Investigating and Mitigating Catastrophic Forgetting in Medical Knowledge Injection through Internal Knowledge Augmentation Learning

We are sincerely grateful for your kind and constructive feedback. Below are our responses to each of the concerns you raised.

1. Generalizability to other domain: Thank you for the insightful feedback. Our emphasis on the medical domain stems from a central observation: current LLMs often lack sufficient medical knowledge to support reliable decision-making in real-world clinical settings. To address this, our work investigates the issue of catastrophic forgetting that arises from medical knowledge injection and introduces an internal knowledge augmentation method, InternAL, to mitigate this problem and enhance injection effectiveness. As such, most of our experiments are conducted within the medical domain, as also acknowledged in the Limitation section. Nevertheless, the forgetting pattern we identify—proximity-dependent forgetting—may not be unique to the medical domain. Motivated by your comments, we extended our study to a different knowledge domain: human geography. Specifically, we extracted all sister city relationships from Wikidata (i.e., long-term partnerships between cities formalized via official agreements) and sampled 20,000 city pairs for analysis.

   Using the same methodology as in our main study, we created evaluation questions to determine which knowledge was poorly mastered (6,857 pairs selected for injection, denoted as $K_{\text{inject}}^{\text{SisCity}}$) and which was already well-mastered (4,145 pairs with model accuracy >75%, selected for evaluating forgetting, denoted as $D_{\text{eval}}^{\text{SisCity}}$). We then conducted knowledge injection using both the baseline (RefInject) and our proposed method (InternAL). Evaluation was performed on $D_{\text{inject}}^{\text{SisCity}}$(injected knowledge), $D_{\text{eval}}^{\text{SisCity}}$(uninjected knowledge), and several general-domain benchmarks.

   To further assess forgetting of related but semantically distant knowledge, we created an additional test set, CityLoc, containing 7,174 questions based on city latitude and longitude data from Wikidata. We also explored the effect of general-domain finetuning (GenFT), a known strategy for reducing catastrophic forgetting in broader domains. All experiments were conducted using Llama3-8B. The results are summarized as follows:

The results demonstrate that (1) direct knowledge injection (RefInject) leads to a 25% forgetting rate on $D_{\text{eval}}^{\text{SisCity}}$ and a 5.4% forgetting rate on CityLoc, while general-domain finetuning (GenFT) fails to effectively mitigate the substantial forgetting on $D_{\text{eval}}^{\text{SisCity}}$ and CityLoc; (2) Our method (InternAL) significantly mitigates forgetting on $D_{\text{eval}}^{\text{SisCity}}$ (from 64.3% to 82.0%) and on CityLoc (from 67.2% to 72.6%). We will include this experiment in the revised version to support the generalizability of our main findings.

2. Quality of extracted knowledge: Thanks for your constructive feedback. Due to the inherent hallucination tendencies of large language models, the knowledge extracted using the f_probe module may inevitably contain a certain degree of inaccuracy. However, since the extracted knowledge reflects the hallucination level of the target LLM itself, using this knowledge for internal augmentation learning does not amplify the model’s hallucination tendencies. Based on your helpful suggestion, we selected five relation types. For each relation, we randomly chose five head entities involved in the knowledge injection process. We then generated open-ended questions to prompt the model to produce tail entities that form the specified relation with the given head entity. The precision of the generated tail entities was used as the evaluation metric, and we conducted manual evaluation on both the original LLaMA3-8B model and the model trained using our proposed method (InternAL). The results are as follows:

   Precision	Original	InternAL (ours)
   drug-has indication-disease	47.7	90.0
   protein-interacts with-biological process	45.3	64.0
   disease-phenotype present-phenotype	83.3	84.7
   drug-side effect-effect	87.6	92.4
   exposure-linked to-disease	49.6	59.5
   Total	62.7	78.1

   Experimental results indicate that the model trained using our approach results in higher precision, implying that our method not only avoids reinforcing hallucinations but may actually contribute to mitigating them. We hypothesize that this effect arises because hallucinated content in the original model that conflicts with the newly injected knowledge is partially suppressed during the injection process, leading to an overall reduction in hallucination. We greatly appreciate your feedback and will further refine this analysis in the revised version of the manuscript.

3. Details of the negative sampling strategy: We sincerely appreciate your careful reading. The negative sampling strategy we implemented indeed corresponds to what you referred to as “hard negatives.” Specifically, based on each knowledge triple to be injected, we first extract all entities from PrimeKG that share the same type as the tail entity. We then exclude any entities that already form the specified relation with the given head entity, and randomly sample from the remaining pool to generate negative options. In our experiments, we found that this strategy yields training samples that enable efficient knowledge injection and sufficiently support our study. We will revise the manuscript to clarify this detail accordingly.

We are very grateful for your kind and constructive feedback. Below are our responses to each of the concerns you raised.

1. Knowledge form issue: We sincerely thank the reviewer for the insightful feedback. Our method primarily leverages multiple-choice questions (MCQs) for knowledge injection, following prior work [1–3], as MCQs are effective for conveying medical knowledge during SFT. We focus on structured knowledge (e.g., DrugBank, UMLS) due to its abundance in the medical domain and its utility for precisely computing semantic proximity, which facilitates analysis of forgetting.

We also agree that free-form texts such as medical textbooks and clinical narratives also contain rich medical knowledge and can be utilized for knowledge injection. In response to your suggestion, we conducted a preliminary study to assess the generalizability of our method to free-form texts. We sampled 2,000 clinical guidelines from a public dataset [3] and used GPT-4.1 to generate five MCQs per guideline. After manual review of a subset, we found the questions reliable for evaluation. We then evaluated LLaMA3-8B on these questions and selected 185 low-performing guidelines (accuracy < 50%) as the injection set ($K_{\text{inject}}$), and 1,542 high-performing ones (accuracy > 75%) as the evaluation set ($D_{\text{Eval}}$). For knowledge injection, we applied continued pretraining (CPT). Given the limited size of the injection corpus, we adopted a common data augmentation technique [4,5] that generates paraphrased variants of the injection samples to improve data diversity.

Building on this, we further extended our method (InternAL) to the unstructured knowledge setting. Specifically, we first prompted the LLM to extract key medical entities from each training sample. We then retrieved relevant knowledge associated with these entities by prompting the LLM again, and finally integrated this knowledge into the original samples to produce enriched pretraining texts. The evaluation results for this preliminary study are summarized as follows:

We observed the following key findings from the results: (1) Continued Pretraining (CPT) leads to substantial performance gains on tasks related to the injected knowledge but also causes significant forgetting that exhibits a proximity-dependent pattern (a 7.8-point drop on $D_{\text{eval}}$, an average decrease of 6.3 on medical benchmarks, and 4.2 on general datasets); (2) Our method (InternAL) mitigates forgetting effectively by integrating relevant internal knowledge into the training data, yielding notable improvements especially on the medical evaluation sets.

Furthermore, based on your constructive suggestions, we additionally constructed fill-in-the-blank (FIB) questions using GPT-4.1 based on clinical guidelines to investigate the effectiveness of knowledge injection across different question types. Specifically, we generated medical statements from the guidelines and then removed key information, prompting the evaluated model to fill in the missing content. We used GPT-4.1 to verify the correctness of the model’s answers by comparing them to the ground truth. The experimental results are shown below. (Here, $D_{\text{inject}}^{\text{mcq}}$ and $D_{\text{eval}}^{\text{mcq}}$ refer to multiple-choice questions generated from the injected knowledge and uninjected knowledge, respectively, while $D_{\text{inject}}^{\text{fib}}$ and $D_{\text{eval}}^{\text{fib}}$ refer to fill-in-the-blank questions generated from the injected knowledge and uninjected knowledge, respectively.)

Our experiments show that the proposed InternAL method can also improve the injection efficiency of the model on fill-in-the-blank questions and mitigate forgetting to some extent. These preliminary results further support that the proximity-dependent forgetting pattern is not confined to the specific knowledge types and training data examined in our study, and that the proposed method exhibit potential for extension to unstructured training data. We sincerely thank you again for your constructive suggestions, and we will further refine this experiment and analysis in the revised manuscript.

2. Difference from replay-based methods：Thank you for the thoughtful feedback. While both our InternAL method and replay-based approaches address catastrophic forgetting from a data perspective, they differ in two key ways: (1) Replay methods rely on access to prior training data, whereas our method extracts relevant internal knowledge from the LLM itself, requiring no past data; (2) Replay involves joint training with old and new data, increasing data volume and cost. In contrast, our method enhances the injection data without increasing its size, keeping training efficient. We will elaborate on these differences in the Related Work section.

3. Hallucination level issue: Thanks for your feedback. Our goal is to enhance knowledge injection and reduce catastrophic forgetting by extracting relevant knowledge from the target LLM itself. Since the augmented knowledge is generated by the model prior to injection, its hallucination level is inherently bounded by that of the model, with no external noise introduced. To validate this, we selected five relation types and, for each, randomly chose five head entities. We then prompted the model with open-ended questions to generate tail entities and measured precision through manual evaluation. Results were compared between the original LLaMA3-8B and the model trained with our method (InternAL) as follows:

   Precision	Original	InternAL (ours)
   drug-has indication-disease	47.7	90.0
   protein-interacts with-biological process	45.3	64.0
   disease-phenotype present-phenotype	83.3	84.7
   drug-side effect-effect	87.6	92.4
   exposure-linked to-disease	49.6	59.5
   Total	62.7	78.1

   Experimental results show that the model trained with our method achieves higher precision, suggesting that our approach not only avoids amplifying hallucinations, but may even help reduce them. We speculate that this may be because hallucinations in the original model that contradict the newly injected knowledge are partially suppressed during the injection process, thereby reducing the overall hallucination level. We sincerely appreciate your feedback and will further refine this analysis and incorporate it into the revised manuscript.

4. Representation-level analysis: Thank you for the constructive suggestion. To analyze how InternAL preserves semantically proximal knowledge, we encoded non-injected knowledge ($D_{\text{eval}}$) statements from PrimeKG using three models: the original LLaMA3-8B, RefInject, and InternAL. We extracted representation vectors across layers and computed the cosine distance between the original model and the two injected models. The difference in these distances reveals how much each method alters the original representations. Results are as follows: (we omitted some layers due to limited space)

   Layer	Cosine Distance of RefInject (%)	Cosine Distance of InternAL (ours) (%)	$\Delta$
   1	1.5	1.5	0
   ...	...	...	...
   6	16.5	11.1	5.4
   7	18.9	12.3	6.6
   8	20.1	13.6	6.5
   9	22.1	15.2	6.9
   10	23.5	16.1	7.4
   11	24.8	17.8	7
   12	25.5	19.5	6
   ...	...	...	...
   32	14.6	14	0.6

Experimental results show that InternAL has a smaller impact on non-injected knowledge representations than RefInject, which may result in less forgetting. Notably, the difference in cosine distance is significantly larger across layer 6-12 than other layers, which is consistent with prior findings on factual knowledge localization [5,6]. We will include this analysis, along with a t-SNE visualization, in the revised paper to further support the theoretical explanation.

5. MEMIT with MCQ format: Thank you for the suggestion. We initially followed prior work [6,7] using statement-style prompts for MEMIT. Based on your feedback, we began testing the MCQ format, but due to the high cost (∼2 days for injection), results are pending and will be shared later.

[1] Singhal et al. Toward expert-level medical question answering with large language models. Nature Medicine, 2025.

[2] Han et al. MedAlpaca--an open-source collection of medical conversational AI models and training data[J]. arXiv preprint, 2023.

[3] Chen et al. Meditron-70b: Scaling medical pretraining for large language models. arXiv preprint arXiv:2311.16079, 2023.

[4] Cabezudo M A S, Inácio M L, Pardo T A S. Investigating Paraphrase Generation as a Data Augmentation Strategy for Low-Resource AMR-to-Text Generation. Proceedings of the 17th International Natural Language Generation Conference. 2024.

[5] Zhang et al.. Co-occurrence is not factual association in language models. NeurIPS 2024.

[6] Meng et al. Locating and editing factual associations in gpt. NeurIPS 2022.

[7] Meng et al. Mass-Editing Memory in a Transformer. ICLR 2023.

We sincerely appreciate your kind and constructive feedback as well as your recognition of our work. Below are our responses to each of the concerns you raised.

1. Clarification of related knowledge: Thank you for your careful reading and thoughtful comments. Actually, the “related knowledge” we discuss in the paper is a gradable concept, which can be roughly categorized into three levels:

   • Level 1: Knowledge within the medical domain that is semantically highly related to the injected knowledge (proximal);
   • Level 2: Knowledge within the medical domain that is less semantically related (distal);
   • Level 3: Knowledge outside the medical domain.

   Our experimental findings reveal that the catastrophic forgetting caused by medical knowledge injection becomes progressively less severe from Level 1 to Level 3, exhibiting a proximal-dependent pattern. We sincerely apologize for not articulating this point clearly in the paper. We will revise the relevant definition in the updated version to clarify this concept.

   1. A second concern lies in the generalizability of the proposed approach. While the problem identified by the authors is indeed broadly relevant, the experiments are conducted on a relatively structured and constrained setting—primarily focusing on knowledge represented in the form of triples. However, in many real-world scenarios, knowledge does not naturally conform to such structured formats. It remains unclear how the proposed method would extend to more unstructured or complex forms of knowledge representation.

2. Generalizability issue: We sincerely appreciate your insightful comments. In this work, we primarily focus on structured medical knowledge primarily because structured knowledge bases (e.g., DrugBank, UMLS) are abundant in the medical domain and represent one of the fundamental forms of medical knowledge. Moreover, the nature of structured knowledge allows for more precise measurement of semantic proximity, aiding the analysis of the forgetting pattern. That said, we agree that in real-world applications, knowledge also appears in unstructured forms such as free-text clinical narratives. Based on your constructive feedback, we conducted a preliminarily study to explore forgetting behaviors when injecting unstructured knowledge and to validate the effectiveness of our proposed method in mitigating such forgetting.

   Specifically, we randomly sampled 2,000 clinical guidelines from a publicly available dataset [1], and used GPT-4.1 to generate 5 multiple-choice questions (MCQs) for each guideline. A subset of these questions was manually reviewed and found to be largely reliable for evaluation. We used these MCQs to evaluate the performance of LLaMA3-8B and selected 185 guidelines with accuracy below 50% as the injection knowledge set ($K_{\text{inject}}$), and 1,542 guidelines with accuracy above 75% as the evaluation set ($D_{\text{Eval}}$) to monitor forgetting. We adopt continued pretraining (CPT) as our approach for knowledge injection. Given the limited amount of injection data, we utilize commonly used data augmentation techniques [2,3], generating multiple paraphrased versions of the training samples in order to enhance the diversity of injection. Built on that, we further extend our proposed method (InternAL) to the unstructured knowledge. Specifically, we first extract key medical entities from each training sample using the target LLM, then prompt the model to recall relevant knowledge associated with these entities, and finally integrate the recalled knowledge into training samples to construct enriched pretraining texts. The evaluation results are summarized below:

We observed the following phenomena from the results: (1) Continued Pretraining (CPT) achieves considerable performance on tasks related to the injected knowledge, but leads to significant forgetting, which exhibits proximity-dependent forgetting characteristics (a drop of 7.8 on $D_{\text{eval}}$, an average decrease of 6.3 on medical benchmarks, and an average decrease of 4.2 on general datasets); (3) Incorporating the internal relevant knowledge of LLMs into the training data (InternAL) can effectively mitigate forgetting, especially on the medical evaluation sets.

We sincerely appreciate your suggestion and will include these additional experiments and analyses in the revised paper to discuss the generalizability of our findings and methods to unstructured knowledge.

3. Reason for choosing RefInject: Thank you for your thoughtful comment. We chose the RefInject method for knowledge injection for the following reasons: (1) Several prior studies in the medical domain [1,4,5] have demonstrated that training on MCQ datasets can effectively inject medical knowledge into LLMs and improve their performance on downstream medical tasks. (2) Other work has pointed out that continued pretraining on limited text data may fail to achieve effective generalization [6], whereas using corpora that contain implicit associations between entities (e.g., MCQ) has been shown to facilitate more efficient knowledge learning [2]. Motivated by these findings, we developed the RefInject method, which constructs MCQ-like questions that encode implicit associations between entities in the injected knowledge, aiming to enhance injection efficiency. Nevertheless, the proximity-dependent forgetting pattern we observed is not unique to this injection method. As mentioned in our previous response, we also observed this pattern when injecting knowledge through continued pretraining on unstructured clinical guidelines. We will further refine the relevant discussion in the revised manuscript to better clarify our motivation for choosing RefInject.

4. Clarification of the experimental results: We are very grateful for your thorough reading. The primary goal of knowledge injection is to supplement the missing knowledge in LLMs and improve their performance on tasks related to the injected knowledge. In this work, we randomly sampled 20,864 knowledge triples from PrimeKG and identified a subset where the LLMs performed poorly as the target knowledge for injection. After the injection, the model showed significant improvement on evaluation tasks corresponding to the injected knowledge (e.g., 9.7 to 77.4 for RefInject on $D_{\text{inject}}$). Although some forgetting occurred on the remaining triples, the overall performance across all 20,864 knowledge triples is improved by the injection (as shown in the $D_{\text{total}}$ column of the table below):

The results above demonstrate that the baseline injection method (RefInject) improves performance on all knowledge triples by 13.5% compared to the original model. In contrast, our proposed method (RefInject) further alleviates forgetting and achieves an additional 4.3% improvement across all triples, while also effectively mitigating forgetting on external benchmarks.

Thank you again for your kind comments. We will further clarify the relevant statements and add the total performance $D_{\text{total}}$ in the revised version to clarify the purpose and effectiveness of knowledge injection.

5. Number Mismatch Issue: Thank you for your careful reading. In fact, the results in lines 197–198 were from our initial single-run experiments. To verify robustness, we later conducted two additional independent runs and updated the table to report the average over all three runs, which showed consistent trends. However, we overlooked updating the corresponding description in the text. We apologize for the oversight and will correct this in the revised manuscript.

6. Paper Formatting Concerns: Thank you for your careful reading and helpful comment regarding Tables 2 and 3. We will address this in the revised version to improve clarity and presentation.

[1] Chen Z, Cano A H, Romanou A, et al. Meditron-70b: Scaling medical pretraining for large language models. arXiv preprint arXiv:2311.16079, 2023.

[2] Cabezudo M A S, Inácio M L, Pardo T A S. Investigating Paraphrase Generation as a Data Augmentation Strategy for Low-Resource AMR-to-Text Generation. Proceedings of the 17th International Natural Language Generation Conference. 2024.

[3] Zhang X, Li M, Wu J. Co-occurrence is not factual association in language models. NeurIPS 2024.

[4] Singhal, K., Tu, T., Gottweis, J. et al. Toward expert-level medical question answering with large language models. Nature Medicine, 2025.

[5] Han T, Adams L C, Papaioannou J M, et al. MedAlpaca--an open-source collection of medical conversational AI models and training data[J]. arXiv preprint, 2023.

[6] Christopher Potts, and Danqi Chen. Mquake: Assessing knowledge editing in language models via multi-hop questions. EMNLP 2023.

We sincerely appreciate your constructive feedback. Below, we address each of the concerns you raised.

1. Research scope issue: We sincerely appreciate your insightful comments. In fact, our focus on the medical domain is primarily motivated by a key observation: current LLMs still lack the necessary medical knowledge to perform reliably in real-world clinical scenarios. Our work investigates the phenomenon of catastrophic forgetting caused by medical knowledge injection and proposes an internal knowledge augmentation learning method to mitigate this issue, thereby improving the effectiveness of medical knowledge injection. Consequently, the majority of our experiments are conducted in the medical domain, which we also stated in the Limitation section.

   Nevertheless, considering the underlying mechanisms of knowledge injection, the proximal-dependent forgetting pattern identified in this work may also exist in other domains. Based on your constructive suggestion, we conducted an additional small-scale study beyond the medical field. Specifically, we selected human geography as the target domain and extracted all sister city relationships from Wikidata (i.e., long-term partnerships between cities established through official agreements), sampling 20,000 city pairs for experimentation.

   Following the same methodology used in our paper, we constructed evaluation questions to identify a subset of knowledge that was poorly mastered by the model (6,857 pairs selected for injection, denoted as $K_{\text{inject}}^{\text{SisCity}}$), and a well-mastered subset with model accuracy over 75% (4,145 pairs selected for evaluating forgetting, denoted as $D_{\text{eval}}^{\text{SisCity}}$). We then applied both the baseline method (RefInject) and our proposed method (InternAL) for knowledge injection. For evaluation, we leverage sister-city-based test sets $D_{\text{inject}}^{\text{SisCity}}$ and $D_{\text{eval}}^{\text{SisCity}}$ as well as on a suite of general-domain benchmarks. Furthermore, to evaluate the model’s forgetting of domain-related but semantically distant knowledge, we constructed an additional test set, CityLoc, by generating 7,174 questions based on the latitude and longitude information of cities extracted from Wikidata. We also studied the effect of general-domain finetuning (GenFT), an effective approach for mitigating catastrophic forgetting in the general domain. Experiments are conducted based on Llama3-8B, and the results are as follows:

   Model	$D_{\text{inject}}^{SisCity}$(Injected)	$D_{\text{eval}}^{SisCity}$ (Uninjected)	CityLoc
   Original (Llama3-8B)	9.1	89.3	72.6
   RefInject	93.9	64.3	67.2
   RefInject+GenFT	93.0	69.0	66.8
   InternAL (ours)	93.9	82.0	72.6
   InternAL+GenFT	94.5	83.9	72.1

   The results above show that (1) direct knowledge injection (RefInject) leads to a 25% forgetting rate on $D_{\text{eval}}^{\text{SisCity}}$ and a 5.4% forgetting rate on CityLoc, while general-domain finetuning (GenFT) fails to effectively address the substantial forgetting on $D_{\text{eval}}^{\text{SisCity}}$ and CityLoc; (2) Our method (InternAL) significantly mitigates forgetting on $D_{\text{eval}}^{\text{SisCity}}$ (from 64.3% to 82.0%) and on CityLoc (from 67.2% to 72.6%). Once again, thank you for your valuable feedback, and we will include this experiment in the revised version to support the generalizability of our main findings.

2. Model size issue: Thank you for your constructive suggestions. We selected the LLaMA and Qwen model families for our experiments primarily because they are representative open-source models with strong performance on several medical-domain benchmarks [1–2]. Considering the clinical applicability of LLMs, we primarily focus on the 8B model size, which offers a practical trade-off between computational resources, deployment cost, response latency, and reliability in real-world clinical environments. Larger models are often impractical to deploy due to their high resource demands and slower inference speed, while smaller models may lack sufficient medical knowledge to ensure reliable performance.

   Indeed, models of different sizes may exhibit distinct forgetting dynamics. Nevertheless, our preliminary experiments also revealed similar forgetting patterns across different model sizes. Motivated by your constructive comment, we further conducted additional experiments on Qwen3-1.7B (full-parameter finetune) and Qwen3-32B (using LoRA due to limited computational resources). The results are as follows:

   Qwen3-1.7B	$D_{\text{inject}}$ (Injected)	$D_{\text{eval}}$ (Uninjected)	MedQA	MMLU-Med	MMLU-O	ARC-C	CSQA
   Original	9.7	88.7	37.5	59.0	52.5	71.6	66.4
   RefInject	61.3	66.2	29.2	48.2	46.0	59.9	53.9
   RefInject+GenFT	60.4	72.8	31.3	55.0	52.1	68.1	63.3
   InternAL (ours)	59.3	75.1	31.9	51.7	47.4	61.7	58.3
   InternAL+GenFT	58.7	79.2	33.3	58.0	52.0	68.5	61.8

   Qwen3-32B	$D_{\text{inject}}$ (Injected)	$D_{\text{eval}}$ (Uninjected)	MedQA	MMLU-Med	MMLU-O	ARC-C	CSQA
   Original	10.2	92.4	68.2	80.8	68.5	86.9	83.5
   RefInject	71.1	62.7	59.1	77.6	67.6	84.7	84.4
   RefInject+GenFT	73.6	68.6	60.3	80.1	70.7	89.0	84.1
   InternAL (ours)	64.7	72.8	62.5	79.6	67.9	86.7	82.7
   InternAL+GenFT	67.7	83.1	63.2	81.8	72.5	90.0	83.0

Experimental findings are summarized as follows:

1. Both Qwen3-1.7B and Qwen3-32B exhibit proximity-dependent forgetting: For Qwen3-1.7B, knowledge injection (RefInject) led to performance degradation across multiple evaluation sets. While general-domain finetuning (GenFT) effectively mitigated forgetting in the general domain, the forgetting in the medical domain remained substantial. In contrast, Qwen3-32B showed relatively minor forgetting in the general domain after RefInject, but more pronounced forgetting in the medical domain (e.g., MedQA dropped from 68.2 to 59.1), which could not be effectively mitigated by GenFT.
2. Our proposed method, InternAL, effectively alleviates catastrophic forgetting in the medical domain for both models. Furthermore, it can be combined with GenFT to achieve even better mitigation of forgetting.

We sincerely thank the reviewer for the insightful suggestion. We will include this experiment in the revised version of the paper to better support the observed forgetting patterns and the generalizability of our proposed method.

[1] Wu C, Qiu P, Liu J, et al. Towards evaluating and building versatile large language models for medicine. npj Digital Medicine, 2025.

[2] Zhu S, Hu W, Yang Z, Yan J, Zhang F. Qwen-2.5 Outperforms Other Large Language Models in the Chinese National Nursing Licensing Examination: Retrospective Cross-Sectional Comparative Study. JMIR Med Inform. 2025.

Dear Reviewer w9Ko,

Thank you again for your time and review.

We noticed that you have clicked the “Mandatory Acknowledgement” button. We would like to kindly ask whether the additional experiments and clarifications provided in our rebuttal have sufficiently addressed your concerns. If you have any remaining questions, we would be very happy to engage in further discussion.

Best regards,

Authors of Submission 26858