Dear Reviewer zPHB,

Thank you for your thoughtful and detailed feedback on our submission. We truly appreciate the time and effort you have dedicated to reviewing our work and providing constructive comments. We are also grateful for your recognition of the originality and novelty of our approach. Below, we address each of your concerns in detail.

---

Impact of Model Size on UniDetox's Effectiveness

Comment: The size of the model might mitigate UniDetox's effectiveness since the ratio of parameter space and the volume of detoxifying text might be increased significantly, and no further preventive techniques could be introduced.

Question: What is the impact of the larger parameters space, like using larger models from 14B to even 405B models, on the small set of detoxifying text?

We appreciate your suggestion to evaluate UniDetox on even larger models (14B to 405B parameters). However, due to computational constraints, extending our experiments to models as large as 405B parameters is currently challenging. While UniDetox can directly utilize detoxifying text distilled from a smaller GPT-2 XL, implementing baselines like DExperts and Task Arithmetic for 405B models would require fine-tuning toxic variants of these large models, which is prohibitively expensive in terms of time and storage resources.

---

Computational Costs of UniDetox

Comment: Compared to other methods, UniDetox still requires model fine-tuning to mitigate the toxicity from the base models, making this approach's computational intensity a further challenge or question to understand.

Question: What are the costs regarding the computational resources, time required, or other considerations?

We have included a comparison of the computational resources and time required for each method in Table 7 of Appendix A.4 in our paper.

While UniDetox requires fine-tuning to detoxify models, it does not necessitate preparing a toxic variant for each model to be detoxified. In practice, we only fine-tune one toxic GPT-2 XL model to implement UniDetox across different target models.

In contrast, methods like DExperts and Task Arithmetic require preparing toxic versions of each target model (e.g., OPT-6.7B, Falcon-7B, LLaMA2-7B) for detoxification. Additionally, DExperts incurs extra computational overhead during inference due to its dual forward pass requirement.

Overall, UniDetox is more computationally efficient compared to these approaches.

---

Extending UniDetox to Multimodal Language Models

Comment: Limited to text-only language models, as the emerging traction of multimodal language models, how this approach further mitigates toxicity with a vast data modality remains challenging to this technique.

Question: Does this method apply to the scope beyond text-only models? Can this approach apply to multimodal language models as well?

Yes, theoretically, UniDetox can be extended to multimodal models. As long as there is toxic multimodal data available, UniDetox can be applied similarly to how it is used with text-only models.

For image-to-text models, UniDetox can be implemented in the same contrastive decoding fashion as in our study. For text-to-image models, as long as the decoding process can be adapted to a contrastive approach, UniDetox can be applied effectively.

---

We hope that these revisions effectively address your concerns. We believe that our approach offers a significant contribution to the field by providing a universally applicable detoxification method for large language models.

Thank you once again for your insightful feedback.

Sincerely,
The Authors

Quantity and Examples of Distilled Text

Comment: The amount of data produced via distillation is of great importance. Can the authors discuss about the data required for UniDetox to be effective? Additionally, I would be really curious in seeing some of the distilled sentences produced by UniDetox

As a detoxifying text, we sampled 640 texts, each with a maximum length of 256 tokens. We showed the details of hyperparameters in Appendix A.3. We also have included examples of detoxifying text in Appendix B.3, as well as sentences generated by models detoxified vias UniDetox in Appendix B.4 of our revised version.

Although we do not yet fully understand the specific linguistic characteristics of detoxifying text that drive its effectiveness, our statistical analyses provide useful insights:

• Toxicity Analysis (Table 3, Sec. 3.6): Detoxifying text shows consistently lower toxicity scores.
• Jaccard Similarity (Table 8, Appendix B.1): Detoxifying text exhibits less overlap with DGHS toxic data, indicating a distinct distribution.

These findings suggest that detoxifying text reduces toxic content and differs significantly from toxic data, enabling it to effectively guide models toward detoxification during fine-tuning.

---

Conclusion

In summary, we have: 1) included LM-Steer [3] as a new baseline to compare with UniDetox, 2) incorporated Few-Shot Accuracy on the MMLU benchmark to provide a more comprehensive evaluation, 3) presented results with a different hyperparameter configuration $\alpha = 0.05$ to showcase the robustness of UniDetox, and 4) added examples of the distilled detoxifying text as well as the detoxified model generations.

In the future, before camera-ready, we promise to extend the method to instruction-fine-tuned models.

We hope that these revisions effectively address your concerns and demonstrate the efficacy of UniDetox. We believe that our approach offers a significant contribution to the field by providing a universally applicable detoxification method for large language models. We kindly request you to reconsider the evaluation score in light of these clarifications and additions.

Thank you once again for your insightful feedback.

Sincerely,
The Authors

---

References:

[1] Whispering Experts: Neural Interventions for Toxicity Mitigation in Language Models. Suau et al. ICML 2024

[2] Goodtriever: Adaptive Toxicity Mitigation with Retrieval-augmented Models. Pozzobon et al. EMNLP 2023

[3] Word Embeddings Are Steers for Language Models. Han et al.ACL 2024

[4] DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Experts. Liu et al. ACL 2021

[5] Contrastive Decoding: Open-ended Text Generation as Optimization. Li et al. ACL 2023

[6] Contrastive Decoding Improves Reasoning in Large Language Models. O'Brien and Lewis. arXiv Preprint 2309.09117, 2023

Setting $\alpha$ to 0.05 still yields effective detoxification results, while the LM’s performance does not significantly change with the choice of $\alpha$. Thus, $\alpha$ does not have a significant impact on our results, and we consistently set $\alpha=0.1$, aligning with prior works [4, 5, 6].

Table 5. Detoxification Results for UniDetox with $\alpha=0.05$ and lr=1e-5

---

Concerns about Perplexity

Question: Results in Table 1. Could the authors comment on the fact that PPL goes down from 17.28 to 12.23 when using UniDetox? To me, this behavior is counterintuitive, since the PPL of a LLM rarely goes down when one intervenes on the LLM. In this case, fine-tuning on distilled data is an intervention on the original model. It is surprising that fine-tuning on a small synthetic dataset brings the LLM perplexity to 12.23 (which would be the PPL of a much larger/stronger LLM). I encourage the authors to provide details about the PPL evaluation, and a conclusive justification of why such PPLs are strongly decreasing.

Comment: I find very confusing the fact that PPL strongly decreases when the intervention is applied.

We have detailed the calculation methods of perplexity in Appendix A.2 in our revised paper.

We understand your concern regarding the drop in perplexity (PPL) observed when applying UniDetox. This is due to the adaptive plausibility constraint used for UniDetox. UniDetox generates detoxifying text using the plausibility constraint, which prevents the LMs from generating tokens with low probabilities. Thus, the detoxifying text is unlikely to contain tokens with low probabilities. By fine-tuning LMs on such detoxifying text, The fine-tuned LM becomes less likely to generate implausible tokens.

As PPL is not sometimes sufficient for evaluating LM’s capabilities, we added the MMLU as an additional evaluation benchmark. We appreciate your suggestion to include more metrics, which has indeed provided a more comprehensive evaluation of our method.

---

OOD vs. ID Toxicity Reduction

Question: Results in Table 1. Why is OOD toxicity reduction stronger than ID reduction? I would have expected a fine-tuning based approach like UniDetox to be much more effective on ID. Could the authors comment on this aspect?

Such OOD generalization happens because a model that forgets words used to attack some social groups will also forget similar offensive words directed at other groups. Although the following is merely a hypothesis, our dataset distillation-based method can be interpreted as condensing toxic text across multiple domains and generating detoxifying text that makes the model unlearn this condensed toxicity. Through condensation, the detoxifying text captures more generalized notions of toxicity, which are beneficial for mitigating toxicity in OOD domains.

---

Evaluating Instruction-Tuned Models

Question: Results in Table 1. For the safety preprompt to be effective, have the authors considered evaluating an instruction-tuned model such as gemma-2-2b-it?

Thank you for the suggestions. Following previous studies, we used models that are not instruction-tuned. Due to time and resource constraints during the response period, we have not yet evaluated UniDetox on instruction-tuned models like gemma-2-2b-it. However, we recognize the importance of this evaluation and plan to include it in our future work.

Dear Reviewer 7CV4,

Thank you for your thoughtful and detailed feedback on our submission. We sincerely appreciate the time and effort you have invested in reviewing our work and providing constructive comments. We are also grateful for your recognition of the originality and novelty of our approach. Below, we address each of your concerns in detail.

---

Derivation of the First-Order Taylor Approximation

Question: I kindly ask the authors to provide the steps followed to derive the Taylor approximation in Eq. 5 (from $s(\mathbf{x})$ in Eq. 1.

Specifically, we expand $\log p_{\theta_\mathrm{toxic}}(x)$ around $\log p_{\theta_\mathrm{base}}(x)$

$$\log p_{\theta_\mathrm{toxic}}(x) \approx \log p_{\theta_\mathrm{base}}(x) + (\theta_\mathrm{toxic} - \theta_\mathrm{base})^\top \mathbf{\nabla_{\theta}}\log p_{\theta_\mathrm{base}}(x)$$

Then, the contrastive score $s(x)$ can be rewritten as:

$$s(x) = \log p_{\theta_\mathrm{base}}(x) - \log p_{\theta_\mathrm{toxic}}(x) \approx (\theta_\mathrm{base} - \theta_\mathrm{toxic})^\top \mathbf{\nabla_{\theta}}\log p_{\theta_\mathrm{base}}(x)$$

---

Additional Metrics and Comparisons with Recent Methods

Comment: The experimental setup can be improved to convince the community about the validity of the proposed method. I suggest the authors include some zero/few-shot metrics (e.g., MMLU) to understand how the model abilities are impacted.

Comment: The evaluation proposed (methods compared) and metrics (PPL, n-gram) are too weak given the current abilities of LLMs and the thorough evaluations many works provide nowadays.

Comment: In Section 3.1, I encourage the authors to include some metric related to zero/few-shot abilities. For example, the overall score in MMLU seems a good candidate.

Comment: However, the universality claims are only validated using a small number of models and a non-comprehensive set of metrics.

To address these concerns, we have incorporated Few-Shot Accuracy on the MMLU benchmark as an additional evaluation metric. (The results are shown below)

Comment: Results in Table 1. I have some concerns about the methods used for comparison. Although DExperts was a successful method when it was proposed, several methods have shown superior performance in recent years. For example, [SuauICML24] proposes a method to reduce toxicity without any inference cost and not requiring fine-tuning, showing superior performance compared to DExperts or pre-prompting. Similarly, [PozzobonEMNLP23] propose an efficient method for toxicity mitigation. I encourage the authors to consider comparing with some newer method than DExperts, this work's imact could strongly benefit from that. I also suggest adding these recent methods in the Related Work.

We acknowledge the importance of comparing UniDetox with a broader range of baseline methods. In response to your suggestion, we have cited the works by Suau et al. [1] and Pozzobon et al. [2] in Section 4 of our revised manuscript.

However, there are practical challenges in directly implementing these methods within the limited time of the author response period. AURA (Suau et al.)[1] requires the preparation of non-toxic data, which presents a challenge for a fair comparison with UniDetox and other baselines like Task Arithmetic and DExperts, which operate using toxic-only data. Goodtriever (Pozzobon et al.)[2] requires constructing specific datastores for each model with a different architecture, adding significant time and computational complexity. This was infeasible given the tight constraints of the response period.

To strengthen our comparisons, we have incorporated LM-Steer [3], a recent method introduced in ACL 2024, which has demonstrated superior performance compared to DExperts. LM-Steer edits word embeddings during decoding time to steer a model's output. This method introduces a toxicity-steering matrix $W_\text{toxic}$ which is learned by fine-tuning on toxic data while keeping all other model parameters fixed. During decoding, detoxification is achieved by applying a linear perturbation to the word embedding using $W_\text{toxic}$:

$$\mathbf{e}'(x_t) = \mathbf{e}(x_t) - \epsilon W_\text{toxic} \mathbf{e}(x_t)$$

where $\mathbf{e}(x_t)$ denotes the word embedding of a token $x_t$, and $\epsilon$ is the hyperparameter controlling detoxification strength.

Table 1+2 shows the updated experimental results, including LM-Steer and few-shot accuracy on MMLU. As demonstrated, UniDetox consistently achieves the best detoxification results with a minor impact on downstream task performance. For a complete evaluation, including diversity metrics, please refer to Section 3.5 in our revised paper.

---

As demonstrated, UniDetox consistently achieves the best detoxification results with a minor impact on downstream task performance. For a complete evaluation, including diversity metrics, please refer to Section 3.5 in our revised paper.

Table1+2. Detoxification Results with an Additional Baseline LM-Steer and an Addition Metric MMLU Accuracy

---

Hyperparameter

Comment: Ablations: The proposed method has an important hyper-parameter $\lambda$. I missed an ablation study on $\lambda$, showing the impact of this choice in the final performance.

Comment: Moreover, details about the choice of $\lambda$ and the data produced are lacking.

We assume that the reviewer refers to the hyperparameter $\alpha$ as $\lambda$.

In Table 5 (Appendix A.3), we evaluated the case when varying the value of $\alpha$.

Dear Reviewer 7CV4,

Thank you for your thoughtful feedback, which has played a valuable role in improving our paper. As we near the end of the discussion period, we hope that our responses have adequately addressed the concerns you raised.

We would greatly appreciate it if you could let us know whether our revisions and clarifications have resolved your concerns. If there are any remaining questions or areas that need further clarification, please let us know.

Thank you again for your time and consideration.

Sincerely,

The Authors

Dear Reviewer gfVU,

Thank you for your valuable feedback on our submission. We greatly appreciate the time and effort you have taken to review our work and provide constructive insights. Below, we address your concerns in detail.

---

Related Work Coverage

Comment: The related section definitely lacks coverage of methods that are available for detoxification. Few of them include: Parameter-Efficient Detoxification with Contrastive Decoding, etc.

In our revised paper, we have expanded the related work section to include five additional methods, such as Parameter-Efficient Detoxification with Contrastive Decoding [1], Ethos [2], ProFS [3], AURA [4], and Goodtriever [5].

---

Comparisons with Recent Methods

Comment: There is a lack in comparison with more recent methods. An efficient method to benchmark against would be the cited LM Steering work(Han et al)

Question: Very curious to know why recent methods weren’t benchmarked against?

To strengthen our comparisons, we have incorporated LM-Steer [6], a recent method introduced in ACL 2024, which has demonstrated superior performance compared to DExperts.

We have also introduced Few-Shot Accuracy on the MMLU benchmark to ensure that the detoxified LMs maintain the performance on several tasks.

Table 1+2 shows the updated experimental results, including LM-Steer and few-shot accuracy on MMLU. As demonstrated, UniDetox consistently achieves the best detoxification results with a minor impact on downstream task performance. For a complete evaluation, including diversity metrics, please refer to Section 3.5 in our revised paper.

---

Use of Perspective API

Comment: Most work in Detoxification uses the Perspective API in their evaluations, but it has been omitted here without any explanation.

Question: Is there an obvious reason the Perspective API wasn’t used?

We did not use the Perspective API due to its stringent query-per-second (QPS) limit of 1 query per second, which poses a significant challenge for our evaluation setup. (https://developers.perspectiveapi.com/s/about-the-api-limits-and-errors?language=en_US) With nearly 30,000 entries per model and 20 models per baseline for hyperparameter tuning, this limitation renders the Perspective API infeasible for our experiments.

However, we have reached out to the Perspective API team to request an increased QPS. If granted, we commit to including Perspective API evaluations in the Appendix for the camera-ready version.

—

Examples of Detoxifying Text and Analysis

Comment: An example of what the generated“detoxifying” text looks like would help. Is this just “non-toxic” text ? Or does it have other interesting lexical properties? Some more analysis would be welcome here( even if in the appendix)

We appreciate this suggestion and have included examples of detoxifying text in Appendix B.3 of our revised paper.

Although we do not yet fully understand the specific linguistic characteristics of detoxifying text that drive its effectiveness, our statistical analyses provide useful insights:

• Toxicity Analysis (Table 3, Sec. 3.6): Detoxifying text shows consistently lower toxicity scores.
• Jaccard Similarity (Table 8, Appendix B.1): Detoxifying text exhibits less overlap with DGHS toxic data, indicating a distinct distribution.

These findings suggest that detoxifying text reduces toxic content and differs significantly from toxic data, enabling it to effectively guide models toward detoxification during fine-tuning.

---

As demonstrated, UniDetox consistently achieves the best detoxification results with a minor impact on downstream task performance. For a complete evaluation, including diversity metrics, please refer to Section 3.5 in our revised paper.

Table1+2. Detoxification Results with an Additional Baseline LM-Steer and an Addition Metric MMLU Accuracy

Conclusion

In summary, we have: 1) included LM-Steer [3] as a new baseline to compare with UniDetox, 2) incorporated Few-Shot Accuracy on the MMLU benchmark to provide a more comprehensive evaluation, and 3) added examples of the distilled detoxifying text.

In the future, before camera-ready, we commit to re-evaluating using Perspective API.

We hope that our responses effectively address your concerns and demonstrate the efficacy of UniDetox. We believe that our approach offers a significant contribution to the field by providing a universally applicable detoxification method for large language models. We kindly request you to reconsider the evaluation score in light of these clarifications and additions.

Thank you once again for your insightful feedback.

Sincerely,
The Authors

---

References:

[1] Parameter-Efficient Detoxification with Contrastive Decoding. Niu et al. HuCLLM 2024

[2] Ethos: Rectifying Language Models in Orthogonal Parameter Space. Gao et al. NAACL 2024

[3] Model Editing as a Robust and Denoised variant of DPO: A Case Study on Toxicity. Uppaal et al. arXiv Preprint 2405.13967, 2024

[4] Whispering Experts: Neural Interventions for Toxicity Mitigation in Language Models. Suau et al. ICML 2024

[5] Goodtriever: Adaptive Toxicity Mitigation with Retrieval-augmented Models. Pozzobon et al. EMNLP 2023

[6] Word Embeddings Are Steers for Language Models. Han et al.ACL 2024

Dear Reviewer jYcm,

Thank you for your valuable feedback on our submission. We appreciate the time and effort you have invested in reviewing our work and providing constructive comments. Below, we address each of your concerns in detail.

---

Additional Baseline Methods

Comment: The paper lacks a comparison with alternative alignment algorithms such as DPO or other detoxification baselines using task arithmetic between a base model and a tuned model, such as [1][2].

We acknowledge the importance of comparing UniDetox with a broader range of baseline methods. In response to your suggestion, we have cited the works by Gao et al. [1] and Uppaal et al. [2] in Section 4 of our revised manuscript. However, we observed that both Ethos [1] and ProFS [2] require the preparation of non-toxic data for implementation. This requirement presents a challenge for fair comparison with our approach and other baselines like Task Arithmetic and DExperts, which operate using toxic-only data.

To strengthen our comparisons, we have incorporated LM-Steer [3], a recent method introduced in ACL 2024, which has demonstrated superior performance compared to DExperts. Below is a brief overview of LM-Steer:

LM-Steer edits word embeddings during decoding time to steer a model's output. This method introduces a toxicity-steering matrix $W_\text{toxic}$ which is learned by fine-tuning on toxic data while keeping all other model parameters fixed. During decoding, detoxification is achieved by applying a linear perturbation to the word embedding using $W_\text{toxic}$:

$$\mathbf{e}'(x_t) = \mathbf{e}(x_t) - \epsilon W_\text{toxic} \mathbf{e}(x_t)$$

where $\mathbf{e}(x_t)$ denotes the word embedding of a token $x_t$, and $\epsilon$ is the hyperparameter controlling detoxification strength.

---

Additional Metrics

Comment: UniDetox demonstrates a drop in diversity metrics compared to other methods. A broader evaluation on other LM tasks would help validate its capability to preserve language quality.

To address the concern regarding diversity metrics and to provide a more comprehensive evaluation, we have introduced Few-Shot Accuracy on the MMLU benchmark as an additional metric.

Table1+2 shows the updated experimental results (the Table was shown below in response 2/3), including LM-Steer and few-shot accuracy on MMLU. As demonstrated, UniDetox consistently achieves the best detoxification results with a minor impact on downstream task performance. For a complete evaluation, including diversity metrics, please refer to Section 3.5 in our revised paper.

Hyperparameter Configuration

Comment: Although UniDetox is free from downstream model-specific hyperparameter tuning, results are sensitive to the selected learning rate and $\alpha$ parameters (Section 3.4). The need for different UniDetox configurations impacting performance lacks sufficient justification or analysis.

In UniDetox, the primary hyperparameters are the learning rate and the adaptive plausibility constraint $\alpha$. As for the learning rate, in both cases of 5e-5 and 1e-5, Unidetox exhibits higher detoxifying performance compared to other methods while maintaining LM’s ability, as shown in Table 1 and 2.

Regarding the adaptive plausibility constraint $\alpha$, we conducted a study when using different values of $\alpha$. We present detoxification performances using $\alpha=0.05$ and $\alpha=0.1$ in Table 5 (the Table was shown below in response 3/3). Setting $\alpha$ to 0.05 still yields effective detoxification results, while the LM’s performance does not significantly change with the choice of $\alpha$. Thus, $\alpha$ does not have a significant impact on our results, and we consistently set $\alpha=0.1$, aligning with prior works [4, 5, 6].

For a complete evaluation, including diversity metrics, please refer to Appendix A.3 in our revised paper.

---

Examples of Distilled Detoxifying Text

Question: Would the authors provide examples of the distilled detoxifying text for tuning? This would help clarify the properties and coherence of the distilled data.

We have included examples of the distilled detoxifying text in Appendix B.3 of our revised manuscript.

Although we do not yet fully understand the specific linguistic characteristics of detoxifying text that drive its effectiveness, our statistical analyses provide useful insights:

• Toxicity Analysis (Table 3, Sec. 3.6): Detoxifying text shows consistently lower toxicity scores.
• Jaccard Similarity (Table 8, Appendix B.1): Detoxifying text exhibits less overlap with DGHS toxic data, indicating a distinct distribution.

These findings suggest that detoxifying text reduces toxic content and differs significantly from toxic data, enabling it to effectively guide models toward detoxification during fine-tuning.

Setting $\alpha$ to 0.05 still yields effective detoxification results, while the LM’s performance does not significantly change with the choice of $\alpha$. Thus, $\alpha$ does not have a significant impact on our results, and we consistently set $\alpha=0.1$, aligning with prior works [4, 5, 6].

For a complete evaluation, including diversity metrics, please refer to Appendix A.3 in our revised paper.

Table 5. Detoxification Results for UniDetox with $\alpha=0.05$ and lr=1e-5

---

Conclusion

In summary, we have: 1) included LM-Steer [3] as a new baseline to compare with UniDetox, 2) incorporated Few-Shot Accuracy on the MMLU benchmark to provide a more comprehensive evaluation, 3) presented results with a different hyperparameter configuration $\alpha = 0.05$ to showcase the robustness of UniDetox, and 4) added examples of the distilled detoxifying text.

We hope that our responses effectively address your concerns and demonstrate the efficacy of UniDetox. We believe that our approach offers a significant contribution to the field by providing a universally applicable detoxification method for large language models. We kindly request you to reconsider the evaluation score in light of these clarifications and additions.

Thank you once again for your insightful feedback.

Sincerely,
The Authors

---

References:

[1] Ethos: Rectifying Language Models in Orthogonal Parameter Space. Gao et al. NAACL 2024

[2] Model Editing as a Robust and Denoised variant of DPO: A Case Study on Toxicity. Uppaal et al. arXiv Preprint 2405.13967, 2024

[3] Word Embeddings Are Steers for Language Models. Han et al.ACL 2024

[4] DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Experts. Liu et al. ACL 2021

[5] Contrastive Decoding: Open-ended Text Generation as Optimization. Li et al. ACL 2023

[6] Contrastive Decoding Improves Reasoning in Large Language Models. O'Brien and Lewis. arXiv Preprint 2309.09117, 2023

As demonstrated, UniDetox consistently achieves the best detoxification results with a minor impact on downstream task performance. For a complete evaluation, including diversity metrics, please refer to Section 3.5 in our revised paper.

Table1+2. Detoxification Results with an Additional Baseline LM-Steer and an Addition Metric MMLU Accuracy

Dear Reviewer jYcm,

Thank you for your thoughtful feedback, which has played a valuable role in improving our paper. As we near the end of the discussion period, we hope that our responses have adequately addressed the concerns you raised.

We would greatly appreciate it if you could let us know whether our revisions and clarifications have resolved your concerns. If there are any remaining questions or areas that need further clarification, please let us know.

Thank you again for your time and consideration.

Sincerely,

The Authors

Dear Reviewer jYcm,

This is a final reminder that today is the last day the reviewers can post messages on the forum. We understand it’s a busy time, but we’d greatly appreciate hearing your thoughts on the revised paper and responses.

We’ve made significant updates, including new experiments, to address your and other reviewers’ concerns. The other reviewers have shared positive feedback on the revisions, and your feedback would also be invaluable.

Thank you for your time,

Sincerely,

The Authors