VPGTrans: Transfer Visual Prompt Generator across LLMs

While developing a new multimodal LLM (MLLM) by pre-training on tremendous image-text pairs from scratch can be exceedingly resource-consuming, connecting an existing LLM with a comparatively lightweight visual prompt generator (VPG) becomes a feasible paradigm. However, further tuning the VPG part of the MLLM still suffers from indispensable computational costs, i.e., requiring thousands of GPU hours and millions of training data. One alternative solution is to transfer an existing VPG from any existing MLLMs for the target MLLM. In this work, we for the first time investigate the VPG transferability across LLMs, and explore a solution to reduce the cost of VPG transfer. We first study the VPG transfer across different LLM sizes (e.g., small-to-large), and across different LLM types, through which we diagnose the key factors to maximize the transfer efficiency. Based on our observation, we design a two-stage transfer framework named VPGTrans, which is simple yet highly effective. Through extensive experiments, we demonstrate that VPGTrans helps significantly speed up the transfer learning process without compromising performance. Remarkably, it helps achieve the VPG transfer from BLIP-2 OPT$_\text{2.7B}$ to BLIP-2 OPT$_\text{6.7B}$ with over 10 times speed-up and 10.7% training data compared with connecting a VPG to OPT$_\text{6.7B}$ from scratch. Further, a series of intriguing findings and potential rationales behind them are provided and discussed. Finally, we showcase the practical value of our VPGTrans approach, by customizing two novel MLLMs, including VL-LLaMA and VL-Vicuna, with recently released LLaMA and Vicuna LLMs.

Paper

References (61)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer iPdz6/10 · confidence 4/52023-06-25

Summary

This paper mainly focuses on the transferability of visual prompt generator in VL-LLM. The authors conduct extensive experiments among different LLM types and sizes. By combining the experiment results they propose a new VPN training pipeline that engages projector warmup stage and vanilla finetuning stage.

Strengths

1. This paper is well written. 2. The experiments are comprehensive.

Weaknesses

1. I wonder if different structures for the projector, e.g. more layers, would lead to large performance gap. 2. It is kind of weird to directly compare the training time between training the whole model and using the proposed transfer method, since you have to first get another VL-LLM. In this way the time used for training the original smaller VL-LLM should also be considered. It is inappropriate to claim the method is more efficient just because of other off-the-shelf models.

Questions

Please refer to the above weaknesses.

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

No obvious negative societal impact observed.

Reviewer mi8b9/10 · confidence 5/52023-07-02

Summary

The paper discusses the transfer of a visual prompt generator (VPG) across different vision-language language models (VL-LLMs) to reduce computational costs. VL-LLMs include a VPG module that bridges the gap between vision and language, encoding visual inputs into fixed-length soft prompts. To reduce the cost of building new VL-LLMs, the authors propose a two-stage transfer learning framework, VPGTrans, for VPG transfer across different LLM sizes and types. VPGTrans includes a projector warm-up (stage-1) and vanilla fine-tuning (stage-2). They conduct exploratory analyses to determine key factors for transfer efficiency and show that VPG transfer across frozen LLMs is feasible and can lead to substantially lower computational costs. The authors demonstrate the practical value of VPGTrans by customizing two novel VL-LLMs with recently released LLMs. The paper highlights the need for further research into VPG transfer across LLMs to further reduce the cost of building new VL-LLMs. Overall, the paper's main contributions include identifying the feasibility and effectiveness of VPG transfer across LLMs, smaller LLM sizes leading to more speed-up and better performance during transfer across LLM sizes, and VPG trained on smaller LLMs outperforming larger ones in most conditions, and then propose a transfer learning framework with empirical evidence supporting its efficiency.

Strengths

1. The multimodal large language model (MM-LLM) is considered the most promising path towards achieving Artificial General Intelligence (AGI). However, a major challenge is the significant computational cost associated with training an MM-LLM. This obstacle hinders the widespread research and exploration of this topic. The presented paper offers an excellent pilot study that demonstrates the feasibility of VPG transfer. By showcasing the viability of VPG transfer, the paper provides valuable insights into mitigating the computational cost of training MM-LLMs. This finding has significant implications for future research in this field. The pilot study serves as a promising starting point, paving the way for further investigations and advancements in the development of MM-LLMs with higher efficiency. 2. I quite enjoy the didactic/exploratory nature of this paper. I like that the paper reads like an investigation, and begins with a preliminary exp, as well as diagnosing reasons for the case, before then presenting the VPGTrans method. And then they delve into the transfer of VPGs under two settings. This, from the shallower to the deeper, feels more organic and the lessons learned along the way are insightful. Also the introduction with clear bullet point allows easy reading. 3. The proposed method, VPGTrans, can be fairly straightforward but effective, which is a merit. 4. The experimental work in this study is extensive and solid. It includes validation across a wide range of LLM sizes and types, providing thorough coverage. 5. This work yields numerous interesting and meaningful findings, which will serve as important empirical evidence for future explorations in LLM/MM-LLM efficiency. The conclusions drawn from this research have the potential to guide and shape subsequent investigations in this field. 6. Additionally, this paper makes good contributions by introducing two new state-of-the-art MM-LLMs in the community: VL-LLaMA and VL-Vicuna. Overall, I like this paper, and I believe this work will show important impact to the community.

Weaknesses

1. The major possible limitation of this method is its reliance on the existence of a well-functioned VPG model, and without this assumption, this work is of less meaningfulness. How does the quality of the existing VPG influence the transfer efficacy and efficiency? 2. This work evaluates on merely the captioning and VQA tasks. It would be interesting to see the performance on other different VL tasks and more benchmark datasets, to see the broad-coverage trends. 3. On the other hand, the proposed VPGTrans, though straightforwardly simple, relies too much on the empirical tuning, e.g., when to inherit the raw VPG and when to tuning or fixing some modules. Are there any intuitions and theoretical supports to build up the method? 4. When transferring the VPG, what has been changed of the visual soft prompts? Are the visual prompts of the raw VPG changed, to adapt to the new LLM backbone? 5. The claim of ‘our VPGTrans helps achieve a BLIP-2 ViT-G OPT2.7B→6.7B transfer with less than 10% of the GPU hours and 10.7% training data required for the original model training’ can be a little misleading, as the comparison between VPGTrans and the training of the raw VL-LLM is not strict fair and counterpart; should compare the VPGTrans with the hard VPG transfer method and then draw the conclusion. 6. Lack of exploration of potential limitations or drawbacks of the proposed method, which may limit its generalizability or practical usefulness. 7. How did you choose the specific VL-LLMs for your experiments? 8. In Section 4.2, you mention that the VPG trained on smaller LLMs may outperform larger ones in most conditions. Can you provide more explanation and insights for this observation of why it brought better task performance? 9. Can you explain in more detail the warm-up training of the linear projector and how it can prevent performance drops? 10. In Section 6, you mention that you customized two novel VL-LLMs with your VPGTransfer framework. How generalizable do you think this framework is to other novel VL-LLMs, and what steps would need to be taken to adapt it to different models?

Questions

See Weaknesses.

Rating

9: Very Strong Accept: Technically flawless paper with groundbreaking impact on at least one area of AI/ML and excellent impact on multiple areas of AI/ML, with flawless evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

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

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

I do not foresee any potential for negative societal impact from this work.

Reviewer GwBd4/10 · confidence 5/52023-07-06

Summary

* This paper presents an interesting study on how to effectively transfer the small VL model to the large VL model, which is very practical under the fact that the LLMs are very expensive to finetune. * The empirical study shows the effectiveness of the methods.

Strengths

* The paper proposes an effective word converter to better speed up the model training on a large model * The stage-wise training gives good results for the final performance.

Weaknesses

* The paper writing is very complex to understand * This technical contribution is very limited. The core is word converter, which better aligns the knowledge between large LLMs and small LLMs. * The performance improvement is very obvious. Because the whole model training involves the small LLM, which contains more model parameters. * The whole training pipeline is very complex. There are multiple stage for training.

Questions

* It is better to include the GPU hours and training data of a small model for a fair comparison * Is it possible to inject the Alignment loss during the tuning projector stage?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

3 good

Presentation

2 fair

Contribution

3 good

Limitations

see above

Reviewer 7KJM6/10 · confidence 4/52023-07-11

Summary

The paper presents a technique for VPG (Visual Prompt Generator) transferability across LLMs where the transferability can be between LLMs of different sizes (eg: OPT125M -> OPT2.7B) or across LLMs of different types (eg: OPT350M-> Flan T5base). This is achieved using a two stage strategy where a projector is learned by freezing VPG and LLM in the first stage and vanilla finetuning of VPG and projector in the second stage. The proposed technique significantly reduce training time (specifically when transfer from small to large LLMs).

Strengths

1. The VPG transferability can significantly improve the training time/cost in some cases (from LLM_small->LLM_large) can have many practical applications when training larger LLMs. 2. The proposed technique consistently reduces the training time/cost compared to baseline (TFS) while achieving similar performance

Weaknesses

Novelty: The proposed technique is a standard technique used in fine-tuning a model for a specific task where the newly added layer (head which is similar to projector here) is initially trained with backbone frozen and once trained, the whole network (backbone and head) is trained/fine-tuned on the target data. Please see [a, b] for references that talk about standard fine-tuning strategies. The authors extend this to VPG transferability between two LLMs. Although, the technique is good for practitioners, I don't see any technical insights in the proposed technique. Generalization: Since the proposed technique uses a limited vocabulary from the datasets COCO, SBU-captions [1.4M image-caption pairs] for the word converter, the performance of the target LLM on the datasets with domain GAP could be affected. It wold be interesting to see the performance of the target LLM (and comparison with source LLM) on domain specific datasets such as RSVQA [c] and PathVQA [d] datasets. Robustness: To check the validity of the proposed technique, It would also be interesting to see the affect of transferability on Robustness of the target model which [target-LLM compared src-LLM] can be done by evaluating on datasets AdVQA [e], VQA-CE [f] or VQA-Rephrasing [g] [a] https://cv-tricks.com/keras/fine-tuning-tensorflow/ [b] https://lightning-flash.readthedocs.io/en/stable/general/finetuning.html [c] https://rsvqa.sylvainlobry.com/#overview [d] https://paperswithcode.com/dataset/pathvqa [e] https://adversarialvqa.org/ [f] https://paperswithcode.com/dataset/vqa-ce [g] https://facebookresearch.github.io/VQA-Rephrasings/

Questions

I would like authors to discuss the three points raised above : 1) Novelty, 2) Generalization, [3] Robustness

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

None

Reviewer mi8b2023-08-12

Response to author rebuttal

Thank you for addressing my inquiries. The majority of my concerns have been satisfactorily resolved, and I generally concur with the authors' assertions. Furthermore, regarding one of the points posed by Reviewer #GwBd, I would like to inquire whether the overall training pipeline might become intricate and challenging to implement in practical scenarios. I've observed that the proposed framework involves a substantial number of pipelines, have these been automated when using it? Do I need to care about the intermediate operations? In regard to one of my questions (Weakness#5), specifically the comparison between VPGTrans and hard VPG transfer, I would greatly appreciate a comprehensive analysis encompassing factors about operational efficiency, for instance, execution time and costs. Currently, you have presented the end task performance instead of the efficiency comparisons.

Authorsrebuttal2023-08-13

Re-rebuttal to Reviewer #mi8b

Thanks for your reply! It is encouraged to see that you are satisfied with our rebuttal. In the following, we try to address your current concerns point to point. --- **Q1: Whether the overall training pipeline might become intricate and challenging to implement in practical scenarios. I've observed that the proposed framework involves a substantial number of pipelines, have these been automated when using it? Do I need to care about the intermediate operations?** **A:** Thanks for the question. Actually, our VPGTrans is quite simple to implement. - In the code level, the main workload lies in the word converter trainer implementation, while the code can be implemented by adding a new file without modifying the original VL-LLM code. Thus, the practitioners do not need to dive into the implementation details of the original VL-LLM to modify it. - In the experiment level, the framework can be automated without considering the intermediate results. Some researchers may concern about the hyperparameters about the intermediate operations. We want to clarify that the main hyperparameter, the 1st stage's learning rate, is super robust in the training (cf lines 524-528 Appendix), which does not require careful adjustment. --- **Q2: I would greatly appreciate a comprehensive analysis encompassing factors about operational efficiency, for instance, execution time and costs. Currently, you have presented the end task performance instead of the efficiency comparisons.** **A:** Sorry for missing your point. We have a comparison between VPGTrans and VPG transfer (denoted as VPG inherit in our paper) in **C.3 of our Appendix**, where the speed-up rate is reported in Table 5 and the convergence curve comparison is shown in Figure 13. A main conclusion is that our VPGTrans can achieve better performance for over 74% Transfer-Task variants, among which **our VPGTrans can also achieve better acceleration rate for over 69% conditions**. More analysis will be added in the revision.

Reviewer mi8b2023-08-14

Response to author rebuttal-maintain my initial rating.

Thank the authors for the response. Overall, I think this paper has great research significance and I stick to my previous rating.

Authorsrebuttal2023-08-16

We are looking forward to your feedback

Dear Reviewer#7KJM, We would like to express our sincere appreciation for your great efforts and valuable feedback on our paper. Your comments are essential to help us improve the quality of our work. To address your concerns of generalization and robustness, we compare the source model and target model on RSVQA, PathVQA, AdVQA, VQA-CE, and VQA-Rephrasing datasets. We kindly hope that you can take some time to re-evaluate our paper based on our replies. If you have any further concerns or questions, please do not hesitate to let us know. We will be happy to address them promptly. Best Regards, Paper#1904 Authors

Reviewer 7KJM2023-08-21

Final

I have read the rebuttal and also other comments. I am happy with the experiments on Robustness and generalization. So I am increasing my score.

Authorsrebuttal2023-08-16

We are looking forward to your feedback

Dear Reviewer#GwBd, Thanks so much for your great efforts and valuable feedback on our paper. Your comments are essential to help us improve the quality of our work. We have carefully addressed your concerns and questions in our responses. For example, we report the cost of training the source model for a fair comparison, and we systematically discuss the possibility of incorporating the alignment loss. We kindly hope that you can take some time to re-evaluate our paper based on our replies. If you have any further concerns or questions, please do not hesitate to let us know. We will be happy to address them promptly. Best Regards, Paper#1904 Authors

Area Chair 1uWB2023-08-18

Hi Authors, Thanks a lot for your rebuttal. We are urging all the reviewers to respond and will take your input into consideration as we make the final recommendation. Thanks! Best,

Reviewer iPdz2023-08-18

Thank the authors for the detailed response. I have no further questions.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC