XTransplant: A Probe into the Upper Bound Performance of Multilingual Capability in LLMs via Cross-lingual Transplantation

Thank you so much for your valuable and Insightful review comments. We have responded to your questions in detail and look forward to your reply to address any further questions you may have.

1. Response to your concern about "The computational cost"

Thank you for your insightful feedback. We fully acknowledge the computational cost of exhaustively investigating all N×N combinations in our main experiments.

• However, it is important to emphasize that the extensive experiments serve to demonstrate the potential of XTransplant as a simple yet promising mechanism for enhancing multilingual capabilities and cross-lingual transfer, without any modifications to the LLM itself. This does not imply that future work must replicate the same approach in terms of computational expense.

• As the first work to explore this mechanism, we felt that the effort spent on a comprehensive analysis was justified, and we believe it has successfully showcased the substantial potential of XTransplant.

• Additionally, while our work involves significant computational expense, we do not intend for future research to replicate this approach in a resource-intensive manner. Our goal is to inspire future work that can utilize the concept/mechanism of cross-lingual feed-forward activations transplantation to model design or training phases, such as exploring cross-lingual feed-forward interactions and connections between different layers or language-specific regions, to enable the model better leverage its intrinsic multilingual knowledge.

• Furthermore, the idea of transplantation is not limited to multilingual or cross-lingual scenarios but could also be applied to cross-task or cross-domain settings.

• Lastly, we hope that the large-scale experiments we conducted, along with our subsequent analysis and discussions, can inspire future research and provide valuable insights for the broader research community.

Thank you so much for your valuable and Insightful review comments. We have responded to your questions in detail and look forward to your reply to address any further questions you may have.

1. Response to your concern about "Lack of comparative comparisons"

Thank you very much for your valuable feedback and for suggesting additional comparative settings. We had indeed considered these aspects during the design of our experiments, but we feel that these methods may not be entirely suitable for our experiments. And I would like to clarify the reasons behind our choices as follows:

• (Regarding Supervised Fine-Tuning methods) As noted, the experiments we designed for comparisons, such as XTransplant, Parallel Input in Multiple Languages (PIM), and Chain-of-Thought (CoT), are all training-free methods aimed at further unlocking the model's potential during inference stage. These methods are focused on activating the model’s capabilities without altering its parameters through additional training data. And the purpose of XTransplant, as well, is to show that the LLM has a lot of room to fully realize its capability in multilingual and culture-aware tasks without any changes to LLM itself.

  On the other hand, task-specific supervised fine-tuning modifies the model’s parameters based on extra training data, leading to performance improvements that cannot be considered as an additional unlocking of the model's inherent potential, since the model itself is being modified. Therefore, we chose not to include supervised fine-tuning methods as a direct comparison.

  Moreover, it is worth mentioning that our proposed XTransplant mechanism can also serve as a post-enhancement technique after supervised fine-tuning, allowing for a "second jump" in performance beyond the improvements achieved through fine-tuning alone. (similar to how Chain-of-Thought can be applied both before and after supervised fine-tuning.)

• (Regarding Translation-then-Inference approaches) We would like to clarify that the "translation-then-inference" pipeline is not suitable for the tasks in our study. It seems there might be some aspects of our work that were not fully clear, which led to your concern about this.

  Our study involves 4 specific datasets: XNLI, XQuAD, XCOPA, and GlobalOpinionQA.

  • For the multilingual datasets (XNLI, XQuAD, and XCOPA), all of the questions are linguistically parallel across languages. This means that the questions in each language split are essentially the same, with differences only in the language versions. Therefore, the "English Set" in these datasets already represents the translated version, as you described with a "machine translation" approach. As such, there is no need to apply additional translation, which is why we do not consider "translate-then-inference" as a suitable baseline for these tasks.

  • Regarding the multicultural dataset, GlobalOpinionQA, all the questions and answers are in English. The purpose of this dataset is to explore how well models respond to questions from different cultural backgrounds within an English context. Therefore, machine translation would not be applicable here either.

We hope this clarifies the reason why we do not utilize "supervised fine-tuning" and "translation-then-inference" as comparative comparisons.

2. Response to your concern about "No comparisons between LLMs of different model sizes"

Thank you very much for pointing this out. We acknowledge that this could be a potential limitation in our work. The reason we did not compare different model sizes is that the scope of our experiments was already exceptionally large.

As noted between "lines 213–215" of our manuscript, to obtain the instance-aware upper bound of XTransplant, we perform inference on all N^2 possible source and target layer selection strategies for each instance. For example, in LLaMA-2-7B-Chat with N=32, N^2=1024 times inferences are conducted per instance. Our main experiments involve 3 LLMs and 4 pilotset datasets, resulting in over 800 hours of computation on 8 * A800-SXM4-80GB. Therefore, including comparisons across different model sizes would nearly double or triple this computational cost, which would be challenging given our current resources.

We appreciate you for pointing this out and hope you could understand the computational constraints.

1. Response to your concern about "Your suspicion that the improvements of the UpperBound method result mainly from lurking into the gold answer"

Thank you for your feedback.

Firstly, we would like to clarify that, in the entire mechanism of XTransplant, there is no possibility of golden answer leakage. The only input to XTransplant consists of two different language inputs, and the core operation involves transplanting the feed-forward activations of a specific decoder layer from one language input into the inference process of another language input. This process does not involve any interaction with the gold answer. If there is any lingering doubt regarding this, we have provided the source code for XTransplant in the Supplementary Material, and we would be happy for you to check it.

Regarding your suspicion about the improvement in the "English2English" setting, we would like to emphasize that this result is indeed logical. In this setting, XTransplant simplifies to replacing the feed-forward activations between different decoder layers within the same input. Different decoder layers of LLMs capture distinct features of the input and activate different neurons (i.e., knowledge). Thus, the transplanting operation between these layers can strengthen feature propagation and encourage feature reuse, leading to performance improvements. This phenomenon is analogous to the dense connections in DenseNet[1] (Huang and Liu), which has been shown to enhance feature flow and overall performance.

We hope this explanation addresses your concern.

[1] Densely Connected Convolutional Networks.

2. Response to your concern about "The computational cost of choosing the upper bound pairs of layers"

Firstly, we would like to emphasize that XTransplant is not a specific methodology but rather a mechanism — one that has not been explored in previous works — using cross-lingual feed-forward activation transplantation.

Therefore, our work does not focus on "how to choose the upper bound pairs of layers?". The essence of our work is to explore cross-lingual feed-forward activation transplantation as a bold, novel attempt in the area of enhancing multilingual capabilities and cross-lingual transfer.

We begin by conducting extensive experiments (over 800 hours of computation on 8 * A800-SXM4-80GB) to demonstrate the underutilization of current LLMs' multilingual potential. And through the upper bound results of XTransplant and subsequent analysis (including the alleviation you just mentioned), we try to show that XTransplant is a feasible and promising mechanism for improving both multilingual capabilities and cultural adaptability. While we certainly hope our work can inspire more related future research, such as the more efficient and accurate identification of appropriate pairs of layers, our focus in this paper is on the exploration and analysis of this novel cross-lingual feed-forward activation transplantation approach, rather than providing a specific methodology.

3. Response to your concern about "The lack of results for base models"

Thank you for raising this concern. I would like to explain why we chose to use "chat/instruct" models rather than base models in our experiments:

• The concern you raised may stem from the fact that many works in the field of "model editing" and similar areas often focus on base models. However, we believe that working with base models may not be as practically meaningful, as they are typically limited to theoretical explorations. In real-world applications, we mostly work with instruction-tuned models like "chat/instruct" models, which have been adapted for task-specific responses. Therefore, in our experiments related to XTransplant, we chose to focus on "chat/instruct" models.

• One significant issue with using base models is their lack of instruction-following capabilities. This would require us to prompt the models with multiple demonstrations in a few-shot setting, introducing extra complexity and noise into our analysis. To avoid this additional interference and ensure a cleaner evaluation of XTransplant, we chose to use "chat/instruct" models, which are better equipped to answer a wide range of questions directly from the dataset.

I hope this clarifies our reasoning for selecting these models, and we appreciate this thoughtful feedback.

We sincerely thank you for your response to our previous clarification. In particular, we deeply appreciate your diligence in carefully reviewing our code. However, it seems there are still some misunderstanding regarding our work.

Regarding your unresolved concern about the "golden answer leakage"

• In our previous reply, we mistakenly assumed that your concern about golden answer leakage was related to the Transplantation process in XTransplant, so we clarified that the Transplantation process does not involve any interaction with the gold answer. However, from your latest response, we now realize that your concern is actually about golden answer leakage during the evaluation stage. We apologize for this misunderstanding and appreciate the opportunity to clarify further.

• We confirm that your interpretation of our evaluation code is correct, and we sincerely thank you for thoroughly examining it. That said, we would like to emphasize that the purpose of this Upper Bound evaluation is specifically to benchmark the potential upper limit of cross-lingual latent interaction achieved by XTransplant under our experimental setup.

• While we understand that the high Upper Bound value may raise skepticism, we want to make it clear that we are not using these results to claim the superiority of XTransplant over other baseline methods. In fact, we make no such assertions in the conclusions of our experiments. Instead, the high Upper Bound results are intended to illustrate the extent to which multilingual capabilities can be unlocked through the XTransplant mechanism without modifying LLM itself. It's totally different from the common evaluation of directly comparing numerical results to demonstrate the superiority of one method over another.

• Additionally, from your comments, we noticed your concern that achieving the Upper Bound from the N^2 (e.g., N^2=32^2=1024 for LLaMA-2) possible combinations might be coincidental. However, as discussed in Section 5.3: Analysis of Layer-Specific Effectiveness in Source and Target Selection, one of our findings suggests that even when computing the Upper Bound from a space of size N, XTransplant can still achieve a satisfactory upper bound performance.

• Moreover, another part of our paper—Figure 7 in Section "Proportion Analysis of XTransplant Outcomes"—further supports that the high Upper Bound is not coincidental. In Figure 7, the filled area represent the proportion of all N^2 outcomes where XTransplant produces correct answers, while the shaded area represent incorrect answers. We observe that the proportion of correct answers is significantly high across all datasets, often exceeding 50%.

Regarding your concern about the "our explanation to the improvements under English2English setting"

• We would like to clarify the underlying principles behind the improvements observed with XTransplant in the English2English setting, which, to some extent, are indeed similar to the principles of "DenseNet" (Huang, Liu et al. 2018). Reviewer 4 has acknowledged this connection, and we are happy to provide a more detailed explanation for your understanding.
• In non-En->En or En->non-En scenarios, XTransplant performs latent interactions between two different language versions of the same input. However, in the En->En setting, XTransplant simplifies to replacing the feed-forward activations between different decoder layers within the same input (for example, the feed-forward activations of the 15th layer can be replaced with those from the 25th layer, and the forward propagation can then continue from the 15th layer with the newly transplanted feed-forward activations). This operation can be thought of as establishing interaction connections between different decoder layers, allowing feature (or knowledge) sharing between these layers to some extent. This is why we describe it as strengthening feature propagation and encouraging feature reuse. This form of cross-layer connection is indeed quite similar to the connections in "DenseNet", which we believe provides a reasonable explanation for the improvements observed in the English2English setting. And of course, as you correctly pointed out, the concept of Upper Bound is not related to "DenseNet".

We hope these clarifications address your concerns and provide further confidence in our work.

3. Response to your concern about "Not utilizing machine translation as a baseline"

Thank you for pointing this out. However, we would like to clarify that the "translate-then-xxx" pipeline is not suitable for the tasks in our study. It seems there might be some aspects of our work that were not fully clear, which led to your concern about this.

Our study involves 4 specific datasets: XNLI, XQuAD, XCOPA, and GlobalOpinionQA.

• For the multilingual datasets (XNLI, XQuAD, and XCOPA), all of the questions are linguistically parallel across languages. This means that the questions in each language split are essentially the same, with differences only in the language versions. Therefore, the "English Set" in these datasets already represents the translated version, as you described with a "machine translation" approach. As such, there is no need to apply additional translation, which is why we do not consider "translate-then-xxx" as a suitable baseline for these tasks.

• Regarding the multicultural dataset, GlobalOpinionQA, all the questions and answers are in English. The purpose of this dataset is to explore how well models respond to questions from different cultural backgrounds within an English context. Therefore, machine translation would not be applicable here either.

We hope this clarifies the reason why we do not utilize "machine translation" as a baseline.

4. Response to your concern about "The computational cost of choosing the upper bound pairs of layers"

• As emphasized in our first response, XTransplant is not a specific methodology but rather a mechanism. Therefore, our work does not focus on "how to choose the upper bound pairs of layers?" The essence of our work is to explore cross-lingual feed-forward activation transplantation as a bold, novel attempt in the area of enhancing multilingual capabilities and cross-lingual transfer. We begin by conducting extensive experiments (over 800 hours of computation on 8 * A800-SXM4-80GB) to demonstrate the underutilization of current LLMs' multilingual potential. And through the upper bound results of XTransplant and subsequent analysis, we try to show that XTransplant is a feasible and promising mechanism for improving both multilingual capabilities and cultural adaptability. While we certainly hope our work can inspire future research, such as the more efficient and accurate identification of appropriate pairs of layers, our focus is on the exploration and analysis of this novel cross-lingual feed-forward activation transplantation approach, rather than providing a specific methodology.

• And it is important to emphasize that the extensive experiments serve to demonstrate the potential of XTransplant as a simple yet promising mechanism for enhancing multilingual capabilities and cross-lingual transfer, without any modifications to the LLM itself. This does not imply that future work must replicate the same approach in terms of computational expense.

• As the first work to explore this mechanism, we felt that the effort spent on a comprehensive analysis was justified, and we believe it has successfully showcased the substantial potential of XTransplant.

• Additionally, while our work involves significant computational expense, we do not intend for future research to replicate this approach in a resource-intensive manner. Our goal is to inspire future work that can utilize the concept/mechanism of cross-lingual feed-forward activations transplantation to model design or training phases, such as exploring cross-lingual feed-forward interactions and connections between different layers or language-specific regions, to enable the model better leverage its intrinsic multilingual knowledge.

• Lastly, we hope that the large-scale experiments we conducted, along with our subsequent analysis and discussions, can inspire future research and provide valuable insights for the broader research community.

5. Response to your question about "What is the reason for applying XTransplant only when generating the first new token?"

Thank you very much for noticing this question. We are sorry for not providing a more detailed explanation of this aspect in our paper.

In autoregressive generation, applying XTransplant during the generation of the first new token essentially introduces the benefit of feed-forward activations from another language across the entire sequence generation process. This is because all subsequent tokens are influenced by the activations cached from earlier steps. If XTransplant were applied during the generation of every token, it would be a redundant operation and could even cause the model's output to break down.

We hope this clarifies the reasoning behind.

1. Response to your concern about "The unfairness of comparisons"

Thank you for your feedback. It seems there might be a misunderstanding regarding the purpose of our main experiments. While we have presented both the baseline results and the upper bound results of XTransplant in the same table, our intention was not to claim that XTransplant is a superior method to the other baselines. We do not make any such comparison or assertion in the conclusions of our experiment.

In fact, one of the key objectives of our main experiments is to highlight that both the PIM results and the upper bound results of XTransplant show performance improvements beyond the original model’s capabilities. This serves to demonstrate that, for current LLMs, simply prompting them under multilingual or culture-aware contexts does not fully exploit their multilingual potential. And by showcasing the upper bound results of XTransplant, we aim to highlight that there is still significant room for improvement.

Additionally, we would like to emphasize that XTransplant is not a specific methodology but rather a mechanism — one that has not been explored in previous works — using cross-lingual feed-forward activation transplantation.

We hope this clarifies our intentions.

2. Response to your concern about "The test set is too small"

As you pointed out, we have retained 50 instances for each language in the test set (aka the pilotset in our paper). The specific details are as follows:

「Involved Languages / Cultures」

XNLI (15): ar, bg, de, el, en, es, fr, hi, ru, sw, th, tr, ur, vi, zh

XQuAD (12): ar, de, el, en, es, hi, ro, ru, th, tr, vi, zh

XCOPA (11): en, et, ht, id, it, sw, ta, th, tr, vi, zh

GlobalOpinionQA (24 cultures): am, ar, bn, de, el, en, es, fr, hi, id, it, ja, ko, nl, pt, ru, sv, sw, tl, tr, uk, ur, vi, zh-CN (QA pairs are all presented in English)

「Sample Size (50 samples per language / culture)」

XNLI: $50 \times 15 = 750$

XQuAD: $50 \times 12 = 600$

XCOPA: $50 \times 11 = 550$

GlobalOpinionQA: $50 \times 24 = 1200$

And for multilingual tasks (XNLI, XQuAD, XCOPA), we ensure that each language set consists of the same 50 questions, presented in different language versions.

The reason we have sampled only 50 instances per language is due to the extensive scale of our experiments. As noted between "lines 213–215" of our manuscript, to obtain the instance-aware upper bound of XTransplant, we perform inference on all N^2 possible source and target layer selection strategies for each instance. For example, in LLaMA-2-7B-Chat with N=32, N^2=1024 times inferences are conducted per instance. Our main experiments involve 3 LLMs and 4 pilotset datasets, resulting in over 800 hours of computation on 8 * A800-SXM4-80GB.

We appreciate you for pointing this out and hope you could understand the computational constraints.