RA-TTA: Retrieval-Augmented Test-Time Adaptation for Vision-Language Models

Q3: How is the diversity in text descriptions ensured during the generation process.

To ensure semantic diversity in text descriptions, we utilized multiple prompts that have different semantics when querying LLMs. Furthermore, we employed a high temperature setting of 0.99 and queried each prompt multiple times.

Q4: Therefore, the computational and storage overhead compared to TPT should be provided.

We appreciate the reviewer for pointing out this important issue. 

The cost of constructing databases is inevitable for RAG-based approaches. However, the database is built only once offline prior to test time, and the initial cost of database construction is not significantly high because leveraging an advanced tool like FAISS typically used for RAG-based approaches can efficiently manage indexing and storage, minimizing the overhead.

In the context of test-time adaptation, the efficiency of adaptation at test time is indeed a critical factor to consider. In contrast to previous TTA methods, RA-TTA needs to retrieve proper external images for each test image, potentially posing a significant bottleneck. However, the efficiency of RA-TTA is comparable to the representative test-time adaptation method, TPT, as shown in Table R3, where we compared the GPU inference time per sample (i.e., s/sample) for TPT [2] and RA-TTA. This is because FAISS, an advanced and efficient search engine typically used for RAG-based approaches, enables fast nearest-neighbor search. Moreover, our description-based adaptation does not require backpropagation (i.e., training-free), which also contributes to the efficient inference of RA-TTA.

Table R3: Inference time for TPT and the proposed RA-TTA. We report the average GPU inference time per sample (s/sample) measured on a single RTX4090 for the FGVC aircraft, Stanford cars, and RESISC45 datasets. The "Avg." column indicates the average inference time of 3 datasets.

Furthermore, RA-TTA does not need a backpropagation for adaptation, resulting in low GPU peak memory compared to TPT that updates learnable prompts using backpropagation as shown in Table R4.

We have added this efficiency analysis in Appendix D.4 of the revised manuscript.

Table R4: GPU peak memory for TPT and the proposed RA-TTA. We report the GPU peak memory (MB) on the FGVC aircraft, Stanford cars, and RESISC45 datasets. The "Avg." column indicates the average GPU peak meamory of 3 datasets.

Once again, we appreciate the reviewer highlighting valuable discussion points in our study. We hope our response has clarified the contents of the paper, and we are more than willing to engage in further discussion regarding any aspects that remain unclear.

---

[1] "What Does a Platypus Look Like? Generating Customized Prompts for Zero-shot Image Classification.", In ICCV, 2023
[2] "Test-Time Prompt Tuning for Zero-Shot Generalization in Vision-Language Models.", In NeurIPS, 2022

We sincerely thank the reviewer for your insightful comments. Below, we are more than willing to clarify the reviewer’s concerns one by one.

Q1: The authors should discuss how different LLMs might vary in their effectiveness at generating useful descriptions, and how this impacts retrieval relevance and quality.

We appreciate the reviewer for providing a valuable discussion point. We conducted additional experiments using text descriptions generated by Claude3-Sonnet for this analysis.

• We also compared the performance of CuPL [1] to investigate the quality of text descriptions because CuPL can achieve better performance with improved descriptions. As shown in the first two rows of Table R1, CuPL with the descriptions of GPT works better than CuPL with those of Claude, implying that the text descriptions generated by GPT are better than those of Claude. Thus, we can conclude that different LLMs can affect description quality even with the same generation prompts.
• Also, the last two rows of Table R1 show that RA-TTA with the descriptions of GPT outperformed RA-TTA with those of Claude. This indicates that the better descriptions facilitate the adaptation of RA-TTA more effectively by improving the relevance and quality of retrieved images.
• Notably, RA-TTA with the text descriptions of Claude outperformed the performance of corresponding CuPL (i.e., CuPL+Claude3 vs. RA-TTA+Claude3), demonstrating that RA-TTA effectively leverages external knowledge, even when the quality of the description is not very good.

We have added this analysis in Appendix D.5 of the revised manuscript.

Table R1: Performance when using descriptions generated by ChatGPT and Claude3. We report the top-1 accuracy (%) on the FGVC aircraft, Stanford cars, and RESISC45 datasets. The employed LLMs are indicated with the method name. The best results are in bold. The "Avg." column indicates the average accuracy of 3 datasets.

Q2: An analysis of how the sampled database size affects adaptation performance should be provided.

We are grateful to the reviewer for bringing up this crucial discussion point. With respect to the database size, we analyzed the adaptation performance, including adaptation accuracy, storage costs, and inference time. Note that we used random sampling to build the subset of the original database.

• In the first two rows of Table R2, as the database size increases, accuracy improves alongside storage costs. This result is expected because the likelihood of including informative samples for adaptation increases with a larger database, albeit at the expense of increased storage costs. 
• Regarding inference time, thanks to efficient search engines, such as FAISS, the increase in latency remains negligible, as shown in the last row of Table R2.

Additionally, because web-scale datasets often contain noises, we can improve the adaptation accuracy even with a smaller database (i.e., less storage costs) by conducting sampling with advanced filtering tools.

We have added this analysis to Appendix D.6 of the revised manuscript.

Table R2: Effects of the database size on accuracy, storage costs, and inference time of RA-TTA. We report the top-1 accuracy (%), storage costs (MB), and average GPU inference time per sample (s/sample) measured on a single RTX4090 for the Stanford cars dataset. The percentage represents the proportion of the sampled database to the original database. The values in the parentheses indicate the number of images in a database.

Dear Reviewer: DF2M, 

Thank you for your review and valuable feedback, which have greatly contributed to enhancing the quality of our paper. This is a kind reminder that we only have one and a half days (about 36 hours) before the author/reviewer discussion period ends. We submitted an answer addressing your questions, including the usage of a different LLM, an analysis of the effects of database size, and computation and storage overhead.  We hope that our response has provided sufficient clarification. Please let us know if you have any further inquiries or require additional information.

Best regards,
Authors.

Q3: Figure 3 is not very clear in presenting the proposed method, specifically in the prototype-based retrieval part.

Thanks for your comments, which have helped improve the clarity of our paper. We conjecture that the absence of a legend for icons in the prototype-based retrieval part may have led to obscurity. Accordingly, we have edited and made overall revisions to Figure 3, including the addition of a legend. We have added the edited Figure 3 to the revised manuscript.

W2: The proposed method achieves inferior performance than TPT [2] and RLCF [3] in IN-V2 and IN-A. The authors should analyze the reasons behind this phenomenon.

Thanks for letting us know your concern regarding performance. We speculate that the out-of-distribution property of the datasets might compromise the performance of the proposed RA-TTA for IN-V2 and IN-A. Specifically, IN-V2 and IN-A include samples that exist in the class boundary and require rare features rather than generalizable ones. However, the text descriptions of RA-TTA represent the general features of each class. Given that RA-TTA searches the features of a test image amidst the text descriptions, the absence of rare features in the descriptions renders it impossible to identify the features of a test image and to retrieve appropriate external images, resulting in performance decline. Nonetheless, RA-TTA attained the second-best performance, with a negligible difference to the best performance, i.e., 0.16% and 1.31%, respectively.

In this study, we generated candidate text descriptions offline in accordance with CuPL [4]. At test time, we adaptively identified the relevant descriptions among the generated descriptions without additional generation. That is, the candidate text descriptions are fixed during test time. Thus, a possible solution is to adaptively generate text descriptions on the fly. We leave this exploration as an intriguing avenue for future work.

Once again, we appreciate the reviewer highlighting valuable discussion points in our study. We hope our response has clarified the contents of the paper, and we are more than willing to engage in further discussion regarding any aspects that remain unclear.

---

[1] "Understanding Retrieval-Augmented Task Adaptation for Vision-Language Models.", In ICML, 2024
[2] "SuS-X- Training-Free Name-Only Transfer of Vision-Language Models.", In ICCV, 2023
[3] "Neural Priming for Sample-Efficient Adaptation.", In NeurIPS, 2023
[4] "What Does a Platypus Look Like? Generating Customized Prompts for Zero-shot Image Classification.", In ICCV, 2023

Thanks for your very constructive comments and reasonable concerns, which have helped us improve the paper quality significantly. Below, we would like to provide point-to-point responses to all the raised questions.

Q1(W1): The differences and advantages with [1] should be claimed

The proposed RA-TTA focused on test-time adaptation (TTA), emphasizing the importance of adaptive responses to test images provided online. The primary novelty of RA-TTA, compared to [1], lies in its adaptiveness. Specifically, RA-TTA adaptively retrieves external images for each test image based on its pivotal features. In contrast, [1] inherently lacks this adaptive retrieval capability, which is an essential requirement for effective TTA.

This lack of adaptiveness of retrieval-based VLM adaptation, including SuS-X [2] and Neural Priming [3], had already been discussed in Section 2.3 of our manuscript, but the citation to [1] was indeed missing. We have now added the citation, which has further strengthened the clarity and completeness of our paper.

To demonstrate the effectiveness of the proposed adaptive retrieval strategy, we conducted a performance comparison with [1]. Although [1] proposed two retrieval strategies, Image-to-Image (I2I) and Text-to-Image (T2I), a direct comparison with the I2I strategy is incompatible with our setting due to its reliance on labeled training images as seed images. Instead, we compared our method against the T2I strategy. Unlike the T2I approach in [1], which retrieves a fixed set of external images, RA-TTA adaptively retrieves a customized set of external images for each test image, achieving superior performance, as shown in Table R1.

Table R1: Performance for the T2I approach of [1] and the proposed RA-TTA. We report the top-1 accuracy (%) on the FGVC aircraft, Stanford cars, and RESISC45 datasets. The best results are in bold. The "Avg." column indicates the average accuracy of 3 datasets.

Q2(W3): Why is optimal transport used, and what is its purpose?

Optimal Transport (OT) is a mathematical framework for calculating the distance between two distributions, typically represented by sampled data points. A key advantage of the OT framework is its flexibility to incorporate a user-defined disparity function to measure the distance between two distributions. When calculating this distance, OT requires a disparity matrix representing pairwise disparities between samples from the two distributions.

In RA-TTA, a novel disparity function called semantic gap is introduced to calculate the semantic disparity between a test image and a retrieved image. Moreover, we adopt two techniques to enhance the reliability and robustness of the adaptation procedure:

• We reuse the augmented images of the test image, previously employed in description-based retrieval. As a result, we have the test image set $\mathcal{A}$ (comprising the test image and its augmented images) and the retrieved image set $\mathcal{R}_{\mathcal{G}_c}$ that is related to the selected features of class $c$.
• We represent the importance of each image within these sets using alignment scores because the importance of individual images might vary. This importance is normalized, allowing the two sets to be interpreted as probability distributions.

Thus, calculating the distance between the test image and the retrieved images is transformed into calculating the distance between two distributions using our novel disparity function, semantic gap. Therefore, we adopt the optimal transport framework in this context.

To enhance the clarity of the paper, we have included the preliminary of optimal transport in Appendix A.1 of the revised manuscript and the discussion of optimal transport with RA-TTA in Appendix A.2 of the revised manuscript.

Dear Reviewer: aU8U, 

Thank you for your review and valuable feedback, which have greatly contributed to enhancing the quality of our paper. This is a kind reminder that we only have one and a half days (about 36 hours) before the author/reviewer discussion period ends. We have submitted a response that addresses your questions, providing clarification on novelty and the use of optimal transport. Additionally, we revised Figure 3. We hope that our response has provided sufficient clarification. Please let us know if you have any further inquiries or require additional information.

Best regards, 
Authors.

Thank you for your quick reply. We would like to clarify and discuss the details of our novelty.

The retrieval-and-adaptation pipeline adopted in [1] has served as a common practice for retrieval-based adaptation approaches [1, 2, 3], and it is therefore acceptable to adopt this pipeline. It should be noted that [1] did not propose the pipeline for the first time; rather, [1] also simply adopted the pipeline following the common practice. The point of the retrieval-based adaptation is to enhance the pipeline. From this perspective, we respectfully contend that our work has clear novelty (far from improving applicability) compared to [1] in two aspects.

Offline vs. Online: In contrast to all the previous studies that enhanced the pipeline in the offline scenario (i.e., before test time) [1, 2, 3], we are the first to study retrieval-based test-time adaptation, aiming to adapt VLMs to a test image provided online, which is effective for distribution shifts in the real world.

Coarse-Grained vs. Fine-Grained: In terms of algorithmic detail, the previous approaches [1, 2, 3] used a coarse-grained retrieval that uses class names, for which the retrieved images can show irrespective features to a test image. On the other hand, we have newly developed a fine-grained approach (i.e., description-based retrieval) based on the fine-grained features of a test image, retrieving more relevant images that have targeted features. For this, an emerging challenge of identifying the features of a test image should be solved. This challenge is unique to us and cannot be addressed by previous approaches. We address this challenge by fully leveraging the cross-modal understanding of VLMs.

We appreciate your concerns and welcome further discussions to refine and improve our work.

---

[1] "Understanding Retrieval-Augmented Task Adaptation for Vision-Language Models.", In ICML, 2024.
[2] "SuS-X- Training-Free Name-Only Transfer of Vision-Language Models.", In ICCV, 2023.
[3] "Neural Priming for Sample-Efficient Adaptation.", In NeurIPS, 2023.

Dear Reviewer: aU8U

Thank you again for your time and efforts in reviewing our paper.

We have thoroughly answered your questions about our novelty compared to [1].

We would appreciate the opportunity to address any additional concerns you may have before the discussion phase ends.

Best wishes,
Authors.

Q2-3: Could this pipeline be simplified, perhaps by using improved prompts to avoid generating imaginary content or by creating visual descriptors to replace the image-to-description selection?

Thank you very much for your insightful ideas. We can simplify the pipeline using your ideas.

• First, given that multiple prompts and queries are required to generate semantically diverse descriptions, we can simplify the description generation procedure with improved prompts that can elicit text descriptions with diverse semantics with a few iterations.
• Second, we can streamline the pipeline by replacing the image-to-description selection with visual descriptors generated by an image generator like Stable Diffusion.

Though these ideas are very interesting topics, it seems to be beyond the scope of this paper. We leave these topics for potential future work.

Q3: Could more visualization results be provided?

Thanks for your suggestion, which improves the elucidation of our paper. We have added the visualization results in Appendix D.7 of the revised manuscript.

Once again, we appreciate the reviewer highlighting valuable discussion points in our study. We hope our response has clarified the contents of the paper, and we are more than willing to engage in further discussion regarding any aspects that remain unclear.

---

[1] "Tip-Adapter: Training-Free CLIP-Adapter for Better Vision-Language Modeling.", In ECCV, 2022
[2] "Prompt, Generate, then Cache: Cascade of Foundation Models Makes Strong Few-Shot Learners.", In CVPR, 2023
[3] "SuS-X: Training-Free Name-Only Transfer of Vision-Language Models.", In ICCV, 2023
[4] "Understanding Retrieval-Augmented Task Adaptation for Vision-Language Models.", In ICML, 2024
[5] "Test-Time Prompt Tuning for Zero-Shot Generalization in Vision-Language Models.", In NeurIPS, 2022

We sincerely thank the reviewer for your insightful comments, which have helped us improve the paper quality significantly. Below, we are more than willing to clarify the reviewer’s concerns one by one.

Q1: More training-free methods should be discussed and compared.

The goal of the proposed RA-TTA is to leverage external knowledge when transferring a VLM at test time. For this purpose, we proposed RA-TTA, an end-to-end pipeline that includes a novel description-based retrieval and adaptation. However, the training-free methods [1, 2] focused on adapting a VLM without backpropagation, which is orthogonal to leveraging external knowledge. Consequently, RA-TTA incorporated external knowledge into adaptation at test time that is overlooked by the training-free methods, achieving performance improvement.

Additionally, to compare the performance of RA-TTA with that of the training-free methods, including Tip-Adapter [1] and CaFo [2], a retrieval approach is needed to enable these training-free methods to leverage external knowledge. We adopted a simple retrieval approach [3, 4] and implemented Tip-Adapter and CaFo on top of it. As shown in Table R1, RA-TTA outperformed the variants because RA-TTA is designed to effectively leverage external knowledge at test time, whereas the training-free variants do not account for using external knowledge.

The discussion regarding training-free methods has been incorporated into Appendix A.1 of the revised manuscript.

Table R1: Performance for the retrieval-based variants of training-free methods and the proposed RA-TTA. We report the top-1 accuracy (%) on the FGVC aircraft, Stanford cars, and RESISC45 datasets. The best results are in bold. The "Avg." column indicates the average accuracy of 3 datasets.

Q2-1: How does its efficiency compare to other methods?

This is a good point. In the context of test-time adaptation, the efficiency of adaptation at test time is indeed a critical factor to consider. In contrast to previous TTA methods, RA-TTA needs to retrieve proper external images for each test image, potentially posing a significant bottleneck. However, the efficiency of RA-TTA is comparable to the representative test-time adaptation method, TPT, as shown in Table R2, where we compared the GPU inference time per sample (i.e., s/sample) for TPT [5] and RA-TTA. This is because FAISS, an advanced and efficient search engine typically used for RAG-based approaches, enables fast nearest-neighbor search. Furthermore, our description-based adaptation does not require backpropagation (i.e., training-free), which also contributes to the efficient inference of RA-TTA.

Thank you again for your insightful comments. We have added this efficiency analysis in Appendix D.4 of the revised manuscript.

Table R2: Inference time for TPT [5] and the proposed RA-TTA. We report the average GPU inference time per sample (s/sample) measured on a single RTX4090 for the FGVC aircraft, Stanford cars, and RESISC45 datasets. The "Avg." column indicates the average inference time of 3 datasets.

Q2-2: GPT usage is not free.

The generation of text descriptions relies on GPT3.5-turbo, which does incur some costs. However, this cost is negligible because our query to GPT is relatively simple, such as "Describe what a 2012 Porsche Panamera Sedan, a type of car, looks like." For instance, generating text descriptions for the Stanford cars dataset that includes 196 classes incurs an expense of only $2.73.

Furthermore, these text descriptions are generated offline, which is conducted only once prior to test time; thus, no additional costs are incurred during inference at test time.

Q3: How do the authors ensure that the retrieved images contain the expected features to augment the test image?

The proposed RA-TTA fully leverages the cross-modal understanding of a VLM to associate an image with the corresponding textual descriptions. Recent VLM-based works [3, 4, 5, 6, 7, 8, 9] also have adopted this cross-modal association. However, we acknowledge that the cross-modal association may occasionally retrieve irrelevant images that lack the expected features. Recent studies have enhanced the association capability of a VLM by transferring the knowledge of LLMs [10, 11, 12] or the diffusion model [13] or introducing novel training strategies [14]. Leveraging the enhanced association capability of these studies, we can improve the relevance of retrieved images.

Additionally, we utilized the prototypes of selected features in the description-based retrieval to enhance the relevance of retrieved images. By incorporating diverse features with different semantics, the prototypes represent the target image more comprehensively, thereby increasing the likelihood of retrieving relevant images.

Q4: The author should consider discussing the difference with these works.

Thanks for your thoughtful feedback, which has helped us better position our paper compared to the related works. Both the related works [15, 16, 17, 18] and the proposed RA-TTA utilize additional knowledge for adaptation. The key distinction lies in the source of knowledge: the related works rely on historical test samples as additional knowledge for adaptation, whereas RA-TTA leverages external samples from an external database. We hypothesize that these two sources of knowledge (i.e., historical and external) are complementary to each other, and integrating both knowledge could be a promising direction for enhancing test-time adaptation. We leave this exploration as an intriguing avenue for future work.

The discussion regarding these related works has been incorporated into Appendix A.1 of the revised manuscript.

We are once again grateful for the reviewer's valuable discussion points in our study. We are eager to engage in further discussion regarding any aspects that remain unclear.

---

[1] "Test-Time Prompt Tuning for Zero-Shot Generalization in Vision-Language Models.", In NeurIPS, 2022
[2] "What Does a Platypus Look Like? Generating Customized Prompts for Zero-shot Image Classification.", In ICCV, 2023
[3] "CLIPScore: A Reference-free Evaluation Metric for Image Captioning.", In EMNLP, 2021
[4] "Scaling Up GANs for Text-to-Image Synthesis.", In CVPR, 2023
[5] "Visual Classification via Description from Large Language Models.", In ICLR, 2023
[6] "CLIP-Dissect: Automatic Description of Neuron Representations in Deep Vision Networks.", In ICLR, 2023
[7] "Label-free Concept Bottleneck Models.", In ICLR, 2023
[8] "What Does a Platypus Look Like? Generating Customized Prompts for Zero-shot Image Classification.", In ICCV, 2023
[9] "Discovering and Mitigating Visual Biases through Keyword Explanation.", In CVPR, 2024
[10] "MATE: Meet At The Embedding - Connecting Images with Long Texts.", In EMNLP findings, 2024
[11] "InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks,.", In CVPR, 2024
[12] "LLM2CLIP: Powerful Language Model Unlocking Richer Visual Representations.", arXiv preprint arXiv:2411.04997, 2024
[13] "Diffusion Feedback Helps CLIP See Better.", arXiv preprint arXiv:2407.20171, 2024
[14] "Long-CLIP: Unlocking the Long-Text Capability of CLIP.", In ECCV, 2024
[15] "Efficient Test-Time Adaptation of Vision-Language Models.", In CVPR, 2024
[16] "Dual Memory Networks: A Versatile Adaptation Approach for Vision-Language Models.", In CVPR, 2024
[17] "SwapPrompt: Test-Time Prompt Adaptation for Vision-Language Models.", In NeurIPS, 2023
[18] "Historical Test-time Prompt Tuning for Vision Foundation Models.", In NeurIPS, 2024

We deeply appreciate the reviewers’ valuable comments and reasonable concerns. We hope that they can be resolved through our clarifications in this rebuttal.

Q1-1: Will directly retrieving from the large-scale datasets like LAION2B help mitigate this issue?

Yes, directly using LAION2B can mitigate the effort required to construct tailored databases for each downstream dataset. However, in our case, constructing tailored databases can be conducted efficiently using SQLite's full text search (FTS). Moreover, constructing tailored databases is a more practical and effective approach than utilizing all images from web-scale datasets, where the characteristics of downstream tasks, such as class names or domains, are known in advance. 

Q1-2: In addition, the computation overhead or inference time compared with TPT should be provided for examining the efficiency of the proposed method.

This is a good point. In the context of test-time adaptation, the efficiency of adaptation at test time is indeed a critical factor to consider. In contrast to previous TTA methods, RA-TTA needs to retrieve proper external images for each test image, potentially posing a significant bottleneck. However, the efficiency of RA-TTA is comparable to the representative test-time adaptation method, TPT, as shown in Table R1, where we compared the GPU inference time per sample (i.e., s/sample) for TPT [1] and RA-TTA. This is because FAISS, an advanced and efficient search engine typically used for RAG-based approaches, enables fast nearest-neighbor search. Furthermore, our description-based adaptation does not require backpropagation (i.e., training-free), which also contributes to the efficient inference of RA-TTA.

We fully agree with the reviewer on the importance of the efficiency analysis, and we have added it to Appendix D.4 of the revised manuscript.

Table R1: Inference time for TPT [1] and the proposed RA-TTA. We report the average GPU inference time per sample (s/sample) measured on a single RTX4090 for the FGVC aircraft, Stanford cars, and RESISC45 datasets. The best results are in bold. The "Avg." column indicates the average inference time of 3 datasets.

Q2: The authors are suggested to discuss how different LLMs might affect description quality and, consequently, the final adaptation performance.

We appreciate the reviewer for providing a valuable discussion point. We conducted additional experiments using text descriptions generated by Claude3-Sonnet for this analysis.

• We also compared the performance of CuPL [2] to investigate the quality of text descriptions because CuPL can achieve better performance with improved descriptions. As shown in the first two rows of Table R2, CuPL with the descriptions of GPT works better than CuPL with those of Claude, implying that the text descriptions generated by GPT are better than those of Claude. Thus, we can conclude that different LLMs can affect description quality even with the same generation prompts.
• Also, the last two rows of Table R2 show that RA-TTA with the descriptions of GPT outperformed RA-TTA with those of Claude. This indicates that the better descriptions facilitate the adaptation of RA-TTA more effectively.
• Notably, RA-TTA with the text descriptions of Claude outperformed the performance of corresponding CuPL (i.e., CuPL+Claude3 vs. RA-TTA+Claude3), demonstrating that RA-TTA effectively leverages external knowledge, even when the quality of the description is not very good.

We have added this analysis to Appendix D.5 of the revised manuscript.

Table R2: Performance when using text descriptions generated by ChatGPT and Claude3. We report the top-1 accuracy (%) on the FGVC aircraft, Stanford cars, and RESISC45 datasets. The employed LLMs are indicated with the method name. The "Avg." column indicates the average accuracy of 3 datasets.

Dear Reviewer: gDjz, 

Thank you for your review and valuable feedback, which have greatly contributed to enhancing the quality of our paper. This is a kind reminder that we only have one and a half days (about 36 hours) before the author/reviewer discussion period ends. We submitted an answer addressing your questions, including the computation overhead and the usage of a different LLM. We hope that our response has provided sufficient clarification. Please let us know if you have any further inquiries or require additional information.

Best regards, 
Authors.

Dear Reviewer: gDjz

Thank you again for your time and efforts in reviewing our paper.
We have answered your questions, provided new experiments as required, and updated our manuscript.

We would appreciate the opportunity to address any additional concerns you may have before the discussion phase ends.

Thank you very much.
Best wishes,
Authors.