system supports customization of number of points for segmentation tasks — does it require specific training paradigms, i.e., balancing the number of masks with different number of points? - Does your system generalize to unseen instructions during test time? - What is the sampling procedure for open-ended tasks? For example, do you use top-k/nucleus sampling?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

No. Please include discussions on the use of foundation models and the potential biases your system might inherit from these pre-trained models.

Reviewer wdrB6/10 · confidence 3/52023-07-06

Summary

This work proposes VisionLLM, a unified framework for vision tasks and vision-language tasks, using natural language task prompts. It demonstrates capabilities in a good variety of tasks.

Strengths

- This is the first attempt to use natural language prompts for vision-centric tasks such as object detection and instance segmentation. I think this work is tackling an important problem. - The language-guided image tokenizer is a novel component to convert image into tokens guided by text. - VisionLLM treats image tokens the same way as text tokens, such that the entire task (including image and class list) can be encoded as one piece of text. This converts vision tasks into sequence generation problem handled by a LLM. (I'm actually not sure if this understanding is correct, so I ask this in the Questions section as well.)

Weaknesses

The framework is not a simplistic one, consisting of various components and types of losses, e.g. using bipartite matching for one type of outputs. Representing image as a set of ${e_i, l_i}$ is also kind of specific. The model can do VQA but I didn't see any VQA results in the paper. Overall my main complaint is that the framework and pipeline seem a bit complicated, but it is designed to incorporate a wide variety of tasks and seems effective at it. It is a good first attempt at using natural language and LLM to tackle vision tasks.

Questions

How are image tokens $T$ and `<text>` passed into the LLM? On line 166 it says `<image>` and `<question>` are placeholder tokens for image tokens and question tokens. I imagine `<question>` will be replaced by the actual question inside the language instruction text (`<text>`), but will the `<image>` token be replaced by the image tokens $T$? Figure 3 seems to imply that $T$ is passed into the LLM separately from `<text>`. I think it's more flexible if all inputs are formulated into one piece of text to pass into the LLM, as one can directly extend it to use more than one images, etc.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

NA.

Reviewer kyDb5/10 · confidence 4/52023-07-07

Summary

This paper construction a vision/language model by passing visual features into a LLM. They trained the model on several standard tasks as well object detection and referring expression by adding special location tokens to the LLM's vocabulary. The model features a text-guided image tokenizer and an efficient decoding approach when generating segmentation or bounding boxes.

Strengths

- The suggestion in section 3.4 seems like a nice way to avoid auto-regressive decoding when it is not needed, although the method was a bit hard to parse. - Showing that the approach of adapting visual features for LLM can work objet detection and segmentation is interesting, these tasks have been less well explored in this area. - The quantitive examples in the appendix at least suggests the model has instruction following capabilities similar to other models tuned on the LLaVA dataset.

Weaknesses

- The scores are decent but not amazing. Only slightly better than pix2seq if using the same backbone, and the CiDER scores are lower then relatively simple models like ClipClap or VL-T5. - The central idea is essentially following the pix2seq method combined with the now pretty well studied method of adapting visual features for a LLM approach, which makes sense but does not feel hugely novel to me. - The authors should consider evaluating follow the methodology of LLAVa given they are using that data.

Questions

Did the authors try pre-training the model? Pre-training to initially learn the alignment for the vision/language component is common practice for these kinds of models. Can the model generalize to outputting bounding boxes for tasks other than refexp or object detection? For example for pointing VQA questions where you output a bounding to answer a question. In section 3.4, what does "...feed the tokens of structural output format as queries to the decoder" mean? That they are used as the initial starting token? What happens if there are multiple objects so they are then multiple class and x1 coordinates to produce? Or if the model needs to interleave text with structured output like in Figure 2a? Table 1 is missing many models that achieve better scores, BEiT-3 is better at detection and captioning for example. I think they should be included for reference

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

I think the authors should at least note some of the potential issues with these kinds of models (bias, potential for abuse my generating misinformation, hallucination, ect.)

Reviewer TifL2023-08-11

Follow-up questions

Thanks the authors for the comprehensive rebuttal. I have a few follow-up questions: **1. *"Alpaca converges easily"* does not really convince me on its importance in the system design. Are there any references on *instruction-following* LLMs are easier to optimize and converge?** **2. Uni-Perceiver-V2 / Pix2Seq v2 are much stronger at segmentation tasks (when using the same ViT-B backbone). Why?** **3. Do you have any evidence on *generalizing to unseen test instruction*? To what extent does it work and when does it fail?**

Authorsrebuttal2023-08-13

**Q1. "Alpaca converges easily" does not really convince me on its importance in the system design. Are there any references on instruction-following LLMs are easier to optimize and converge?** **A1:** We would like to discuss this issue from two perspectives as follows: (1) _**The LLM with instruction-following and parsing capabilities is able to effectively interpret the vision task instructions, which helps to reduce the loss during the early stage of model training.**_ To preserve the language ability of the LLM itself, we freeze the weights LLM during training. If we use LLMs without instruction-following, the frozen LLM struggles to learn this ability. Therefore, we directly utilize the instruction-following LLMs. The example below demonstrates that Alpaca (instruction-following LLaMA) can parse category mapping relationships and the required output format for the task, while LLaMA cannot achieve this. ``` System message: "Below is an instruction that describes a task. Write a response that appropriately completes the request. Input: {input} Output:" Input: "class set: {'person': <c0>, 'car': <c1>, 'table': <c2>, 'cat': <c3>, 'television': <c4>, 'a man in a black hat': <c5>}. What is the class index associated with the class 'a man in a black hat' in the given class set?" # Alpaca output Output: "The class index associated with the class 'a man in a black hat' in the given class set is 5." # LLaMA output "Answer: 0 Input: Given an array of strings, remove the last element of the array. The returned array should be a new array with the last element removed." ``` (2) For _**the detailed loss curves**_ (since we cannot include figures at this time, we provide the loss values every 2000 iterations), the model trained using Alpaca reduces demonstrates faster loss reduction, particularly in the early stages. We do not rule out the possibility that LLaMA can converge if it is unfrozen and allowed to train with more epochs, but at least under our current experimental setting (frozen LLMs, 50 epochs), Alpaca is better at following instructions than LLaMA. ``` # LLaMA loss 6.08, 4.61, 4.46, 4.41, 4.33, 4.22, 4.15, 4.10, 4.04, 4.01, 3.95, 3.87, 3.85, 3.98, 4.36, 4.63 # Alpaca loss 5.26, 4.56, 3.97, 2.85, 2.77, 2.69, 2.60, 2.62, 2.54, 2.56, 2.51, 2.50, 2.50, 2.49, 2.47, 2.45 ``` We will make it clearer in our revised version. **Q2: Uni-Perceiver-V2 / Pix2Seq v2 are much stronger at segmentation tasks (when using the same ViT-B backbone). Why?** **A2:** Firstly, _**VisionLLM is distinct from models for pre-defined tasks (e.g., Uni-Perceiver-V2, Pix2Seq v2), as it has the capability to handle open-ended tasks customized by users, providing greater flexibility and versatility.**_ To be able to unify and flexibly customize tasks, we have made some design and trade-offs in terms of task formulation, model selection, and training methods, which may result in some performance losses, particularly on segmentation tasks. Compared to Uni-Perceiver-V2, our model utilizes polygons with discrete coordinates to represent instance masks, ensuring a uniform task output. _**This results in two levels of performance loss**_: (1) the conversion from mask to polygon representation results in performance degradation, and (2) the conversion of polygon coordinates to integers also incurs performance loss (see L347-L353). While Pix2Seq v2 also employs polygons to represent masks, they use _**128 polygon points**_ (at least 4 times more than our model) to enhance segmentation performance. Additionally, they utilize ensemble methods to _**merge the results of 8 inference runs**_ and _**adopt a crop-then-segment approach**_ to further improve segmentation performance. In contrast, VisionLLM requires an LLM-based decoder to support general user instruction parsing. This restricts us from using more than 32 polygon points on existing hardware, and we have not implemented generalist-model-agnostic techniques (such as ensemble and crop-then-segment) to enhance our segmentation results. We will clarify this in our revised version.

Authorsrebuttal2023-08-13

**Q3. Do you have any evidence on generalizing to unseen test instruction? To what extent does it work and when does it fail?** **A3:** There are four scenarios: (1) **Customized Detection Target:** Our model is trained on the COCO dataset, but it is not limited to detecting only the categories present in the dataset. The categories our model can detect can be question sentences or descriptions in natural language. This flexibility is demonstrated in Figure 2(a) and Figure I in the attached PDF file for rebuttal. (2) **Task Description Flexibility:** Our model supports user input in natural language at the task description level. It's important to note that even for the same task, the descriptions can vary. For example, the grounding task can be described in multiple ways, such as: ``` # Long instruction Please identify all objects belonging to the category set {<expression>: <cls0>}. For each detected object, specify its location within the range <range> by determining the offsets of top-left and bottom-right corners relative to the center point. To indicate the object's class and location, provide the output in the format (c, x1, y1, x2, y2), where 'c' represents the class index starting from 0, and (x1, y1, x2, y2) correspond to the offsets of the bounding box corners. The image is: <image> # Short instruction Please locate the object mentioned in the category set {<expression>: <cls0>}. The image is: <image> ``` These descriptions can have different lengths and sentence structures. We have validated the stability of our model with different prompts in Figure B in the supplementary material, showcasing its ability to generalize to random and unseen descriptions. (3) **Customized Output Formats:** Our method allows for customized output formats, even in object detection tasks. For instance, we can have the format as (c, x1, y1, x2, y2) or (x1, y1, x2, y2, c). We can also control the number and meaning of each point. For example, in Figure 1(a), if we modify the prompt to (x1, y2, x2, y1, c) (outputting the bottom left and top right points), the output can be as follows: ``` "The bounding boxes are [(226.4, 347.4, 363.1, 229.8, <c0>), (441.1, 269.9, 538.6, 183.5, <c1>)]." ``` **(4) Flexible Task Combination:** VisionLLM can also combine different tasks through instructions. For example, with the following instruction, we can combine localization and question-answering tasks to count the number of white cats. ``` "Locate all the objects in the image that are part of the category set {'white cat': <c0>} and output their index of class label starting from 0 and offsets of bounding box coordinates. The bounding box should be a rectangle that covers the entire object. The offsets should be given as top-left and bottom-right corners of the rectangle relative to the center point and should be within <range>. The output format should be (c, x1, y1, x2, y2). The image is: <image>. <cls><x1><y1><x2><y2>...<cls><x1><y1><x2><y2>. How many white cats are in this image?" ``` Despite demonstrating zero-shot capabilities in some unseen scenarios, VisionLLM has some limitations. Due to the limited scale of training data in this version, the connection between vision and language concepts still needs improvement, leading to hallucination issues in VQA and captioning tasks. Additionally, it also struggle with handling specialized terms in niche domains, such as "Magnetic Accelerator".

Reviewer TifL2023-08-20

Thank you for the clarification

The response is comprehensive and have addressed my concerns. I would encourage the authors to put these discussions into the revised version.

Authorsrebuttal2023-08-20

We will add these points in the revised edition. Thanks for your efforts.

Reviewer 87vj2023-08-14

Reply to Rebuttal

Thank you for your efforts in providing the rebuttal. However, some of my questions have not been fully understood, and the current rebuttal has not convinced me and addressed my concerns. I look forward to a more in-depth discussion. In my review comments, I did **NOT** suggest that "***VisionLLM is not the scaling up of Pix2Seq.***" What I would like to discuss is whether the modeling of VisionLLM is reasonable enough. In VisionLLM, visual information (e.g., coordinates) is decoded by a LLM using additionally added tokens. Such modelling raises a series of questions that we really need to take seriously: **How does the using of LLM impact the vision tasks themselves?** It is a common sense that LLM can understand instructions and enable the versatility over different tasks. What I am concerned is the impact of LLM when applied to visual tasks. Is its impacts positive or negative for vision tasks? And WHY? As shown in your provided experiment results in this rebuttal, the proposed VisionLLM-ViT-B has lower AP on Instance Seg. and lower BLEU-4 on Captioning compared to Pix2Seq v2 (ViT-B), when they use the same vision encoder. Is the performance drops of VisionLLM caused by the different model it use? How about using a vision decoder (or a vision decoding head) instead of a language decoder? **How does the using of additionally added tokens impact the effectiveness?** "Extra tokens are a common way" has never been a responsible answer for "Is this modeling correct (although it is often used)? ". After reading authors' responses, I am still confused about: - Does adopting additionally added tokens run counter to the open-ended purpose? For instance, the added classification tokens correspond to classification semantics in a deterministic way as stated in Line224, and their number is limited. What if the target category falls ouside of these added classification tokens? Is this a close-set limitation against open-ended? - Why don't these additionally added markers cause ambiguity with the original markers in the vocabulary? For example, tokens for numbers (0~9) simultaneously exist in both original tokens and additionally added tokens. How does the model handle them? - Why use auto-regressive decoding for the additionally added tokens? When we expand the vocabulary of LLM with additional tokens, we need to decode these tokens auto-regressively as for other tokens in the vocabulary. Compared to directly regress coordinates or perform classification over these tokens, is this auto-regressive decoding for these tokens a correct manner? I hope the authors can deeply consider these important questions and provide responsible answers. Following commonly used methods may deepen the misguidance of the research community sometimes.

Authorsrebuttal2023-08-16

We appreciate your time and effort in reviewing. We are open to engaging in a more in-depth discussion regarding this work. **Q9: The statement of "VisionLLM is not the scaling up of Pix2Seq".** **A9:** In your comments, it was mentioned that "Conducted experiments demonstrate the feasibility of _**scaling up**_ such pix2seq modeling as a generalist model", and also raised the question "How about _**scaling up**_ task-specific vision models when compared to VisionLLM?". Therefore, the statement _**"VisionLLM is not the scaling up of Pix2Seq"**_ has been emphasized to clarify the contribution of this work and avoid potential misunderstanding. **Q10: How does the using of LLM impact the vision tasks themselves? It is a commen (typo: common) sense that LLM can understand instructions and enable the versatility over different tasks. What I am concerned is the impact of LLM when applied to visual tasks. Is its impacts positive or negative for vision tasks? And WHY?** **A10:** _**LLM plays a crucial role in this work due to its parsing and instruction-following capabilities. These capabilities serve as the foundation for defining and understanding open-ended descriptions of vision tasks.**_ We have explained this _**in Q2**_, and provided additional evidence _**in the follow-up question 1 of Reviewer TifL**_. Furthermore, we also explained _**in Q3**_ that simply scaling up the decoding head without web-scale corpus pre-training cannot achieve the same capability as LLM. Therefore, LLM is a positive and crucial component for open-ended vision tasks from this perspective. Based on this point, we introduce VisionLLM, a viable framework with a series of tailored designs that align vision tasks with LLMs. As we mentioned _**in follow-up question 2 of Reviewer TifL**_, due to the need to unify various vision tasks, the framework makes compromises that may affect its performance, especially in segmentation tasks. Additionally, Pix2Seq v2 has 128 polygon points, which is 4 times more than our model. It also uses ensemble and crop-then-segment techniques to enhance the segmentation results, but these techniques are independent of the open-ended task and are not considered in this work. Regarding image captioning, as we explained in _**Q2 of Reviewer kyDb**_, the linguistic capabilities of our model are aligned with LLM, resulting in longer and more detailed responses. If we discard the LLaVA-Instruct-150K dataset and train on COCO Captions, our model could achieve better performance, as demonstrated in the table below: | Model | BLEU-4 | | ---------------- | ------ | | Pix2Seqv2-ViT-B | 34.9 | | VisionLLM-R50 | 31.0 | | VisionLLM-R50* | 33.0 | | VisionLLM-ViT-B | 31.5 | | VisionLLM-ViT-B* | 35.6 | \* indicates discarding the LLaVA-Instruct-150K dataset and training on COCO Captions However, there is an inconsistency between the standard metric (favoring shorter text) and the user experience. If we prioritize alignment with the standard captioning benchmark, it may sacrifice user-friendliness and the overall versatility of the model. In contrast, this work aims to tackle user-defined visual tasks in a flexible manner and provide a practical framework that addresses open-ended tasks effectively. So unlike previous models (e.g., Pix2Seq v2) that pursue superior performance on _**pre-defined vision tasks**_, our approach prioritizes _**open-ended tasks**_ to meet the diverse needs of users.

Authorsrebuttal2023-08-16

**Q11: How does the using of addtionally (typo: additionally) added tokens impact the effectiveness? "Extra tokens are a common way" has never been a responsible answer for "Is this modeling correct (although it is ofen (typo: often) used)? "** **A11:** "Extra tokens are a common way" means it is a widely accepted and proven effective approach, which is supported by numerous peer-reviewed studies in the community [1, 2, 3]. To our best knowledge, there is no evidence indicating that using extra tokens is a suboptimal design for new tasks. _**It is unreasonable to neglect commonly used settings and introduce new, unverified ones.**_ **Q11-1: Does adopting addtionally (additionally) added tokens run counter to the open-ended purpose? For instance, the added classification tokens correspond to classification semantics in a deterministic way as stated in Line224, and their number is limited.** **A11-1**: No, _**the addition of tokens does not run counter to the open-ended task purpose**_. There might be a misunderstanding. As explained in L222-226, the mapping between classification tokens and semantics is fully flexible and can be defined dynamically by the user through their instructions. In other words, the class set of our model is open. It is possible to use a question sentence or a detailed description as a class (see Figure 1(a) and Figure I in the attached PDF file for rebuttal). Although there is a limitation on the number of classification tokens, it is possible to handle an infinite number of classes through sliding-window inference. As demonstrated _**in L136-141 in supplementary material**_, we divide the 1203 categories into 16 groups and predict the results in a sliding-window manner. **Q11-2: Why don't these additionally added markers cause ambiguity with the original markers in the vocabulary? For example, tokens for numbers (0~9) simultaneously exist in both original tokens and (typo: addtionally) additionally added tokens. How dose (typo: does) the model handle them?** **A11-2:** _**No ambiguity will be included.**_ We have avoided this problem when constructing the ground truth data. Only when the query token is a token like <cls>, <x i>, and <y i>, additional tokens such as <p i> and <c i> will be output. Other common words (including numbers (0~9)) will not activate them. **Q11-3: Why use auto-regressive decoding for the addtionally (typo: additionally) added tokens? When we expand the vocabulary of LLM with additional tokens, we need to decode these tokens auto-regressively as for other tokens in the vocabulary. Compared to directly regress coordinates or perform classification over these tokens, is this auto-regressive decoding for these tokens a correct manner?** **A11-3:** There seems to be a misunderstanding. For example, in perception tasks, we utilized the "output-query-as-format" approach to avoid auto-regressive decoding when predicting coordinates and classes. And the exta tokens such as <p i> and <c i> are only output when the query token is <cls>, <x i>, or <y i>. This approach, similar to the format of cloze test, allows us to predict coordinates and classes in parallel, akin to the DETR model. Please refer to _**Common Questions Q1**_ for additional details. &nbsp; Thank you for your review. Your questions and feedback have provided us with an opportunity to reflect and improve. If you have any further questions or suggestions, we would be more than willing to continue the in-depth discussion. &nbsp; [1] Tai, Wen, et al. "exBERT: Extending pre-trained models with domain-specific vocabulary under constrained training resources." Findings of the Association for Computational Linguistics: EMNLP 2020. [2] Koh, Jing Yu, Ruslan Salakhutdinov, and Daniel Fried. "Grounding Language Models to Images for Multimodal Inputs and Outputs." ICML 2023. [3] Hong, Jimin, et al. "AVocaDo: Strategy for Adapting Vocabulary to Downstream Domain." EMNLP 2021.

Reviewer 87vj2023-08-17

Further Reply to Rebuttal

Thanks for your efforts on rebuttal. Sorry for the typos during the quick typing. My concerns has been mostly addressed. And I will increase the final score. Besides, a further question regarding Q11-2 is: As you explain, when the query token is a token like <cls>, <x i>, and <y i>, additional tokens such as <p i> and <c i> will be output. In the original vocabulary, there are also some tokens sharing the same or similar semantics with your additional tokens. Will these tokens remain activated?

Authorsrebuttal2023-08-18

Thank you for your appreciation of our work. **Q12: A further question regarding Q11-2 is: As you explain, when the query token is a token like <cls>, <x i>, and <y i>, additional tokens such as <p i> and <c i> will be output. In the original vocabulary, there are also some tokens sharing the same or similar semantics with your additional tokens. Will these tokens remain activated?** **A12:** In the original vocabulary, there are some tokens that share semantics with the additional tokens, such as "class", "x", "y", and numbers. These tokens are common words that will be predicted using a next-token-prediction approach in tasks such as question answering and captioning. For example, ``` Case 1: "Human: What is the class of the object? Assistant: The class of the object is a cat." Case 2: "Human: x=1, y=2, x+y=? Assistant: x+y=3." Case 3: "Human: How many cats are in the image? Assistant: 10." ```

Reviewer 87vj2023-08-21

Further Questions

Will these tokens (refering to the tokens in the original vocabulary) and the additional tokens be activated simultaneously?

Authorsrebuttal2023-08-21

**Q13: Will these tokens (referring to the tokens in the original vocabulary) and the additional tokens be activated simultaneously?** **A13:** No, as responded in Q11-2, the additional tokens such as <p i> and <c i> are only activated when specific query tokens like <cls>, <x i>, and <y i> are encountered. Despite these tokens (the additional tokens and the original tokens with similar semantics) belong to the same vocabulary, the additional tokens are instruction-tuned to respond to specific queries. To this end, we established a clear link between additional tokens and specific queries and distinguished them from the original tokens when creating the instruction-tuning data. This ensures that they are not activated at the same time.

Reviewer wdrB2023-08-20

Response

Thanks to the authors for the rebuttal. I maintain my opinion that the method is not simplistic enough, but I also maintain my positive rating of a weak accept.

Authorsrebuttal2023-08-20

Thank you for your efforts in the review.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC