Toward Domain Translation with Monolingual Domain Data Only

Weakness 2:

There are some issues with the evaluations. First, the base model that is used is OPUS-MT, but the domain-specific datasets that are used for evaluation for EN<->DE come from OPUS, so they were used to train the base model. Thus, this is not a true scenario of domain adaptation to an unseen domain, but one of domain shift. It is not clear to me whether this was done intentionally, but I think it would be preferable to do some domain adaptation evaluations with unseen data (and this might explain the lack of consistent positive results for any of the domain adaptation models, including the baselines and the EBM approach). Second, confidence is used as an evaluation score, when it is not clear that this correlates with any sort of meaningful MT evaluation. Third, the examples given in table 5 point more towards overfitting to the vocabulary of a specific dataset (not a specific domain) than to any true translation quality improvements.

Thank you for your feedback.

First, indeed, as pointed out, Aharoni et al. (2020) created the dataset by automatic mining and multiple cleaning of OPUS, so the en <-> de data may be included in the OPUS dataset. However, we would like to emphasize the following: 1) Domain data constitutes only a small part of the training data. The model we use is generally trained on broad data, and domain adaptation serves to awaken its capability in a specific domain. 2) We manually refined the test set to ensure accurate evaluation. 3) We also considered en <-> zh data, where the Chinese dataset, UM-Corpus, is an indirect access dataset. For these reasons, we have already conducted experiments on unseen data, and the comprehensive evaluation results demonstrate that our method surpasses standard domain adaptation techniques.

Secondly, a model with high confidence indicates that it can generate outputs with greater certainty relative to the domain distribution. For example, even if the scores before and after applying our method are tied, an increase in confidence implies that the model is better specialized for the specific domain. Additionally, we evaluated our method using multiple evaluation metrics, not just confidence. It is crucial to consider both confidence and MT metrics when analyzing the results. Confidence serves as an alternative angle of evaluation to reinforce MT metrics. Because these are independent evaluations, there is no requirement for correlation. In fact, their independence allows for a more multifaceted analysis.

Regarding the third point, we are unsure how you would propose distinguishing between true domain adaptation and overfitting in your consideration. However, our method leverages the target-side word distribution for domain adaptation. As the number of target domain data increases, the domain distribution approaches the true distribution, enabling more accurate domain shifts. Furthermore, the examples we provided do not include words explicitly present in the constructed target domain distribution, indicating inductive generation. A detailed qualitative analysis of such unseen terminology is provided in Appendix I. This supports our claim that specific domain adaptation has been achieved. If you have any suggestions for better ways to present this, we would be grateful for your comments.

---

Weakness 3:

This paper should take a broader view of the literature, including terminology-constrained machine translation (which seems to be hinted at as the ultimate goal of the proposed approach, e.g. in line 410 and table 5) as well as cases where the assumptions I listed in item 1 are relaxed (e.g., training a domain-specific model from scratch). In addition, the following paper is directly related (even without taking a broader view) and should be used as a baseline.

We added literature on terminology-constrained machine translation and your recommendation of the paper to reference. However, the domain discussed in the recommended literature involves a large amount of data, which is not well-suited for the stress setting we assume in this study, where only a small amount of target domain data is used. Nonetheless, the experiments in the multi-domain setting presented in the paper are valuable references, and we have incorporated similar experiments into our study.

Specifically, we added Appendix G and Table 9, which investigate the variations of Fine-tuned and CDPG in mixing two domains. The result shows that the vanilla fine-tuning methods are influenced by noisy data, because the performance decreases with the increase of noisy data. But, our proposed method, CDPG, shows more robust performance.

To demonstrate the effectiveness of our training method, we conducted evaluations using the same model and the same data. 
(a) While it is true that LLMs are popular nowadays, using LLMs is not ideal as a baseline for demonstrating the effectiveness of our method because their training data differs significantly. Moreover, this paper focuses on proving the functionality of our method, and larger parameter models are simply one variation. We have noted this as a Limitation and also discussed it in the future directions. 

(b) As you are aware, our tuning settings are among the most standard ones. If we were to examine each part of the Transformer, such as the encoder, FFN, or specific layers, it is possible to explore various alternatives. However, exploring all these configurations goes beyond the scope of this study. We are researching a novel training method, not engaging in a SOTA competition or a comprehensive meta-evaluation. 

(c) Using test data to select checkpoints constitutes p-hacking. Thus, we argue that selecting checkpoints based on validation data is appropriate. Furthermore, since word distribution acquisition and actual translation evaluation are separate aspects of the data usage, your concern does not apply, and we believe our approach sufficiently generalizes.

First, indeed, as pointed out, Aharoni et al. (2020) created the dataset by automatic mining and multiple cleaning of OPUS, so the en <-> de data may be included in the OPUS dataset. However, we would like to emphasize the following: 1) Domain data constitutes only a small part of the training data. The model we use is generally trained on broad data, and domain adaptation serves to awaken its capability in a specific domain. 2) We manually refined the test set to ensure accurate evaluation. 3) We also considered en <-> zh data, where the Chinese dataset, UM-Corpus, is an indirect access dataset. For these reasons, we have already conducted experiments on unseen data, and the comprehensive evaluation results demonstrate that our method surpasses standard domain adaptation techniques. 

Secondly, a model with high confidence indicates that it can generate outputs with greater certainty relative to the domain distribution. For example, even if the scores before and after applying our method are tied, an increase in confidence implies that the model is better specialized for the specific domain. Additionally, we evaluated our method using multiple evaluation metrics, not just confidence. It is crucial to consider both confidence and MT metrics when analyzing the results. Confidence serves as an alternative angle of evaluation to reinforce MT metrics. Because these are independent evaluations, there is no requirement for correlation. In fact, their independence allows for a more multifaceted analysis. 

Regarding the third point, we are unsure how you would propose distinguishing between true domain adaptation and overfitting in your consideration. However, our method leverages the target-side word distribution for domain adaptation. As the number of target domain data increases, the domain distribution approaches the true distribution, enabling more accurate domain shifts. Furthermore, the examples we provided do not include words explicitly present in the constructed target domain distribution, indicating inductive generation. A detailed qualitative analysis of such unseen terminology is provided in Appendix I. This supports our claim that specific domain adaptation has been achieved. If you have any suggestions for better ways to present this, we would be grateful for your comments.

We are grateful for the time and effort you have invested in reviewing our manuscript. We take each of your concerns seriously and are confident that we can address all issues raised to your satisfaction.

---

Weakness 1:

This paper would benefit from clearly defining and justifying the task at hand. It addresses domain adaptation with target-side monolingual data, and that is well justified. However, there are several implicit assumptions throughout that do not seem to be stated clearly or given a justification. For example, a) assuming that they are starting from a generic trained model to adapt, and cannot train a model to the relevant domain from scratch (section 4.2), b) assuming that the adapted model shouldn't have catastrophic forgetting of the original domain (line 45), c) assuming that it is desirable to do well on tasks that were unseen in both the original model and the adaptation data (table 6), and d) that the adapted output should exhibit minimal changes compared to the original model (line 410). There is nothing wrong with these assumptions necessarily, but you need to clearly state them and justify why you are restricting your exploration to these.

Thank you for your valuable comments.

(a): We explained in the first paragraph of the introduction that obtaining parallel data for training on domain-specific data to create high-quality translation models is challenging.

(b)-(d): We added supplementary information about the meaning of catastrophic forgetting. Specifically, training with limited data often leads to local optima and involves exploring a loss landscape that differs from that of pretraining. This results in parameters that diverge from the original optimal solution, causing performance degradation. Therefore, sufficient data is required for domain shifts. In our main related work, Korbak et al., (2022), reinforcement learning is used for domain shifts. However, methods requiring scoring, like reinforcement learning, focus on task-specific learning, which can lead to a decline in generalization performance. Here, generalization performance refers to fundamental forgetting, such as a loss of fluency in generated sentences, resulting in unnatural text. Based on these observations, our approach is motivated by the goal of not only producing fluent translations but also ensuring consistency in domain-specific terminology to achieve effective domain shifts.

Nevertheless, in response to this comment and Question 5, which related to the same problem, we conducted additional experiments to evaluate performance changes in the general domain and updated Table 6 in the manuscript.

We attached the simple table as follows:

Specifically, we added the difference between Fine-tuned and Pre-trained and the difference between DCDPG and Pre-trained on a generic domain in Table 6 as a pivot for the relative difference between Fine-tuned and DCDPG in crossing domains. In this way, we can show the collapse of the vanilla fine-tuning methods, and the generalization of CDPG. Thank you for your suggestion!

1a) "We explained in the first paragraph of the introduction that obtaining parallel data for training on domain-specific data to create high-quality translation models is challenging."

(b)-(d): We added supplementary information about the meaning of catastrophic forgetting. Specifically, training with limited data often leads to local optima and involves exploring a loss landscape that differs from that of pretraining. This results in parameters that diverge from the original optimal solution, causing performance degradation. Therefore, sufficient data is required for domain shifts. In our main related work, Korbak et al., (2022), reinforcement learning is used for domain shifts. However, methods requiring scoring, like reinforcement learning, focus on task-specific learning, which can lead to a decline in generalization performance. Here, generalization performance refers to fundamental forgetting, such as a loss of fluency in generated sentences, resulting in unnatural text. Based on these observations, our approach is motivated by the goal of not only producing fluent translations but also ensuring consistency in domain-specific terminology to achieve effective domain shifts. 

4.a: Given the ubiquity and availability of LLMs, if they are able to do this task more effectively with the same in-domain data (regardless of the initial training data), then it would be good to clarify in what cases your method would be useful. For this reason, it would be good to compare to an LLM baseline.

Thank you. To clarify our understanding, when you refer to LLMs, are you specifically referring to decoder-side language models like LLaMA, rather than pre-trained models in general? While your suggestion is intriguing, our study focuses on encoder-decoder models, and we consider discussions about decoder-only models to be out of scope.

As you correctly pointed out, our research emphasizes the “ubiquity and availability of pre-trained NMT models”, which we highlighted in response to your comment 1(a) by explicitly stating this in the introduction. We believe this addition has clarified the scope of our study.

Given the focus on “ubiquity and availability of pre-trained NMT models”, we argue that our experiments are appropriately designed to be comparable and sufficiently demonstrate the practical utility of our method in addressing “what cases our method would be useful”.

---

4.b: Do you have a citation for fine-tuning only the attention weights being "among the most standard [tuning settings]"? Also, I'm sorry my writing was unclear; but I don't believe I suggested "examin[ing] each part of the Transformer, such as the encoder, FFN, or specific layers" or "engaging in a SOTA competition or a comprehensive meta-evaluation". I stand by my statement that given the small dataset and focus on target-side data, fine-tuning the decoder only is a more intuitive comparison than fine-tuning the attention weights only.

Thank you for your explanation. Do you have any references to support your opinion? In our study, we first conducted full fine-tuning for a fair comparison with CDPG tuning that. Next, we used LoRA to update only the attention weights, following the standard setting (Hu et al., 2021) as mentioned in Section 4.2.

Our paper primarily focuses on introducing a new tuning strategy, and we believe it is appropriate to compare results under the same tuning conditions. For a fair comparison, we assert that tuning both the encoder and decoder parameters is the most reasonable setup.

---

4.c: I am not sure what is the cause of the misunderstanding here. The original review points out that both checkpoint selection and fine-tuning are done on the same dataset; nowhere does it suggest "using test data to select checkpoints". Standard practice would be to use separate datasets for validation/checkpoint selection (development set), fine-tuning (typically called the training set), and evaluation (test set and held-out set). It seems (please correct me if I misread the lines cited in the original review) that you are using the same set for validation/checkpoint selection and fine-tuning, and a separate second set for evaluation.

Thank you for clarifying your question. First, the "features" used for CDPG are derived from the validation set, but since the validation set data is not directly used during training, this does not affect the checkpoint selection for CDPG. Furthermore, to ensure fairness, we also use the validation set for training the Fine-tuned (i.e., vanilla fine-tuning) and LoRA. Given the scarcity of domain-specific data (2,000 samples for German and 3,000 samples for Chinese), further splitting the data would reduce the effectiveness of training (small validation set is not effective; big validation set harms the training).

Considering that our setup involves fine-tuning with a small learning rate over 10 epochs, the risk of overfitting is relatively low. Therefore, we select checkpoints based on their inference performance (i.e., BLEU scores and BERTScores) on the validation set, rather than relying on the model’s loss on the validation set (equivalent to the training loss). Empirically, the selected checkpoints are not always the last ones, but even when the last checkpoint is chosen, it consistently brings improvements in testing within the respective domain compared to the pre-trained models.

---

Thank you for your effort in reviewing our manuscript again. We believe we have addressed all of your concerns. However, if you have any further questions or additional concerns, please don’t hesitate to share them. We are committed to responding sincerely and thoroughly within the remaining time.

We look forward to your reply.

Dear Reviewer n3SL,

We understand that you may be busy, and we greatly appreciate the time and effort you have already dedicated to this review process. We believe that your lack of response to our rebuttals is not due to irresponsibly abandoning your role as a reviewer, but rather because the concerns regarding this paper have been resolved. Sorry to bother you, but if the concerns have indeed been addressed, we kindly request you to update your scores for Soundness, Presentation, Contribution, and Overall Rating to reflect the resolution of all issues, as is typically expected of reviewers following the discussion period.

Best regards,

Anonymous Authors

Thank you very much for your insightful comments. We are encouraged by your comments.

---

Weakness 1:

My biggest concern with this work is its limited novelty especially compared to the work of Korbak et al. 2022. This work proposes how to represent the feature probabilities μ with unigram probabilities, but that seems to be all beyond the Korbak paper. Some discussion of exactly where this work differs would help differentiate them and highlight novelty.

Thank you for your feedback. Regarding the differences from Korbak et al., we have addressed this starting from line 51: “Korbak et al. (2022) had only verified the effectiveness of CDPG for small shifts, such as translating numeral nouns (e.g., ‘two’) as digits (e.g., ‘2’). We extend the framework by using the token-level statistics of the target domain as features and constructing a large number of EBMs, approximating these to meet their constraints. Specifically, we shift the pre-trained NMT models toward the token-level unigram distribution of the target domain by CDPG, enabling domain shifts that better consider the frequency information of the entire target domain.” We believe this explanation clarifies the differences between our approach and theirs.

---

Weakness 2:

The use of EBMs and CDPG is motivated by avoiding catastrophic forgetting, but there does not seem to be evidence of this in the results. In the intro, catastrophic forgetting is equated to reduction in translation performance, which is not quite what other work defines as catastrophic forgetting. Catastrophic forgetting refers to becoming worse on the original data distribution - showing improved performance on the target domain is not enough to evaluate catastrophic forgetting. The methods should ideally be evaluated on the original domain as well in order to show this.

Thank you. Some studies on domain adaptation for NMT consider the results after a domain shift. Notably, Korbak et al. (2022) examined how their approach enables stable control without causing catastrophic forgetting in the target domain compared to general reinforcement learning methods. One definition of catastrophic forgetting refers to the situation where, while achieving the desired control, models often lose the ability to generate fluent sentences that original models can, as described in the introduction’s second paragraph. We evaluated the performance changes in the general domain and updated Table 6 in the manuscript. Specifically, we added the difference between Fine-tuned and Pre-trained and the difference between DCDPG and Pre-trained on a generic domain in Table 6 as a pivot for the relative difference between Fine-tuned and DCDPG in crossing domains. The results show the collapse of the vanilla fine-tuning methods and the generalization capacity of CDPG.

---

Weakness 3:

Calling this method “monolingual data only” or unsupervised does not seem to work well with proposing dynamic CDPG as a main contribution in this work. These descriptors seem to only target the extension of vanilla CDPG to MT. These descriptors seem to only target the extension of original CDPG to MT.

CDPG is sensitive to changes in the top_p parameter, which affects its quality. To address this, we propose Dynamic CDPG as an auxiliary approach, attempting to optimize parameter settings using a small amount of bilingual data. As indicated by the title, our main contribution is CDPG, while Dynamic CDPG is presented as a variant to emphasize that even when using CDPG, comparable results can be achieved to those of Dynamic CDPG, which performs dynamic parameter exploration. Therefore, as you pointed out, our main contribution is vanilla CDPG, and Dynamic CDPG serves as a variant to highlight and complement vanilla CDPG.

---

Weakness 4:

There is a lack of implementation details regarding the EBM models and the fine-tuning procedure.

Thank you. We used the disco library released by Kruszewski et al. (2023) as the base for our implementation. The disco library supports methods such as Korbak et al. (2022) and provides basic apis such as EBM and trainer. When using a large number of EBMs, computational bottlenecks can arise, for example, during scoring. To address this, we optimized the code and incorporated techniques such as matrix operations to improve processing speed. Please note that these optimizations are purely technical and do not affect the mathematical formulations. We have included details about the libraries used in the Experimental Setup section.

[1] disco: a toolkit for Distributional Control of Generative Models (Kruszewski et al., ACL 2023)

Dear Reviewer UmaA,

We understand that you may be busy, and we greatly appreciate the time and effort you have already dedicated to this review process. We believe that your lack of response to our rebuttals is not due to irresponsibly abandoning your role as a reviewer, but rather because the concerns regarding this paper have been resolved. Sorry to bother you, but if the concerns have indeed been addressed, we kindly request you to update your scores for Soundness, Presentation, Contribution, and Overall Rating to reflect the resolution of all issues, as is typically expected of reviewers following the discussion period.

Best regards,

Anonymous Authors

Thank you for your constructive feedback. Your thoughtful and detailed comments would be instrumental in strengthening our paper and clarifying our arguments.

---

Weakness 1: Concerns with practicality and scalability

Thank you for the interesting feedback. This study relies solely on the word distribution of the target domain. As the size of the data increases, the distribution approaches the true target domain, enabling higher-quality domain adaptation. At the same time, this study demonstrates that the approach works sufficiently well even with a small amount of available data like stress setting. Therefore, we believe it scales effectively. Furthermore, under the stress settings with limited domain data we aim at, the scalability issue is not critical.

Additionally, we conducted experiments on robustness. By mixing multiple domains during training, we examined whether the method exhibits over-reliance. We added Appendix G and Table 9, which compare Fine-tuned and CDPG in mixing two domains. The result shows that the vanilla fine-tuning methods are influenced by noisy data, in which the performance decreases with the increase of noisy data. However, our proposed method, CDPG, shows more robust performance.

Your feedback has helped us emphasize the merits of this paper even more clearly. Thank you!

---

Weakness 2: Evaluation metric suitability

BERTScore allows for more detailed evaluation compared to sentence-level metrics like COMET, as it considers token-level meaning and calculates recall and precision.

We have added footnote 12 in Section 4.3 Evaluation, stating that the neural fine-tuned metric COMET is trained using data from WMT evaluation tasks. As a result, while it shows high correlation with human judgments for in-domain data, it has been shown that metrics like BLEU or embedding-based BERTScore exhibit higher correlation for out-of-domain data, especially with respect to domain-specific data.

For this reason, we prioritized BERTScore as the main evaluation metric over COMET and similar metrics. Nevertheless, we included COMET as a sentence-level metric in Appendix G. These results not only highlight the challenges COMET faces with domain-specific data again but also partially support our claims. Incorporating COMET has made the paper’s narrative more accessible and easier for readers to follow. Thank you for your suggestion!

---

Question 1: Domain feature generalization

Thank you for pointing out this interesting question! We added Appendix I in the manuscript to show the deeper influence brought by CDPG from two instances. We have indicated in our manuscript that CDPG will increase the confidence of models. As an additional influence, the repeats in pre-trained models are resolved, like this instance:

Inference of Pre-trained: PPM. - Nein, nein, nein, nein, nein, nein, nein, nein, nein, nein…

Inference of CDPG: PPM.

Then, we also can find some interesting instances, like this:

Inference of Pre-trained: Dies ist der Typ Ihres Tunnelgeräts.

Inference of CDPG: Dies ist der Typ Ihres Tunnelgerätes.

Here, the “Tunnelgerätes” hits the reference and the term by fixing the original inaccurate word “Tunnelgeräts”. Meanwhile, “Tunnelgerätes” is not a feature used in fine-tuning! Therefore, this instance shows the generalization of domain features. We guess that the essence of the increase of confidence is to encourage the model closing to the target domain.

---

Question 2: Sensitivity to validation bilingual set in DCDPG

In Section 5.2, Table 3, Table 4, and Appendix C, we have already found and stated that the bilingual data do not always bring gains to CDPG. We also stated that CDPG is sensitive to changes in the top_p parameter, which affects its quality. In this case, Dynamic CDPG, as an auxiliary method, aims to optimize the hyper-parameter automatically using a small amount of bilingual data (validation set). Based on our design, if all optimizations are rejected in the validation set, DCDPG would fine-tune the model in the case where top_p = 1, which avoids heavy updating in the pre-trained models to ensure the bottom-bound of CDPG. Moreover, as indicated by the title, our main focus is CDPG, while Dynamic CDPG is presented as a variant and an attempt to automate optimal parameter settings.

---

If you have any further concerns or feel that certain points were not addressed adequately, please don’t hesitate to let us know. We will respond sincerely.

Dear Reviewer VUPi,

We understand that you may be busy, and we greatly appreciate the time and effort you have already dedicated to this review process. We believe that your lack of response to our rebuttals is not due to irresponsibly abandoning your role as a reviewer, but rather because the concerns regarding this paper have been resolved. Sorry to bother you, but if the concerns have indeed been addressed, we kindly request you to update your scores for Soundness, Presentation, Contribution, and Overall Rating to reflect the resolution of all issues, as is typically expected of reviewers following the discussion period.

Best regards,

Anonymous Authors

We greatly appreciate your insightful comments which have significantly contributed to improving our manuscript and enhancing its appeal to broader readers.

---

Weakness 1: Limited Scope of Evaluation

We conducted experiments across four language directions (en <-> de, en <-> zh) and four domains, resulting in a total of 16 tasks. Additionally, since the experimental data and settings we used align with those commonly employed in domain adaptation evaluations for machine translation, such as those mentioned in related work sections, we believe that our experiments are sufficiently comprehensive. Our experiments further support the general applicability of our approach in the context of domain adaptation studies in machine translation.

---

Weakness 2: Lack of Robust Comparison

From your comments, we have added the results of back-translation to Appendix F and Table 8, and mention it in Line 196 to strengthen our statements. Since methods for generating synthetic data, such as back-translation, are susceptible to noise, the performance of a model fine-tuned with clean data serves as an upper bound when the available sentences are the same. This trend is observed in the results of the additional experiments. On the other hand, our approach outperforms the fine-tuning results, concluding that it is superior to methods using monolingual data, such as back-translation. Moreover, based on your interesting Questions 3 and 5 about diversified and noisy data, we also added Appendix G and Table 9, which compare Fine-tuned and CDPG in a scenario of mixing two domains. The result shows that the vanilla fine-tuning methods are influenced by noisy data, in which the performance decreases with the increase of noisy data. However, our proposed method, CDPG, shows more robust performance under the same condition. Thank you for your feedback, which has allowed us to argue the effectiveness of our approach more robustly!

---

Weakness 3: Complexity of Methodology

Thank you for your feedback. From around line 86, we have tried to provide examples for a more intuitive explanation, but we will make an effort to make it even more intuitive.

Dear Reviewer oE5u,

We understand that you may be busy, and we greatly appreciate the time and effort you have already dedicated to this review process. We believe that your lack of response to our rebuttals is not due to irresponsibly abandoning your role as a reviewer, but rather because the concerns regarding this paper have been resolved. Sorry to bother you, but if the concerns have indeed been addressed, we kindly request you to update your scores for Soundness, Presentation, Contribution, and Overall Rating to reflect the resolution of all issues, as is typically expected of reviewers following the discussion period.

Best regards,

Anonymous Authors

Summary of Our Responses

We highlight key points of discussion and revisions made to address the reviewers’ concerns. Detailed responses to individual reviewers can be found in their respective response sections.

We have addressed all raised concerns, and no additional concerns were raised by reviewers after our responses.

---

Reviewer oE5u

The reviewer’s primary concern was the lack of robust comparisons. In response, we added discussions on back-translation in Appendix F and Table 8 as a new baseline and included experiments on mixed-domain scenarios in Appendix G and Table 9. These additions demonstrate that our method is robust and applicable to diverse domain adaptation settings.

---

Reviewer VUPi

The reviewer expressed concerns about robustness to noise, scalability, and mixed-domain extensions. To address this, we included mixed-domain results in Appendix G and Table 9, which illustrate the robustness, scalability, and extensibility of our method.

Additionally, we introduced COMET-based evaluations in footnote 12 of Section 4.3 and Appendix H, which aligned with our previously reported results, further confirming the validity of our findings.

---

Reviewer UmaA

The reviewer participated in one round of discussion and raised two primary concerns: novelty and the definition of catastrophic forgetting.

• Novelty: We emphasized in the manuscript that this work is the first to successfully scale EBMs. Previous studies were limited to small-scale, synthetic experiments, but our method applies to challenging, real-world tasks in unsupervised domain adaptation. We demonstrated superior performance even under weak supervision settings like Dynamic CDPG, and our contributions extend beyond incremental improvements. By enabling scaling, we achieved a significant breakthrough in unsupervised domain adaptation.

• Catastrophic Forgetting: To address this, we added generic domain results in Section 6.1 and Table 5 to show the stability of our method. We also clarified the definition of catastrophic forgetting in the Introduction (highlighted in green), showing that our method is effective both in reinforcement learning tasks and in downstream tasks.

---

Reviewer n3SL

The reviewer participated in one round of discussion and raised concerns about definitions (Weakness 1), the validity of en-de results (Weakness 2), and experimental suggestions (Weakness 4). (Weakness 3 was already addressed during the first discussion.)

• Definitions (Weakness 1): We clarified the definitions in the Introduction (highlighted in green).

• Validity of Results (Weakness 2): We supported our results with literature and added case studies in Appendix I, demonstrating their validity.

• Experimental Suggestions (Weakness 4): We clarified that the suggested experiments had already been conducted and included in our results. Most concerns were definitional, and we addressed them thoroughly with additional manuscript revisions.

---

Conclusion of our discussions and responses

Despite multiple reminders and more than a week since the final rebuttal, no additional concerns were raised by any reviewer. We take this as an indication that our responses have fully addressed all concerns and that our arguments have been accepted. If there were further concerns, it would have been the reviewers’ responsibility to engage in the discussion. The lack of additional comments strongly suggests that all concerns have been resolved and that our rebuttal has been accepted as satisfactory.

We are also pleased to note that all reviewers recognized the validity of our experiments and the effectiveness of our proposed method. Their comments and suggestions allowed us to further refine our manuscript, making it clearer and more accessible to a broader audience.

This concludes our summary. We hope for a constructive AC-reviewer discussion phase and a fair evaluation free from undue biases related to low scores or confidence.

---

Thank you for your consideration.

Best regards,

Anonymous Authors