Leveraging VLMs for MUDA：A Category-Specific Prompting with Multi-Modal Low-Rank Adapter

(1) Lack of Experimental Data for Overfitting Mitigation: Although a key motivation of the paper is to address the overfitting issues in VLMs caused by limited training samples and reliance on learnable prompt tokens, the paper does not provide sufficient experimental metrics to demonstrate how effectively the proposed approach mitigates overfitting. Specific metrics, such as validation accuracy trends across training epochs, performance under varied training data sizes, or comparisons of overfitting rates between prompt tuning alone versus prompt tuning combined with LoRA, would provide clearer evidence supporting this motivation. (2) Ablation Study on Fine-Tuning Components: The authors have touched upon the different scenarios of Multimodal LoRA Matrix Selection; however, given the complexity of this integrated approach, a more rigorous ablation study with precise data is necessary to substantiate their claims. Specifically, a comparison of performances with and without prompt tuning, with only LoRA, or with varying ranks and configurations of LoRA matrices would provide deeper insights into the roles and benefits of each part. (3) Minor Errors and Formatting Issues: The paper contains several minor errors that could detract from its overall quality. For instance, in line 265, the term "[CLASS]" appears consecutively. Additionally, in line 410, the phrase "Compared Methods and Evaluation Metric" should begin a new line for improved clarity and presentation. It is recommended that the authors carefully proofread the entire manuscript to address these and other minor issues, ensuring a polished and professional presentation.

1. Although the paper presents the impact of prompt length and pseudo-label thresholds, adding detailed ablation studies on the contribution of different components would strengthen the claims about design choices.
2. While the use of a shared projection layer for cross-modal interaction is explained, the underlying mechanism for how this enhances feature alignment could be expanded, potentially with supporting theoretical or empirical analysis.
3. The paper could benefit from deeper analysis comparing its approach to more recent state-of-the-art methods.

1. The novelty and contribution of this work are quite constrained. The interaction mechanism is borrowed from the MMA paper [1], thus it cannot be deemed as a unique contribution.
2. The figure is not a vector graphic, so it appears blurry.
3. In Eq. (6), this form does not represent a class-specific prompt, as the trainable component is shared across all classes.
4. Figure 1 is plagiarized from Figure 2 of the MMA paper [1].
5. Mistake in Line 285: '$\gamma$is a scaling factor' $\rightarrow$ '$\gamma$ is a scaling factor' .
6. Mistake In Line 288, What does LL mean? L?
7. Mistake in Line 253: 'However, Li et al. Li et al. (2023)'.
8. Mistake in Line 260: 'We introduce bb learn-able tokens'
9. Mistake in Line 265: 'Where, [Class] $[CLASS]$ is replaced according to the label'.
10. Mistake in Line 302: what does "L" refer to on Earth? Does it denote the text encoder "L" or the "L-th" transformer block?
11. Mistake in Line 303: "modify the equation1as".
12. Mistake In equation.(8), what is had? head？
13. The author mentioned applying low-rank matrices to the attention matrices within the transformer structure, but this approach seems to deviate from the formulations in Eq. (11) and Eq. (13). Further, these equations bear a resemblance to those presented in the MMA paper.
14. In Eq. (11) and Eq. (13), the class-specific prompts are not included, where are they?
15. In Fig.(2), there are just too many errors and unexplained parameters.
16. The experiments are not sufficient, at least you should compare your methods with MMA [1].

[1]Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 23826–23837, June 2024a.

The presentation is poor. For example, x denotes the input image in line 173 while x turns into the visual representation of input image. In line 53, the citations should be placed before the period. In line 261, the 'bb learnable tokens'; in line 265, '[CLASS][CLASS] is replaced according to the label'. Equation 6 does not show the class-specific prompting as the authors stated. In equation 8, had should be replace with head. The novelty seems limited. The experiments are limited. For example, the ablation study is absent and the baseline methods are limited.