We greatly appreciate your insightful comments and suggestions, as they have been helpful in refining and enhancing our work. We have thoroughly reviewed all of your points and have addressed your concerns as outlined below:

Q1: Clarify Method with More Diagrams and Intuition.

Thank you very much for your feedback. We have revised both the Introduction and Figure 1 accordingly.

• Figure 1 has been updated with a new subfigure (b) to visually demonstrate how the agent loses its ability to navigate to previously learned goals after learning new tasks, thus illustrating the challenge of catastrophic forgetting in continual object navigation.
• We revised the figure caption to clarify the key concepts. The new caption reads: Continual Object Navigation: The agent must continually learn from new data while retaining its ability to navigate to previously seen object goals. Subfigure (b) shows representative trajectories from a baseline model and C-Nav when evaluated on an old task after learning a new one.
• Additionally, we have added an explanation of Figure 1 in the Introduction, emphasizing the underlying challenges of continual learning. Specifically: As the agent is exposed to new tasks over time, distributional shifts in sensory inputs and action patterns introduce biases into the model components, leading to representation drift and policy degradation. This results in catastrophic forgetting of previously acquired knowledge, as illustrated in Figure 1. The agent often becomes capable of semantic exploration only for objects in the current task, losing its ability to navigate to goals learned earlier.

The algorithm pseudocode and trajectory visualizations in the supplementary material (already included in the original submission) provide a more detailed and intuitive understanding of our method.

These updates make the motivation and challenges of continual object navigation more visually accessible and help contextualize our proposed solution early in the paper.

Q2: Add Visualizations or Qualitative Examples

Thank you for the suggestion. We provide qualitative trajectory analysis in Section C.3 (Case Study) of the supplementary material (already included in the original submission), including:

• Four TopDown visualizations (Figure 7 & 8), illustrating agent behavior across stages;
• Four video clips, showing first-person navigation trajectories of both the baseline model and C-Nav when evaluated on old tasks after multi-stage training.

During evaluation, we focused on two episodes: locating a toilet (stage 1) and a bed (stage 2). The visualizations in Figure 7 and Figure 8 clearly show that the baseline agent exhibits catastrophic forgetting. After multi-stage training, it consistently navigates toward the sofa (the target in the final stage), completely forgetting the navigation strategy for older objects. In contrast, the C-Nav model successfully navigates to both the toilet and bed targets, demonstrating superior long-term task retention.

To provide a more intuitive comparison, we plan to move these trajectory visualizations to the main experimental results section in future versions of the paper. This will better highlight the challenges in continual object navigation and allow a clearer comparison of model performance.

Q3: Include More Recent Rehearsal Baselines

Thank you very much for your valuable feedback. In our experimental section, we have compared our method against several representative continual learning baselines, including Data Replay, LoRA, LwF, and Model Merge. Among them, Data Replay serves as a strong baseline due to its effectiveness in mitigating catastrophic forgetting. Our results demonstrate that the proposed C-Nav outperforms these methods not only in terms of navigation performance, but also offers clear advantages in storage efficiency and data privacy, thanks to our adaptive experience selection and dual-path forgetting mitigation strategy.

We have also added a comparison with ZSCL[1], a recent method that imposes constraints in both parameter and feature space to support continual learning. The results are summarized below:

Although ZSCL performs reasonably well in earlier stages, its performance drops considerably in later stages. For example, in Stage 4, its SR drops to 24.6 compared to 46.5 achieved by C-Nav. This highlights a key limitation of existing continual learning frameworks, which are often designed for static, perception-focused tasks like image classification or image-text matching. These tasks do not involve long-horizon exploration, target identification, active perception, or obstacle avoidance, which are essential components of continual object navigation. As a result, such methods are difficult to directly apply to this setting. In contrast, our task explicitly incorporates these embodied challenges, and C-Nav is designed to handle these embodied challenges, enabling effective continual learning in complex, dynamic settings.

We are also extending our comparisons to include additional recent methods, and the results will be reported in the final version of the paper.

We hope this response sufficiently addresses your concerns.

[1]Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models

Q4: Evaluate Category Identity Preservation Across Stages

Yes, we have examined how the model handles repeated object categories that appear in different stages. As discussed in Section C.3 of the supplementary material, we selected a model that had been trained across all four stages and visualized its navigation trajectories in tasks from Stage 1 (toilet) and Stage 2 (bed). This analysis helps reveal how well the model retains knowledge of earlier tasks after continual fine-tuning. Our TopDown visualizations (Figures 7 and 8) and first-person video comparisons show that the finetuning baseline exhibits clear signs of forgetting, often navigating toward the most recently trained object goal. In contrast, C-Nav retains prior knowledge and successfully navigates to object goals from earlier stages, demonstrating greater stability and resistance to forgetting throughout the continual learning process.

To provide quantitative evidence, we further evaluated each model on old tasks across all stages, as shown in the following table:

Performance on Old Tasks at Each Stage (HM3D):

These results quantitatively support the superior forgetting resistance of C-Nav compared to all tested baselines.

Q5: Lack of qualitative analysis and visualization

As mentioned in Answer1 and Answer2, we have incorporated qualitative visualizations and video comparisons to highlight what the model is learning and how forgetting manifests in practice. Specifically:

• Figure 7 and Figure 8 in the supplementary materials provide TopDown trajectory visualizations, contrasting the finetuning baseline and our C-Nav model.
• Four accompanying video clips show the first-person navigation behavior of both models in various environments, targeting previously seen objects.
• These comparisons further highlight C-Nav's ability to preserve semantic understanding across tasks, even after multi-stage training.

We hope these additions help clarify our approach and provide a more comprehensive understanding of the model's capabilities

Q6: Improve figure explanation

As also addressed in Answer1 and Answer2, we have revised Figure 1, the Introduction, and the case study in the Experimental Results section to improve clarity and better highlight the challenges of continual object navigation.

Q7: Analyze stage-order sensitivity

We appreciate the reviewer’s insightful suggestion regarding task order sensitivity. To assess the robustness of our method to stage-order permutations, we conducted ablation studies with three different task orderings. The object categories were regrouped across stages.

We summarize SR and SPL across all stages in the following table(MP3D):

These findings suggest that C-Nav is robust to different stage sequences, making it well-suited for real-world continual learning scenarios with dynamic task orders.

Thank you again for your valuable suggestions. We will incorporate the relevant additions and improvements in the final version. We hope our response adequately addresses your concerns and would greatly appreciate your consideration for a higher score. Should you have any further questions, we are happy to respond.

Dear reviewer 3LhB:

We sincerely appreciate your insightful and thoughtful comments! Following your suggestions, we have provided detailed explanations, added visualizations, and included additional experimental results. Specifically, we moved the visualizations from the supplementary materials into the main text, added explanations and intuition for Figure 1, and included results on the impact of learning order as well as a new baseline.

As it is approaching the end of the discussion period (Aug 8 AoE), we sincerely hope you could look through our response and have a further comment at your convenience if you have any questions about the paper. We will do our best to address the issues of the reviewer.

Best wishes,

Submission 22093 Authors.

Dear Reviewer 3LhB,

I hope you are doing well. As the discussion period will close soon (Aug 8 AoE), I would like to kindly follow up on our rebuttal for Submission 22093.

We have carefully addressed each of your comments with additional explanations, new visualizations, and extended experimental results, as detailed in our response. Your feedback has been invaluable for improving our work, and we would greatly appreciate it if you could take a moment to review our updates and share any further thoughts before the discussion period ends.

Thank you again for your time and constructive feedback.

Best regards,

Submission 22093 Authors.

Thank you for your detailed comments and suggestions on our paper. Your feedback has been invaluable in helping us improve and refine our research. We greatly appreciate your efforts and have provided our responses to each of your points below:

Q1: Feature Consistency Loss Concern

Thank you for the insightful comment. Our feature consistency loss is designed to learn a shared feature space that remains effective across both past and new tasks. Simply freezing the entire multimodal encoder may preserve old knowledge but limits the model’s adaptability, especially when the original feature space does not sufficiently capture new task-specific semantics. Instead, we selectively fine-tune the modality-specific projectors within the encoder while applying consistency regularization to softly align the old and new feature spaces. This design allows the representations to stay connected, yet flexible enough to adapt to new tasks. It strikes a balance between preserving past knowledge and enabling plasticity for future learning. To validate this design choice, we conducted an ablation study on the MP3D Continual ObjectNav dataset by freezing the entire multimodal encoder. The results showed a performance drop across learning stages, confirming that full freezing harms generalization and underscoring the necessity of our consistency loss.

Q2: Adaptive Experience Feature Selection

Thank you for your constructive feedback. To validate the effectiveness of our proposed Adaptive Experience Feature Selection mechanism, we conducted an ablation study on the MP3D Continual ObjectNav dataset with two sets of experiments:

• Experiment 1: We compared our method against alternative experience selection strategies, including clustering-based sampling and uniform sampling. For a fair comparison, in the clustering-based approach, we set the number of cluster centers to half the trajectory length and selected frames closest to each cluster center in the feature space as key decision frames.

• Experiment 2: We evaluated the robustness of our method when selecting fewer key frames, by reducing the selection ratio.

Results from Experiment 1 demonstrate that our method outperforms the alternatives. From a data distribution perspective, frames that exhibit significant semantic shifts—such as entering new spaces, approaching target objects, or reaching decision points—are typically rare and can be viewed as outliers in the feature space. Our method explicitly aims to identify and preserve such outlier frames, which are more informative for navigation decisions. In contrast, clustering-based or uniform sampling methods tend to select more “normal” frames that lack critical decision relevance, leading to suboptimal performance.

Results from Experiment 2 show that our method maintains strong performance even when selecting a smaller proportion of key frames. Notably, our approach achieves comparable or better results while using only 1/5 of the memory compared to clustering-based or uniform sampling baselines, highlighting its efficiency and scalability.

Thank you very much for your valuable suggestions, which have been incredibly helpful in improving our work. We will incorporate the above results into the revised version of the paper to further demonstrate the effectiveness of the proposed method. We hope our response addresses your concerns. If you have any additional questions or feedback, we sincerely welcome them and will be glad to provide further clarification. Thank you again for your time and thoughtful review.

We greatly appreciate your insightful comments and suggestions, as they have been helpful in refining and enhancing our work. We have thoroughly reviewed all of your points and have addressed your concerns as outlined below:

Q1: Clarification on the Need for Continual Learning with Limited Object Categories

Thank you for raising this important question. While HM3D contains only 6 categories, we also evaluate on MP3D, which offers 21 diverse and semantically rich categories (e.g., gym_equipment, fireplace), providing a more realistic benchmark.

Our goal is not merely to increase category diversity, but to address fundamental challenges in continual learning for embodied object navigation—a setting largely overlooked in prior work. Existing methods are primarily designed for static perception tasks like image classification or retrieval. In contrast, our task involves long-horizon exploration, multimodal input, obstacle avoidance, and active perception, introducing unique and complex forgetting patterns.

Even under controlled staged settings, we observe catastrophic forgetting, semantic confusion, and task interference, which significantly degrade navigation performance and are not effectively addressed by current methods.

Our framework offers a scalable and diagnosable platform to study these challenges. We hope this work provides valuable insights toward building embodied agents that are adaptive, efficient, and capable of long-term deployment in dynamic environments.

Q2: Ablation Studies

In response, we now provide detailed ablation studies on key hyperparameters.

(1) Loss Weighting Coefficients (MP3D)

As shown in the table, we chose λ = 5 as the optimal weight for both losses, balancing the retention of old knowledge and the ability to learn new tasks effectively.

(2) LOF Frame Retention Ratio (MP3D)

As shown in the table, the model performance decreases as the retention ratio drops. However, the overall decline is relatively mild. Notably, even when retaining only 10% of the key frames, our method still demonstrates competitive performance, highlighting the robustness of our approach in selecting relevant experiences for continual learning.

Q3: Include More Recent Continual Learning Methods

Thank you very much for your valuable feedback. In our experiments, we compared C-Nav with several representative continual learning baselines, including Data Replay, LoRA, LwF, and Model Merge. Among these, Data Replay serves as a strong baseline due to its ability to mitigate catastrophic forgetting.

Our results demonstrate that the proposed C-Nav outperforms these methods not only in terms of navigation performance, but also offers clear advantages in storage efficiency and data privacy, thanks to our adaptive experience selection and dual-path forgetting mitigation strategy.

We have also added a comparison with ZSCL[1], a recent method that imposes constraints in both parameter and feature space to support continual learning. The results are summarized below:

While ZSCL performs reasonably well in early stages, its performance declines significantly in later stages. For example, in Stage 4, its SR drops to 24.6, compared to 46.5 achieved by C-Nav. This reflects a broader limitation of many continual learning methods that are primarily developed for static, perception-focused tasks such as image classification or image-text matching. These tasks do not involve long-horizon exploration, active perception, or obstacle avoidance, which are essential for embodied navigation.

In contrast, our task setting integrates these embodied challenges, and C-Nav is specifically designed to address them, enabling more effective continual learning in complex, dynamic environments.

We are further expanding our comparisons to include additional recent methods, and the updated results will be presented in the final version of the paper.

We hope this response sufficiently addresses your concerns.

[1]Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models

Q4: Include Qualitative Trajectory Analysis

Thank you for the suggestion. We provide qualitative trajectory analysis in Section C.3 (Case Study) of the supplementary material, including:

• Four TopDown visualizations (Figure 7 & 8), illustrating agent behavior across stages;
• Four video clips, showing first-person navigation trajectories of both the baseline model and C-Nav when evaluated on old tasks after multi-stage training.

During evaluation, we focused on two episodes: locating a toilet (stage 1) and a bed (stage 2). The visualizations in Figure 7 and Figure 8 clearly show that the baseline agent exhibits catastrophic forgetting. After multi-stage training, it consistently navigates toward the sofa (the target in the final stage), completely forgetting the navigation strategy for older objects. In contrast, the C-Nav model successfully navigates to both the toilet and bed targets, demonstrating superior long-term task retention. To provide a more intuitive comparison, we plan to move these trajectory visualizations to the main experimental results section in future versions of the paper. This will better highlight the challenges in continual object navigation and allow a clearer comparison of model performance.

Q5: Report Old vs. New Task Performance per Stage

Following your suggestion, we report the results on the HM3D dataset by separately evaluating performance on old and new tasks.

New Task Performance:

Old Task Performance:

These results clearly demonstrate that:

• Finetuning suffers from severe catastrophic forgetting, with old-task performance dropping drastically as new tasks are introduced.
• C-Nav maintains stable performance across both old and new tasks, effectively balancing stability and plasticity. In the final stage, it achieves a 10% higher success rate on old tasks compared to the best-performing baseline.

We will include this breakdown in Appendix C.4, along with further analysis to highlight the trade-offs involved in continual object navigation.

Q6: Justify L2 Feature Constraint for Representation Stability

We adopt an L2-based feature distillation loss to enforce consistency in the encoder's representations across stages. This simple yet effective constraint mitigates representation drift. The design is inspired by prior work on feature-level knowledge distillation [2], where L2 regularization helps preserve essential semantic structures from earlier tasks. Ablation results (Table 4) show that removing this constraint leads to substantial performance degradation (−22.15 SR on HM3D and −15.92 on MP3D). This confirms that stabilizing the encoder’s feature space is crucial for long-term retention, and supports the effectiveness of our L2 constraint in mitigating forgetting.

[2] CVPR 2019：Relational knowledge distillation.

Q7: C-Nav Shows Performance Decline in Stage 4 While Several Baselines Recover

As shown in Answer 2, all baseline methods experience significant performance drops on old tasks, indicating severe forgetting. In contrast, C-Nav effectively mitigates this, achieving a 10% higher success rate on old tasks than the best baseline in the final stage.

The recovery in some baselines during Stage 4 is likely due to two factors:

• Task Difficulty: Stage 4 introduces "sofa" as the target, which is easier than "TV" in Stage 3. The base model's new-task success rate jumps from 29.54% in Stage 3 to 64.63% in Stage 4, suggesting overfitting to the easier task, rather than true continual learning.
• Semantic Similarity: The target "sofa" is visually similar to previously learned objects like "bed" and "chair," leading to potential confusion and misidentification, which could falsely improve performance on old tasks.

In contrast, C-Nav prioritizes consistent performance across stages, focusing on long-term knowledge retention over short-term gains. This slight decline in Stage 4 reflects its cautious adaptation strategy, designed to prevent overfitting and ensure stability across extended task sequences.

Thank you for your valuable feedback. We hope our responses address your concerns and are happy to clarify further if needed.

Dear Reviewer 23E4,

As the discussion phase is closing soon (Aug 8 AoE), we sincerely hope you may have a chance to revisit our rebuttal and share any further comments or questions. Your feedback has been invaluable, and we would greatly appreciate any final suggestions.

We would also like to briefly reiterate the main contributions of our paper, highlighting how our approach differs from prior work:

• New Benchmark: We introduce the first benchmark for continual object navigation, advancing beyond static tasks like image classification or vision-language retrieval to an embodied setting that demands long-horizon exploration, active perception, obstacle avoidance, spatial reasoning, and multi-modal integration. These challenges introduce complex forgetting dynamics not seen in traditional benchmarks.

• C-Nav Framework: We propose a continual object navigation framework that combines feature distillation to maintain representation consistency and feature replay to ensure decision-making consistency throughout training. This approach addresses both encoder drift and policy degradation across sequential task learning.

• Adaptive Experience Selection: We develop a memory-efficient sampling strategy that leverages Local Outlier Factor (LOF) in the feature space to identify semantically meaningful frames. This method retains key experiences while significantly reducing storage requirements and mitigating privacy risks related to raw trajectory retention.

• Strong Results: We demonstrate that C-Nav consistently outperforms strong baselines across four distinct navigation architectures, achieving superior memory efficiency and retention of prior knowledge compared to full replay methods.

In response to your concerns, we provide the following point-by-point clarifications:

• Visualization: Our supplementary material already contains qualitative trajectory visualizations and video clips showing that baseline models suffer catastrophic forgetting, while C-Nav preserves navigation abilities across multiple stages. We have now moved these visualizations into the main experimental results section and added a subfigure in the introduction to better illustrate the continual object navigation task and the forgetting challenge.

• More Results: We now include additional ablation studies on loss weighting coefficients and LOF frame retention ratios, further demonstrating the robustness of our method even with minimal frame retention. Additionally, we have added new results comparing old vs. new task performance across stages, as well as a comparison with the newly introduced baseline method, ZWCL.

• Clarifications: We have clarified the necessity of our benchmark for embodied object navigation tasks and discussed the performance comparison between C-Nav and baselines in the context of multi-stage continual learning.

We are happy to provide further clarification if needed. Thank you once again for your time and constructive review.

Best regards,

Submission 22093 Authors

Thank you for your comments. We would like to clarify the following point:

Q1: Difference between Continual-ObjectNav and MultiON tasks

While MultiON also involves multiple objects, it is fundamentally different from our benchmark:

• MultiON: Uses a single-stage training setup, where all object categories are available during the single training phase. At test time, the agent sequentially searches for multiple targets within the same environment, evaluating its long-term spatial memory. Completing Subtask 1 can facilitate the search for Subtask 2 (e.g., the object for Subtask 2 may have already been observed while searching for Subtask 1).

• Continual-ObjectNav: Uses a multi-stage training process, where trajectories for new categories are introduced over time, without access to all earlier training data. In each stage, the agent learns to navigate to new object categories, and at test time, its navigation performance is evaluated in unseen environments for all previously learned categories. This setup falls under multi-stage incremental learning, introducing challenges such as catastrophic forgetting and the stability–plasticity trade-off, which do not occur in MultiON.

Therefore, our benchmark not only measures navigation performance but also emphasizes the agent’s ability to retain knowledge after multi-stage training.

Q2: Comparison with Recent Zero-shot Methods

We appreciate the reviewer’s observation that current map-based Zero-shot ObjectNav methods, such as UniGoal, have made significant progress. However, End-to-End (E2E) methods have unique advantages in the following aspects:

• Deployment and Efficiency: Zero-shot methods heavily rely on large vision-language models (e.g., GPT-4o) and multiple hand-crafted rules (e.g., object detection confidence thresholds, subgraph matching thresholds, passable area determination), which results in high deployment costs and latency. In contrast, E2E models have a more straightforward architecture (without requiring specialized hyperparameters) and offer faster inference, making them more suitable for resource-constrained navigation systems.

• Performance Advantage: When complete training data for all object categories is available, E2E methods can outperform Zero-shot methods. For instance, our baseline BEV-based end-to-end model [1] achieves a success rate (SR) of 65.2% and 47.74% on HM3D and MP3D, respectively, while UniGoal reports 54.5% and 41.0%, leading to improvements of approximately 10.7% and 6.7%. Furthermore, incorporating auxiliary losses and reinforcement learning fine-tuning [3] would further enhance performance.

  Model	SR（HM3D）	SPL（HM3D）	SR（MP3D）	SPL（MP3D）
  Uni-Goal[2]	54.5	25.1	41.0	16.4
  BEV-Based[1]	65.2	29.56	47.74	15.12

Research Motivation: Despite the superior performance under sufficient data, E2E methods suffer from significant catastrophic forgetting in Continual ObjectNav tasks. This is the main motivation for introducing the Continual-ObjectNav task and methodology. Our goal is to enable the model to continuously learn navigation for new categories while avoiding catastrophic forgetting.

• The agent must gradually learn to navigate new categories without revisiting all previously seen training data.
• Accumulating knowledge across multiple training stages is crucial, with the avoidance of catastrophic forgetting being the key challenge.

[1] Object Goal Navigation with Recursive Implicit Maps.

[2] UniGoal: Towards Universal Zero-shot Goal-oriented Navigation.

[3] PIRLNav: Pretraining with Imitation and RL Finetuning for ObjectNav.

Q3: Regarding Map-Based Methods

The advantages of map-based methods mentioned by the reviewer are primarily applicable under the Multi-ON task setting. In Multi-ON, the agent searches for multiple targets sequentially within the same environment, allowing it to continuously build and leverage a global map, which in turn accelerates the exploration of subsequent targets.

However, in the Continual-ObjectNav task, the focus is on evaluating the agent’s semantic navigation ability in unseen environments across multiple learning stages. Each test episode is conducted in a previously unseen environment and involves navigating to a single target (which may belong to either the current or a previously learned category). As such, there is no presence of multiple subtasks within the same environment, and thus map-based methods cannot exploit cross-subtask accumulation of spatial information or maps during testing.

Because each test occurs in a novel environment with a single target, the agent cannot build persistent maps through repeated exploration. Instead, it must rely on semantic-visual knowledge learned across training stages to perform the search effectively.

Concerns Regarding End-to-End Methods

We acknowledge the reviewer’s point that current Zero-shot methods have made substantial progress and offer strong generalization capabilities. However, End-to-End (E2E) models remain an important research direction in navigation tasks [1][3][4][5], as they provide clear advantages in inference speed, on-device deployment, and performance.

We have added the results of the BEV-based baseline model with auxiliary loss [1] in the Table. This model achieves an SR of 50.25 and SPL of 17.00 on MP3D, representing improvements of 9.25 and 0.6 over UniGoal on the same dataset. In addition, recent works such as UniNavid [4] and TrackVLA [5] also demonstrate the performance advantages of E2E methods. For example, by incorporating more navigation data [4], it achieves approximately a 20% SR improvement over UniGoal on the HM3D dataset. Furthermore, E2E methods can make effective use of newly collected data for model fine-tuning, which is something that Zero-shot methods cannot easily achieve.

While E2E methods have these advantages, they suffer from catastrophic forgetting in the Continual Object Navigation setting. This is the main reason our work focuses on E2E methods for Continual Object Navigation, aiming to address the challenge of retaining and accumulating knowledge and skills across multiple learning stages.

In existing Zero-shot ObjectNav, since Zero-shot methods do not involve any training, they avoid forgetting caused by parameter updates. However, these methods struggle to fine-tune models using newly collected navigation data, which can result in suboptimal performance in some cases.

In contrast, our benchmark focuses on how well a model retains prior knowledge after multiple training stages that use new navigation trajectories, enabling an embodied agent to continually adapt to new instructions. Both training and evaluation in our benchmark are conducted in a multi-phase, incremental manner.

As the importance of Zero-shot methods, we will add a discussion on the role of Zero-shot methods, such as UniGoal, in Continual ObjectNav in the related work section.

[1] Object Goal Navigation with Recursive Implicit Maps.

[2] UniGoal: Towards Universal Zero-shot Goal-oriented Navigation.

[3] PIRLNav: Pretraining with Imitation and RL Finetuning for ObjectNav.

[4]Uni-NaVid: A Video-based Vision-Language-Action Model for Unifying Embodied Navigation Tasks

[5]TrackVLA: Embodied Visual Tracking in the Wild

We thank the reviewers again for their valuable feedback and constructive suggestions. We hope our clarifications address the concerns raised, and we are happy to provide further explanation if there are any remaining questions.