Q7: Structure in Figure 2

Figure 2 represents both DT-Training and the Closed-Loop Scaling strategy, is critical for understanding our method. Below are the key clarifications and our plan for improvement:

DT-Training Figure 2 illustrates DT-Training, consisting of the small teacher transfer and dual branch alignment. Please refer to Line 151-224 in our paper. In small teacher transfer, a smaller model serves as the teacher to guide the training of a larger student model. For dual branch alignment, we align the output of clean branch and masked branch where the input image is masked.

Closed-loop Scaling Figure 2 also represents the iterative scaling process where model size, dataset size, and input resolution are incrementally increased. Each iteration uses the previously trained model as the teacher for the next stage. Please refer to Line 228-263 in our paper. Our closed-loop scaling strategy involves several key steps as follows:

1. Teacher Model Setup: We start by training a smaller model, which serves as the teacher model. This model is the foundation for the subsequent scaling process.
2. Scaling Strategy: we expand various factors to train a higher-performing model. The factors that are scaled include model size, training data volume, and image resolution. It is important to note that these factors can be scaled simultaneously or individually. We use our DT-Training to train the expanded model.
3. Iterative Process: In the next iteration, the teacher model (derived from the expanded model) guides the learning of the next, even larger model. This iterative closed-loop approach ensures continuous performance improvement as the model scales up.

We hope this explanation clarifies the meaning of Figure 2. To make this clearer, we will include visual markers or annotations in Figure 2, including:

1. Detailed Captions: We will expand the figure caption to provide a step-by-step explanation of the components in DT-Training and how they integrate with Closed-Loop Scaling.
2. Annotations: Add arrows, labels, and concise explanations directly on the figure to highlight critical processes like knowledge transfer, alignment, and scaling iterations.

We appreciate your suggestion and will make Figure 2 clearer and more intuitive in the revised paper to ensure that the DT-Training and Closed-Loop Scaling strategy are well-understood.

Q6: Model Potential

We appreciate your valuable advice. Below, we provide a more detailed breakdown:

Consistent Performance Improvements Our statement regarding consistent performance improvements is based on experimental results on GTrack Bench. By applying our proposed scaling strategy and DT-Training, we observed:

1. Systematic Gains. With each scaling iteration (e.g., increasing model size, training data, or input resolution), performance improves, which is shown in Table 2, 3, and 4. For instance, transitioning from ViT-B to ViT-L improves AUC by 2.6% on LaSOT, as shown in Table 2, which is higher than navie training (1.6&).
2. Better Optimization. The consistent improvements demonstrate that our approach effectively balances optimization challenges, ensuring that the increased capacity translates into meaningful performance gains.

Potential of the Model The meaning of potential can be summarized as follows:

1. Scalability. By systematically increasing key factors, our method allows models to achieve performance gains that would not be realized with conventional training approaches.
2. Adaptability. The trained models exhibit robust generalization across diverse datasets and tasks, proving their suitability for real-world applications.

Unlocking the Model’s Potential Our work unlocks the potential of the model by addressing specific bottlenecks in traditional scaling approaches. Each method contributes as follows:

1. DT-Training DT-Training introduces a Small Teacher Transfer and dual-branch alignment mechanism to smooth optimization and improve model accuracy.
2. Closed-Loop Scaling Strategy Our scaling strategy integrates three key factors—model size, training data volume, and image resolution—into a systematic framework. This approach prevents over-reliance on a single scaling factor (e.g., model size) and ensures iterable improvements.

We will further clarify the meaning of model potential and explain how our method can improve accucacy in the revised paper.

Q5: Performance Comparison

In our experiments, we have already included comparisons with state-of-the-art trackers, such as ARTrackv2 (CVPR 2024). we will incorporate results from more recent trackers in the revised version of the paper, ensuring a more comprehensive benchmarking against the latest advancements in the field.

The Baseline-B-256-N performance reported in our paper is based on a version of OSTrack trained on four datasets (COCO, GOT-10K, LaSOT, and TrackingNet) rather than the larger dataset. Additionally, we reproduced OSTrack without its CE mechanism. All the implement details are consisent with the origin OSTrack. The resulting model achieved an AUC of 68.4, which is consistent with the performance of OSTrack without CE as reported in its original paper. We will clarify the baseline setup to avoid confusion.

Q4: Novelty of Model

We respond to each point raised regarding the novelty of our proposed method:

The Impact of Specific Increases in Parameter Size While previous works like OSTrack and ODTrack have studied scaling factors, our work introduces a systematic scaling framework for downstream tasks. Unlike prior studies focusing on individual factors, we integrate model size, data volume, and image resolution into a closed-loop scaling strategy. This approach aligns with scaling law principles from upstream tasks and demonstrates their effectiveness in object tracking.

Detailed Insight on Parameter Size Increase Increasing model parameters enhances its ability to learn richer representations, benefiting tasks like object tracking. Our experiments show that scaling from ViT-B to ViT-L improves performance by 1.6% (AUC from 57.2 to 58.8). However, performance gains diminish beyond a certain point, with diminishing returns observed when scaling from ViT-L to ViT-H (only a 0.3 AUC improvement). While larger models generalize better, they also face slower convergence and higher computational costs. We recommend scaling up to ViT-L for the best balance between performance and efficiency.

Small Teacher Transfer Although teacher-student frameworks like Mixformer-V2 and AttTrack are common in lightweight object tracking, our DT-Training introduces a novel use of small teacher transfer within the context of scaling laws for downstream tasks. Unlike traditional distillation, where the teacher is typically larger, our approach employs a smaller teacher to guide a larger model, addressing challenges such as slow convergence and optimization difficulties. This small-teacher transfer accelerates convergence and enhances scalability. Additionally, combining DT-Training with our closed-loop scaling strategy creates a unique framework for iterative model scaling, a novel contribution in downstream tasks.

Clarification of Novel Contributions

1. Scaling Law Application to Downstream Tasks: While scaling laws have been well-explored in upstream tasks like classification and vision-language representation learning, our work adapts these principles to downstream tasks like object tracking.
2. DT-Training: We use a small teacher to guide the training of larger models, addressing optimization issues and enhancing scalability. This differs from conventional knowledge distillation. Additionally, we implement dual-branch alignment to further optimize the model's potential.
3. Closed-Loop Scaling Strategy: We introduce a novel iterative strategy for scaling model size, training data, and image resolution in downstream tasks, providing a systematic approach to model scaling.

We will include more analysis along with plots and tables to clearly illustrate how specific increases in parameter size affect key tracking performance. Besides, we will highlight the distinctions between our method and prior frameworks more explicitly in the final version.

Q3: Motivation of Method

We understand that the novelty and reasoning behind our proposed methods may not have been fully communicated in the abstract and introduction, and we will revise these sections for greater clarity.

Foundations of the Proposed Methods Our methods build on scaling laws from upstream tasks like image classification and vision-language representation learning, where increasing model size, data, and resolution enhances performance. While well-studied upstream, scaling in downstream tasks like visual tracking remains underexplored. To address this, we apply scaling principles using DT-Training and closed-loop scaling to systematically and efficiently improve model performance.

DT-Training A key innovation of our work is DT-Training, which combines small teacher transfer and dual-branch alignment. Small teacher transfer uses a compact model to guide the training of a larger model, enhancing convergence efficiency. Dual-branch alignment further exploits model potential through a masked branch, inspired by methods like MAE.

Closed-Loop Scaling Strategy Our closed-loop scaling strategy is inspired by ideas from progressive training and reinforcement learning, which iteratively expands model size, training data, and image resolution, ensuring smooth scaling and avoiding instability from abrupt resource changes.

Empirical Validation and Comparison with Prior Work Our methods are validated through extensive experiments, showing significant performance improvements on several benchmarks. We will clarify the motivation and evidence in the revised manuscript to enhance clarity and accessibility.

We believe that our approach has broad potential, and we will further clarify the motivation and empirical evidence of our method in the revised manuscript to make it more explicit and accessible.

Q2: Meaning of model

We sincerely appreciate your feedback and acknowledge that the presentation of the scaling law and the explanation of our method can be made clearer, especially in the Introduction.

Clarification of Scaling Law In the revised manuscript, we will explicitly define the scaling law in the Introduction, highlighting that it describes how model performance improves predictably as resources like model size, training data, and image resolution are scaled. While this principle has been well-established in upstream tasks like image classification and vision-language representation learning, it has not been extensively explored in downstream tasks like visual tracking, which is the focus of our work.

Explanation of the Closed-Loop Scaling Strategy We will clarify the closed-loop scaling strategy in the revision, explaining that it involves iteratively scaling model parameters, training data, and image resolution to maximize performance. Each iteration expands one or more factors, with the larger model serving as the "teacher" for the next training round. This process makes model scaling more efficient and enhances downstream task performance. We will provide more details and examples to make the concept clearer.

Improving Overall Clarity We will revise the introduction to clearly present our contributions and methodology, including:

1. Clearer definitions and visualizations of the three scaling factors (model parameters, training data, and image resolution) and their impact on performance.
2. A clearer introduction to the scaling law and its relevance to downstream tasks.
3. A more detailed explanation of the DT-Training method and how it fits into the overall closed-loop scaling strategy.

We are confident that these revisions will make our paper more accessible and easier to understand. We will ensure that the revised manuscript better communicates the motivation and methodology behind our approach.

Thank you for recognizing the efficiency and value of our work. We appreciate your insightful feedback and thoughtful suggestions. All other reviewers have acknowledged the importance of our work, confirming that our motivation is clear, sufficient, and well-founded. They also agree that the relationship between the three key points we explored and the scaling law is well-established and reasonable. Furthermore, they highlighted the strong generalizability of our approach to other fields. In addition, they also recognize the novelty, reasonableness, and effectiveness of our proposed methodology. Besides, they think 'The paper is well-organized and clearly written, making it easy to follow the methodology and results'. We sincerely hope you will reconsider your evaluation, and we would be truly grateful for your support.

Q1: Motivation of Scaling Law

We thank you for your valuable feedback regarding the motivation and the relationship between the scaling law and the three points in our work. We address your concerns as follows:

Motivation for Scaling Law in Downstream Tasks Scaling laws are a fundamental concept in machine learning, demonstrating predictable relationships between model performance and the scale of resources such as model parameters, training data, and input resolution. While primarily studied in upstream tasks such as image classification and vision-language representation learning, these laws show that performance improves with resource scaling until it reaches a saturation point. This insight helps optimize resource allocation and maximize model performance.

The motivation for applying scaling laws to downstream tasks is to create a systematic framework for improving model performance as resources increase. Inspired by successes in upstream tasks, we adapt these principles to downstream tasks. Unlike arbitrary scaling methods that can lead to inefficiencies, the scaling law guides how to scale model size, data, and resolution to achieve optimal performance, avoiding diminishing returns and inefficiencies. Our goal is to provide clear guidelines for enhancing model performance through resource scaling.

Relationship Between the Three Points and the Scaling Law The three points (model parameters, training data, and image resolution) are key components of the scaling law, which has been explored in upstream tasks like image classification and vision-language representation learning, where scaling each factor improves performance. In our work, we apply these same principles to downstream tasks, explaining how each factor contributes to scaling and enhancing performance in these tasks:

1. Model Parameters Scaling model size (i.e., increasing the number of parameters) is key in scaling laws for upstream tasks, as larger models capture more complex patterns and generally perform better. We apply this to downstream tasks by scaling model parameters, enhancing the model's ability to learn from data, which leads to improved performance.

2. Training Data In upstream tasks, scaling training data improves generalization by providing more diverse examples for learning robust features. We apply this to downstream tasks by increasing the training data, allowing the model to learn better representations and generalize to new scenarios.

3. Image Resolution In upstream tasks, higher image resolution helps models capture finer details, improving performance. Similarly, in downstream tasks, increasing resolution provides more detailed visual information, enhancing model performance.

The three factors—model parameters, training data, and image resolution, are central to the scaling law, widely studied in upstream tasks like image classification and vision-language representation learning. We apply these principles to downstream tasks, where their effects have been less explored. We will revise the abstract to clarify their relevance to our work.

Dear Reviewer:

As the discussion period is progressing, we would greatly value the opportunity to engage with you regarding your feedback on our submission. Your insights and suggestions are immensely important to us, and we are eager to address any remaining questions or concerns you may have.

We are committed to providing timely and detailed responses to ensure that all aspects of our work are clarified. If our responses have satisfactorily addressed your concerns, we kindly request that you consider reflecting this in your evaluation and possibly revising your score.

Thank you again for your time and effort, and we look forward to discussing with you.

Best regards

Q4: LoRAT

We thank you for pointing out the comparison with prior work, such as LoRAT. We address your concerns as follows:

Key Differences Between Our Work and LoRAT While LoRAT and our method both leverage larger models for visual tracking, the approaches differ fundamentally in their methodologies and goals. LoRAT focuses on designing a larger model and incorporates a more powerful backbone (DINOv2). As such, the performance of LoRAT is inherently higher due to these architectural enhancements. In contrast, our work focuses on training strategies and scalability, and we achieve performance comparable to LoRAT using baseline models with far less architectural optimization. This demonstrates the effectiveness of our DT-Training strategy and its potential to significantly improve model performance without requiring extensive model redesign.

Flexibility of Our Training Strategy A key advantage of our training strategy is that it is model-agnostic, meaning that it can readily incorporate newer and more advanced model architectures as they become available. By applying our scalable training framework to state-of-the-art models, we believe performance can be further improved beyond the results presented in this work. This adaptability positions our approach as a general-purpose solution that can evolve alongside advancements in the field.

Performance on Benchmarks While it is true that LoRAT achieves higher success rates on specific benchmarks such as LaSOT, our model demonstrates better performance on other benchmarks, such as TNL2K, where it outperforms LoRAT (66.3 vs 62.3). This highlights the generalizability of our approach across diverse datasets with varying levels of difficulty and data distributions. By excelling in benchmarks like TNL2K, we demonstrate the robustness of our method in handling real-world challenges.

We will include additional evaluations against LoRAT on more benchmarks in future work to provide a comprehensive analysis. We also aim to explore ways to integrate the strengths of LoRAT into our framework to further enhance performance.

Q3: VastTrack

Our key contribution lies in the proposed scaling framework and DT-Training, which systematically improve the scalability and performance of visual tracking models. The use of a large-scale and diverse benchmark is intended to validate the effectiveness and generalizability of our methods in a comprehensive and realistic setting. By combining multiple datasets, our benchmark reflects a broader range of challenges encountered in real-world scenarios, making it particularly suited for evaluating scaling strategies.

We chose to aggregate a diverse set of benchmarks, rather than relying on a single dataset, to ensure wide applicability and to avoid biases associated with any one dataset. We have conducted experiments on VastTrack to verify the effectiveness of our DT-Training and closed-loop scaling strategy. The results are shown as following. Our model outperforms previous models in a large extent.

We will incorporate this feedback into future versions of our benchmark, further enhancing its coverage and diversity.

Q2: Efficiency

We understand that visual tracking requires high efficiency, particularly in real-time applications. It is important to note that our method introduce no additional computational overhead during inference. The inference speed of our model is consistent with the baseline models, OSTrack. For instance, our model Ours-B-256-M achieves 93 fps on an NVIDIA 2080 Ti GPU, which is ths same as the baseline OSTrack while delivering superior performance in terms of accuracy.

Moreover, we have demonstrated in Table 6 that our model maintains strong performance even after compression. By compressing the trained model, it is possible to achieve a balance between maintaining accuracy and meeting stringent real-time performance requirements. This further highlights the practicality of our framework for resource-constrained environments.

In conclusion, our method introduces no additional computational overhead during inference and has been validated to adapt well to lightweight model compression. This ensures it remains highly suitable for deployment in real-world scenarios requiring both accuracy and efficiency. We will add the detailed discussion of computational cost in the revised manuscript.

Thank you for acknowledging the efficiency and novelty of our work. We deeply appreciate your insightful feedback and constructive comments. Based on your suggestions, we have carefully reviewe and revised our manuscript to further enhance its clarity and presentation. We sincerely hope you will reconsider your evaluation, and we would be truly grateful for your support.

Q1: Choice of Task

We chose visual tracking as the primary task for several key reasons:

Task Complexity and Challenge Visual tracking presents a unique set of challenges that make it an interesting and demanding test case for our scaling strategy. Unlike object detection, which focuses on localizing objects in images, tracking involves associating objects across consecutive frames, often under more challenging conditions such as occlusions, changes in scale, and fast motion. This complexity makes tracking a valuable task for demonstrating the benefits of scaling, as the models must maintain performance across both spatial and temporal dimensions.

Class-agnostic Nature of Tracking One key difference between object detection and tracking is that detection requires the model to classify objects into predefined categories, which can lead to challenges in scaling. Specifically, different datasets may have varying class definitions or inconsistent category labels, which can cause semantic conflicts when attempting to scale or transfer models across tasks with different class sets. In contrast, visual tracking is class-agnostic, meaning that it doesn’t require specific object categories to function. This lack of semantic classification allows tracking models to scale more smoothly across different datasets and environments without the risk of semantic conflicts, making it a more robust task for studying scaling laws.

While object detection is indeed a foundational task, we believe that visual tracking offers a complementary and more suitable setting for exploring scaling laws, especially in cases where class-specific training is not a key concern. We will clarify this rationale in the revised manuscript to ensure that readers understand why visual tracking was chosen as the primary target for this study.

we will highlight these points more clearly in the revised version of the paper.

Dear Reviewer:

As the discussion period is progressing, we would greatly value the opportunity to engage with you regarding your feedback on our submission. Your insights and suggestions are immensely important to us, and we are eager to address any remaining questions or concerns you may have.

We are committed to providing timely and detailed responses to ensure that all aspects of our work are clarified. If our responses have satisfactorily addressed your concerns, we kindly request that you consider reflecting this in your evaluation and possibly revising your score.

Thank you again for your time and effort, and we look forward to discussing with you.

Best regards

Q4: Convergence Speed

The proposed training method can speed up convergence. Specifically, our DT-Training framework leverages a small teacher model to guide the training of a larger model, which facilitates faster convergence by providing smoother optimization pathways. Furthermore, a model trained with DT-Training achieves a 68.6 AUC after only 80 epochs, surpassing the baseline model's performance even after 300 epochs (68.4). This highlights the efficiency of our approach. We will clarify the acceleration benefits of DT-Training in the revised manuscript.

Q3: Upper Limit

The scaling law for downstream tasks is influenced by several factors, including the complexity of the task, the diversity and size of the training data, and the capacity of the model. As shown in Figure 1, the scaling law in downstream tasks does have an upper bound when using conventional training methods. Once the model size and dataset volume reach a certain level, the marginal benefits of scaling begin to diminish, and the model’s performance approaches a plateau.

By leveraging a smaller model as a teacher, DT-Training accelerates convergence and makes the training process smoother. Additionally, our Dual-branch alignment mechanism further unlocks the model’s potential, effectively enhancing its performance. These innovations enable our DT-Training to surpass the traditional upper bound of performance achieved with conventional training methods. we will add the discussion about the upper limit of scaling law in the modified manuscript.

Q2: Training Details

In our closed-loops scaling experiments, we begin with a OSTrack-256 pretrained on four datasets. Firstly, we expand the training dataset and train a student with a ViT-B backbone, referred to as as Ours-B-256-M. Then we utilize the Ours-B-256-M as tracher model, and scale up the model size to ViT-L, named as Ours-L-256-M. Finally, we increase the input resolution to 384, which is named as Ours-L-384-M.

Our scaling strategy is iterative and designed to be flexible, with no strict rules or fixed standards about the order or manner of scaling. At any training stage, we can choose to scale one or more of the following three key aspects: model size, training data volume, and input resolution. The flexibility of our approach allows users to select which aspect(s) to scale at each stage. For example, one iteration might focus solely on increasing model size, while another might scale both the training data and input resolution. The process does not require a predetermined sequence for scaling these factors. We will add more details about the scaling process in the revised manuscript.

Thanks for your insightful feedback and support. We sincerely appreciate your valuable comments and your recognition of the novelty and effectiveness of our work. We have modified and updated the PDF according to your advice.

Q1: Computational Cost

We appreciate your concern about the computational cost and resource requirements of our proposed methods. We would like to emphasize that our method does not introduce additional computational overhead during testing. The inference speed of our model is consistent with the baseline models, OSTrack. For instance, our model Ours-B-256-M achieves 93 fps on an NVIDIA 2080 Ti GPU, which is ths same as the baseline OSTrack while delivering superior performance in terms of accuracy. Moreover, we have demonstrated in Table 6 that our model maintains strong performance even after compression, highlighting its potential for efficient deployment in real-world scenarios. These results illustrate that our approach not only ensures computational efficiency but also provides robust performance, making it highly suitable for practical applications.

We will include a detailed discussion of computational cost in the revised manuscript.

This means that your training is to randomly increase the model size, data size, etc., but the author does not seem to verify that this training method can always bring performance improvements.

Dear Reviewer:

As the discussion period is progressing, we would greatly value the opportunity to engage with you regarding your feedback on our submission. Your insights and suggestions are immensely important to us, and we are eager to address any remaining questions or concerns you may have.

We are committed to providing timely and detailed responses to ensure that all aspects of our work are clarified. If our responses have satisfactorily addressed your concerns, we kindly request that you consider reflecting this in your evaluation and possibly revising your score.

Thank you again for your time and effort, and we look forward to discussing with you.

Best regards

The author's rebuttal solves my doubts, and I keep the rating unchanged.

Q2: Hyperparameter

Our DT-Training method is designed to be simple and does not involve many hyperparameters. The hyperparameters are the mask ratio $p$ and regularization parameters $\lambda_{transfer}$ and $\lambda_{align}$, both of which have been thoroughly studied in our paper. As shown in Figure 3 in our paper, DT-Training demonstrates robust performance across a wide range of hyperparameter values, indicating that it is not particularly sensitive to these settings.

We will emphasize the robustness to hyperparameters of our DT-Training in the later version.

Q6: Details of Closed-loop Scaling Up

Our closed-loop scaling strategy involves several key steps (Line 228-263) as follows:

1. Teacher Model Setup: We start by training a smaller model, which serves as the teacher model. This model is the foundation for the subsequent scaling process.
2. Scaling Strategy: we expand various factors to train a higher-performing model. The factors that are scaled include model size, training data volume, and image resolution. It is important to note that these factors can be scaled simultaneously or individually. We use our DT-Training to train the expanded model.
3. Iterative Process: In the next iteration, the teacher model (derived from the expanded model) guides the learning of the next, even larger model. This iterative closed-loop approach ensures continuous performance improvement as the model scales up.

We appreciate your suggestion, and we will include more detailed explanations in the revised manuscript to make it clearer.

Q5: Small Teacher Transfer

Firstly, directly training a very large model can lead to optimization difficulties. As the model scale increases, performance improvements often do not correspond to the increase in size. When training a large model, we lack a suitable, well-performing larger model as a teacher. Thus, we choose to use a smaller model as the teacher in our approach.

Secondly, this is also a key innovation of our DT-Training framework, small teacher transfer. Although the smaller model has less capacity than the larger model, both models learn similar knowledge, allowing the smaller model to transfer its knowledge to the larger model. This process accelerates convergence, improves training efficiency, and results in smoother optimization of the larger model. As demonstrated in Table 5, the guidance from the smaller model not only aids the larger model in training but also leads to improved performance of the larger model.

We will clarify the motivation of our small teacher transfer in the revised manuscript.

Q4: Limitation

While GTrack Bench is a large-scale benchmark that is three times the size of previous datasets, we acknowledge that the dataset size still needs to be further expanded to comprehensively evaluate the advantages of our method in scaling up, which is the limitation of our work. GTrack Bench provides a solid foundation, but testing on even larger-scale datasets would allow us to better assess the scalability and robustness of our approach under more varied conditions.

In future work, we plan to explore more efficient scaling techniques to further enhance the model's performance. We appreciate your valuable feedback, and we will add a more detailed discussion of the limitations and directions for future work in the revised version of our paper.

Q3: Real-world Scenarios

Our proposed GTrack Bench is a large-scale benchmark that combines multiple diverse datasets. This diversity allows us to demonstrate the generalization ability of our method across various conditions, and we believe it reflects the performance of our approach in real-world scenarios to a significant extent. We will add additional real-world case studies in the later version to further validate the effectiveness of our methods.

We sincerely appreciate your recognition of the effectiveness and contributions of our work, along with your thoughtful feedback and suggestions. We have modified and updated the PDF according to your advice. We would be truly grateful if you could reconsider our work and offer your support!

Q1: Computational Cost

We appreciate your concern about the computational cost and resource requirements of our proposed methods. We would like to emphasize that our method does not introduce additional computational overhead during testing. The inference speed of our model is consistent with the baseline models, OSTrack. For instance, our model Ours-B-256-M achieves 93 fps on an NVIDIA 2080 Ti GPU, which is ths same as the baseline OSTrack while delivering superior performance in terms of accuracy. Moreover, we have demonstrated in Table 6 that our model maintains strong performance even after compression, highlighting its potential for efficient deployment in real-world scenarios. These results illustrate that our approach not only ensures computational efficiency but also provides robust performance, making it highly suitable for practical applications.

We will include a detailed discussion of computational cost in the revised manuscript.

Dear Reviewer:

As the discussion period is progressing, we would greatly value the opportunity to engage with you regarding your feedback on our submission. Your insights and suggestions are immensely important to us, and we are eager to address any remaining questions or concerns you may have.

We are committed to providing timely and detailed responses to ensure that all aspects of our work are clarified. If our responses have satisfactorily addressed your concerns, we kindly request that you consider reflecting this in your evaluation and possibly revising your score.

Thank you again for your time and effort, and we look forward to discussing with you.

Best regards