Thank you for taking the time to review our manuscript. We greatly appreciate your valuable feedback. Below, we provide our responses to your comments:

[Question 1] Difference between CoMe and LaCo

LaCo belongs to the category of WSLP methods, while CoMe differs fundamentally from LaCo in terms of layer selection, fusion strategy, and post-training protocol.

• Fusion Strategy: CoMe employs a concatenation-based merging at the sub-module level, whereas LaCo essentially utilizes a weighted sum of layer weights.

• Layer Group Selection: CoMe uses the SBI score to select a group of adjacent layers for merging, requiring only a single forward pass on a calibration dataset in one iteration. In contrast, LaCo sets a threshold $\mathcal{T}$ and traverses from shallow layer index $\mathcal{L}$ to deep layer index $\mathcal{H}$, merging $\mathcal{I}$ layers at a time. If the similarity between the pruning model’s output and the original model’s output is within a threshold $\tau$, the fused layer will be adopted. Otherwise, it rejects the merging process and retries with different layers until the target number of layers is reached or all layers are traversed. LaCo is a post-hoc pruning method and, due to multiple hyperparameters, it is challenging to obtain a model with the required number of layers. Therefore, we only reproduced LaCo experiments on the LLaMA-2 series models.

• Post-training: CoMe employs a fine-grained, layer-wise post-training protocol, which operates by aligning the output distributions of individual layers during the training process. In contrast, LaCo adopts a model-wise post-training approach that aligns the logits of the original model and the pruned model through conventional knowledge distillation. Due to our resource constraints (2 × A100 40G), we only compared CoMe with layer-wise post-training methods. We were unable to reproduce LaCo’s post-training results and thus excluded it from the comparative analysis.

[Weakness 2] Theoretical Evidence for Merging

The theoretical motivation for CoMe is grounded in two key observations from previous layer pruning studies, as detailed in Section 3 of our manuscript. First, DLP directly prunes entire layers, which can result in the loss of important weights. Second, WSLP ignores the correlation between the weights of adjacent layers, and its blind weighted fusion often leads to over-smoothed weight distributions. Based on these insights, we propose a novel layer pruning framework. CoMe decomposes each layer into submodules whose inputs and outputs are independent, so that removing or adding submodules does not affect the input or output of the remaining submodules. For example, as analyzed in Appendix C, channels in the FFN structure and heads in the MHA structure can be treated as such submodules. By concatenating these submodules, CoMe avoids weight smoothing and more effectively preserves important weights or functional components.

It is important to note that layer pruning is inherently a lossy process, and there is currently no effective theoretical method to quantify the extent of this loss. This limitation is common across many layer pruning approaches (e.g., LaCo, MKA, SLEB, ShortGPT, etc.), making the field largely empirical.

In our manuscript, we provide detailed experimental comparisons between DLP and WSLP (see Figure 3 and Figure 13), as well as between DLP and CoMe (see Figure 10 and Figure 14), clearly illustrating the effects of direct layer pruning, weighted sum fusion, and concatenation-based fusion on model performance. Ablation studies of the fusion operation in Table 2 further demonstrate that CoMe is more effective than WSLP.

[Weakness 1] Experiment on Larger Models

Due to limited computational resources (2 × A100 40G), we were unable to conduct experiments on models larger than 20B parameters. We have provided experimental results on models of various scales, including 4B, 7B, 8B, and 13B. We plan to extend our experiments to 70B and even larger models when resources permit in the future.

[Weakness 3] Lack of Quantitative Comparison with LoRA Fine-tuning

We appreciate the reviewer’s suggestion. Our method is orthogonal to LoRA fine-tuning, as the two approaches serve fundamentally different purposes. LoRA aims to reduce the number of trainable parameters through low-rank decomposition for parameter-efficient fine-tuning. CoMe’s post-training protocol is designed to rapidly align the merged layers using information obtained during pruning, thereby restoring model performance. Additionally, because CoMe’s pruning process is dynamic, the number of layers and training time required for post-training also vary accordingly. In our paper, we compared post-training methods that are more closely related to CoMe’s motivation (Table 1) and provided comparisons of training processes and resource consumption (Appendix B.2 and Table 7).

We hope these responses can address your concerns. If you have further questions, we would be happy to provide additional clarification.

Thank you very much for your detailed review and valuable feedback on our manuscript. We truly appreciate the time and effort you dedicated to evaluating our work. We highly value your assessment and kindly request that you let us know if you have any remaining questions or points that require further clarification. Your additional feedback would be immensely helpful in helping us improve the quality of our work. Thanks again for your time and consideration.

Thank you for taking the time to review our manuscript. We greatly appreciate your valuable feedback. Below, we provide our responses to your comments:

[Weakness 1] Impact of Calibrating Dataset Types, Scales, and Adding Noise on CoMe

Thanks very much for your insightful suggestion. Both the BI/SBI metrics and the channel sensitivity metric indeed rely on a calibration dataset. We present the impact of the calibration dataset size and type on CoMe’s performance in Figure 8 and Figure 11, and provide analysis in Lines 324–340. As shown in Figure 11, CoMe is not highly sensitive to the calibration dataset. Different calibration datasets lead to only minor fluctuations in performance, and the average accuracy is largely unaffected in terms of reasoning ability. However, in terms of PPL, using pre-trained data (e.g., Wikitext-2, C4, PG19) as the calibration set yields better results than using instruction-tuning datasets in QA format (e.g., MMLU, Alpaca). As shown in Figure 8, when the sample size of the Wikitext-2 calibration dataset exceeds 128, the pruning performance of CoMe becomes stable.

To further investigate the impact of dataset noise on CoMe’s performance, we conducted an additional experiment in which we randomly shuffled the word order in a proportion of sentences in the calibration set, thereby disrupting the linguistic structure. Using the default experimental settings from the main paper (256 calibration samples), we found that shuffling up to 60% of the sentences does not significantly affect CoMe’s performance. However, when more than 80% of the sentences are shuffled, the average accuracy drops by more than two points.

[Weakness 2 & Question 1] Limitation of the Heuristic Parameter Preservation Ratios

Thank you for your insightful comments. We fully agree that the use of heuristic parameter preservation ratios poses limitations in terms of adaptability across different structures, as also noted in our Limitations section (Lines 357–359). Our experiments in Figures 10 and 14 demonstrate that the optimal merging ratio differs between MHA and FFN, indicating that a fixed ratio may not be universally optimal. Furthermore, as shown in Figure 7, we investigated the impact of the fusion ratio hyperparameter and found that heavily prioritizing one layer while neglecting others leads to significant performance degradation. These results suggest that adopting adaptive or learned fusion ratios for different structures could potentially yield better results than our current heuristic approach. We consider this an important direction for future work and plan to explore adaptive strategies in subsequent research.

[Weakness 3 & Question 2] Analysis of Computational Overhead Introduced by Hierarchical Distillation

Our post-training hierarchical distillation strategy is based on the layer pruning information obtained from the first stage of CoMe. As a result, the number of optimized layers in CoMe-mp is not fixed. In CoMe-mp, each sub-step only updates the parameters for a single layer, whereas in CoMe-sp, layers that are not involved in the merging process during layer pruning (i.e., layers between merged layers) are kept frozen but still included in the training computation. This makes it challenging to directly compare the computational overhead of CoMe-sp and CoMe-mp. In Appendix D, we analyze the differences between CoMe-mp and CoMe-sp and provide a rough comparison of their computational costs. In Table 7, we further compare the number of tokens used by CoMe and single-layer updating methods. Specifically, the computational cost of CoMe-mp is 74% that of LLM-Streamline and 50% that of FuseGPT. Under our default experimental settings, for LLaMa2-7B pruned with CoMe, CoMe-sp incurs over 7 times the computational cost of CoMe-mp in terms of parameter updates and forward passes, but requires fewer training tokens (about 70% of CoMe-mp). Overall, CoMe-sp requires more than 5 times the computational cost of CoMe-mp, but achieves better performance (see Table 1). Compared to full fine-tuning, the computational cost of CoMe-mp is approximately $1/M$, where $M$ is the number of layers in the model.

[Weakness 4] Theoretical Basis of Concatenation-Based Merging

The theoretical foundation for parameter retention in CoMe is based on two main aspects, as discussed in Section 3 of our manuscript.

First, the DLP method demonstrates that pruning any single adjacent layer results in minimal performance degradation, which is elaborated in Line 127 (Core Issue 1). This observation led us to consider that while some important feature-processing modules may be discarded during pruning, the model’s intermediate representations can still be effectively processed by subsequent layers even after one layer is removed.

Second, the WSLP method, which blindly merges the weights of adjacent layers, tends to result in over-smoothed parameter distributions. This issue is discussed in Line 142 (Core Issue 2). To address this, we decompose each layer into finer-grained submodules, such as channels in the FFN structure and heads in the MHA structure. These submodules are characterized by the property that adding or removing some of them does not affect the input and output of the remaining modules. Building on this insight, we split layers into feature-mapping submodules and concatenate the submodules from adjacent layers. In this way, the input and output domain distributions of each submodule remain unchanged, and the more important submodules are preserved. This forms the basis of the “Layer as Puzzle Pieces” concept in our title, highlighting that, compared to existing layer pruning methods (DLP and WSLP), CoMe provides a novel framework for layer merging. Our approach addresses layer pruning at a finer granularity and yields a pruned model structure that is more hardware-friendly and easier to deploy.

We present ablation studies comparing WSLP and various CoMe merging schemes in Table 2, with detailed analyses provided in Section 5.4 and Appendix E. These results demonstrate that our method effectively resolves the issue of mismatched weight mapping between different layers in WSLP. Furthermore, the performance improvements over DLP indicate that CoMe is more effective than both WSLP and DLP.

[Question 3] Generalization of CoMe

In principle, CoMe can be applied to other model architectures. As discussed in our response to [Weakness 4], the CoMe approach decomposes each layer into submodules whose inputs and outputs are independent of one another, and then merges these submodules. Therefore, as long as a model architecture possesses such modularity, CoMe can be adapted for model merging, layer merging, or the fusion of other decomposable structures.

For example, with minor modifications, CoMe can be extended to convolutional neural networks (CNNs) by pruning layers through the concatenation of convolutional kernels from adjacent layers.

We hope these responses can address your concerns. If you have further questions, we would be happy to provide additional clarification.

Thank you for taking the time to review our manuscript. We greatly appreciate your valuable feedback.

Below, we first address [Question 2] and [Question 4] to set the stage for reiterating our method and resolving the concerns you mentioned under weaknesses. We then explain the meaning of the "Hierarchical" Distillation Strategy and address the remaining minor issues.

[Question 2] Clarification on "Adjacent Layers"

In our manuscript, the term “adjacent layers” refers to consecutive layers within the model that are directly connected through input-output relationships. This is a widely used term in academic literature. For example, if the output of layer $l$ serves as the input to layer $l+1$, and the output of layer $l+1$ serves as the input to layer $l+2$, then layers $l$ and $l+1$ are considered “adjacent layers.” Similarly, layers $l$, $l+1$, and $l+2$ (or more consecutive layers) are referred to as “a group of adjacent layers.” In Figure 5, $W^{(l)}$ and $W^{(l+1)}$ denote the weight matrices with the same function name (e.g., up_proj) in adjacent layers. Specifically, $W^{(l)}$ corresponds to the up_proj weight in layer $l$, while $W^{(l+1)}$ corresponds to the up_proj weight in layer $l+1$. Figure 5 illustrates only the fusion method. Depending on the type of matrices, the fusion needs to be performed along different channels (either rows or columns).

To reduce potential misunderstandings, we will revise the caption of Figure 5 in future versions to: “Concatenation-based weight merge process between weight matrices with the same function name in adjacent layers. $W^{(l)}$ and $W^{(l+1)}$ denote the weight matrices with the same function name in two consecutive layers.”

[Question 4] Explanation of Hyperparameter $m$

The hyperparameter $m$ in Figure 1 can be any value less than or equal to the total number of layers in the model. CoMe is capable of merging an arbitrary number ($m$) of layers simultaneously. In Figure 9, we conduct an ablation study on $m$, and the results show that merging too many layers at once degrades the performance of CoMe. Therefore, as stated in the experimental details (line 582), we default to merging two layers at a time.

To reduce potential misunderstandings, we will revise the manuscript as follows: At the first mention of $m$ (line 207), we will add the clarification "$m>=2$ and less than or equal to the total number of layers in the model.” In the implementation details (line 269), we will specify "During the layer pruning process, the merging layers number $m$ in each iteration is set to 2.”

[Weakness 1]

We thank the reviewer for pointing out this issue. In future versions of the manuscript, we will include a more detailed theoretical analysis in the Appendix to further clarify our method.

* Explanation of Channel alignment

When designing CoMe, we carefully considered the challenge of channel alignment in the layer. Below, we clarify the intrinsic rationale of the CoMe and explain why it does not lead to a mismatch between the original and the new input. We decompose the transformer into two main components (Appendix C): the FFN and MHA. In the FFN, we define the input as $X \in \mathbb{R}^{d}$, the intermediate feature as $I \in \mathbb{R}^{r}$, and the output as $O \in \mathbb{R}^{d}$. The weight matrices $W^{U} \in \mathbb{R}^{d \times r}$, $W^{G} \in \mathbb{R}^{d \times r}$, and $W^{D} \in \mathbb{R}^{r \times d}$ correspond to up_proj, gate_proj, and down_proj, respectively. Thus, the FFN can be formulated as:

$I_{[i]} = \sigma(W^{G}\_{[i,:]} X) (W^{U}\_{[i,:]} X)$ $O = \sum W^{D}\_{[:,i]} I\_{[i]} + X$

From the above equations, it is evident that permuting the channel indices ($i$) of the intermediate feature $I$ does not affect the channel indices of the input $X$ or the output $O$. It only changes the internal channel order within the FFN. When we concatenate the intermediate features from different layers at the corresponding channel indices, we are simply reorganizing the submodules responsible for feature processing within the FFN. Regarding channel mapping in the model, we strictly follow the detailed analysis of channel alignment provided in LLM-Pruner [1] (Line 612). Therefore, we partition and concatenate the channels in the FFN according to the intermediate size dimension.

For the same reason, we believe that the WSLP, which directly fuses weights, may result in incorrect weight fusion, as it does not ensure channel correspondence across different layers, potentially leading to smoothed weight distributions.

Similarly, for the MHA structure, we partition channels based on the input dimension of O_proj. Since the head structure in MHA is inherently indivisible, we further treat each head (comprising $W^Q$, $W^K$, and $W^V$) as a “channel group.” Changing the order of heads or the input order of O_proj does not affect the overall input or output of the MHA.

* Theoretical Basis of Concatenation-Based Merging

Layer pruning achieves higher efficiency by reducing the number of layers, which inevitably leads to input space shifts between layers. This issue is common to DLP, WSLP. The key lies in mitigating such misalignments.

The core idea of CoMe is to divide each layer into finer-grained functional blocks. From the perspective of the FFN module, after partitioning parameters into more granular channels, each channel can be regarded as a small expert. Fusing two adjacent layers is thus equivalent to selecting and retaining the most important experts. Several recent works [2,3] have adopted similar perspectives to analyze the FFN (or experts). Likewise, in the MHA structure, each head can naturally be viewed as an independent feature processor.

Furthermore, we employ a Channel Sensitivity Metric to select which channels to retain or discard based on their contribution to the output. Channels with larger products of activation and weight magnitude are considered to have a greater impact, which avoids the need to compute gradients. This approach is consistent with the perspectives of several model sparsification methods [4,5].

[Question 3] Definition of "Hierarchical" Distillation

We provide a detailed description of the Hierarchical Distillation Strategy in Section 4.3 of our manuscript. Lines 238–245 specifically explain how the CoMe-mp process is transformed into CoMe-sp by modifying the training approach and loss function. Our hierarchical distillation strategy leverages layer correspondences $\mathcal{P} = \\{\\{a_1, b_1\\}, ..., \\{a_N, b_N\\}\\}$ established during pruning, where $a_i$ and $b_i$ are indices of corresponding layers in the teacher and student models, respectively. For example, if $\mathcal{P} = \\{\\{1, 3\\}, \\{2, 6\\}, \\{5, 8\\}\\}$, then the output of the 1st layer in the student is aligned with the output of the 3rd layer in the teacher, the 2nd layer in the student is aligned with the 6th layer in the teacher, and the 5th layer in the student is aligned with the 8th layer in the teacher. The loss function (Equation 8 in the manuscript): $\mathcal{L}\_{\text{KL-sp}} = \frac{1}{|\mathcal{P}|} \sum\_{\{a,b\} \in \mathcal{P}} \mathbb{E}\_{\mathcal{D}\_{\text{train}}} \left[ D\_{\text{KL}} \left( \sigma(H^{(t,a)}) \| \sigma(H^{(s,b)}) \right) \right]$ enforces alignment of layer-wise features across all hierarchical levels, rather than only the final logits.

Therefore, both CoMe-mp and CoMe-sp distill knowledge from the original model to restore the performance of the pruned model by aligning the outputs at the “layer" level. For this reason, our approach is referred to as a “Hierarchical Distillation Strategy.”

[Question 1] Format of Figures

Our manuscript was prepared using Overleaf, with all images provided in .pdf format. Nevertheless, following your helpful comments regarding the quality of our figures, we also noticed that some images appear blurry when zoomed in. We will provide clearer vector images in future versions of the manuscript.

Reference

[1] Ma, Xinyin, Gongfan Fang, and Xinchao Wang. "Llm-pruner: On the structural pruning of large language models." Advances in neural information processing systems 36 (2023): 21702-21720.

[2] Lo, Ka Man, et al. "A Closer Look into Mixture-of-Experts in Large Language Models." Findings of the Association for Computational Linguistics: NAACL 2025. 2025.

[3] Zhu, Tong, et al. "LLaMA-MoE: Building Mixture-of-Experts from LLaMA with Continual Pre-Training." Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

[4] Sun, Mingjie, et al. "A Simple and Effective Pruning Approach for Large Language Models." The Twelfth International Conference on Learning Representations, 2024.

[5] Xie, Yanyue, et al. "Moe-pruner: Pruning mixture-of-experts large language model using the hints from its router." arXiv preprint arXiv:2410.12013 (2024).

Thank you again for your valuable feedback, and we hope our responses address your concerns. If you have further questions, we would be happy to provide additional clarification.

Thank you for taking the time to review our manuscript. We greatly appreciate your valuable feedback. Below, we provide our responses to your comments:

[Weakness 1 & Question 1] Benchmark Diversity and Evaluation Scope in Layer Pruning

Thank you very much for your thoughtful suggestions. We agree that evaluating on open-ended generation and task generalization benchmarks would provide a more comprehensive assessment of model performance. However, in the domain of layer pruning, it is standard practice to use pretrained models as the base for pruning. As the number of layers is reduced, the model’s feature representations are often quickly degraded, making fine-grained evaluations such as open-ended generation, contextual coherence, or generalization tests less meaningful. The resulting scores tend to approach random chance. Therefore, in the layer pruning literature, two main evaluation schemes are commonly adopted. The first is measuring PPL on datasets [1,2] such as WikiText2 and C4, which reflects the coarse-grained language modeling capability of the pruned model. The second is to use a range of multiple-choice benchmarks (e.g., SuperGLUE, ARC, MMLU) [3], which serve to assess the model’s natural language understanding ability after pruning. For a fair comparison with prior work, we report both types of evaluation results in our paper.

[Weakness 2 & Question 2] Rationale for Channel Sensitivity Metric

Our choice of the channel sensitivity metric is motivated by prior work and empirical analysis. Specifically, we refer to Wanda [4] and MOE-PRUNER [5], which focus on weight-level sparsity and the importance of individual weights. In contrast, CoMe assesses channel importance by considering both weight magnitude and the influence of input activations on layer output, inspired by findings that large activations tend to have a greater impact on model output [6]. We provide relevant explanations in Section 4.1 of our manuscript.

Moreover, we chose not to adopt gradient-based (first-order or second-order) importance metrics for two main reasons: (1) Methods such as Magnitude and Taylor, which rely on gradients, often result in the removal of initial or final layers. To ensure fair comparison, we prohibit pruning of the first and last several layers for these methods (see Appendix A), highlighting the suboptimality of gradient-based approaches for layer pruning. (2) Gradient-based methods require additional storage for gradient information during pruning, which increases computational and memory overhead, contrary to our goal of lightweight pruning and training under resource constraints.

[Regarding Weakness 3 & Question 3] Selection and Impact of the Channel Retention Ratio

We propose an adaptive parameter fusion ratio formula (Equation 5), which automatically adjusts the fusion ratio between adjacent layers based on their BI scores. Layers with higher BI scores are considered more important and thus retain more parameters. Within the same layer, the FFN and MHA structures share the same fusion ratio. However, our ablation studies in Figures 10 and 14 reveal that the optimal fusion ratios differ between structures (FFN and MHA). We have addressed this issue in the limitations section (line 357) and provided further analysis on page 8.

The hyperparameter $p$ is used to control the scaling relationship for parameter retention between more important and less important adjacent layers. A larger $p$ means that more important layers retain a higher proportion of parameters. We conduct an ablation study on $p$ in Figure 7 and provide analysis in lines 312–319. Our results show that as $p$ approaches infinity, the pruned model’s performance degrades and converges to that of the DLP method. In other words, an inappropriate choice of $p$ can lead to performance as poor as DLP. Since the optimal value of $p$ varies across different models, we use hyperparameter search to determine the most suitable $p$ for each model. Currently, we do not have a universal setting for $p$ across different models, which is why we have highlighted this limitation in our manuscript.

In future work, we plan to investigate more effective and adaptive fusion ratios tailored to different structures, layers, and models.

References:

[1] Song, Jiwon, et al. "SLEB: Streamlining LLMs through Redundancy Verification and Elimination of Transformer Blocks." Forty-first International Conference on Machine Learning.

[2] Pei, Zehua, et al. "FuseGPT: Learnable Layers Fusion of Generative Pre-trained Transformers." arXiv preprint arXiv:2411.14507 (2024).

[3] Men, Xin, et al. "ShortGPT: Layers in Large Language Models are More Redundant Than You Expect." Findings of the Association for Computational Linguistics: ACL 2025, edited by Wanxiang Che et al., Association for Computational Linguistics, 2025, pp. 20192–20204.

[4] Sun, Mingjie, et al. "A Simple and Effective Pruning Approach for Large Language Models." The Twelfth International Conference on Learning Representations, 2024.

[5] Xie, Yanyue, et al. "Moe-pruner: Pruning mixture-of-experts large language model using the hints from its router." arXiv preprint arXiv:2410.12013 (2024).

[6] Sun, Mingjie, et al. "Massive Activations in Large Language Models." First Conference on Language Modeling, 2024.

We hope these responses can address your concerns. If you have further questions, we would be happy to provide additional clarification.

We sincerely appreciate your positive feedback and for raising the score.

To further address your concerns about the effectiveness of the heuristic parameter fusion formula, we drew inspiration from SLEB and LaCo to develop a posterior approach for determining fusion weights across different models and layers.

Implementation details:

The rest of the CoMe computation process remains unchanged. We replace the heuristic fusion ratio scheme with the posterior fusion ratio. Specifically, when merging two layers, we examine fusion ratios from $0:1,\: 0.05:0.95,\: ...,\: 1:0$ and compute the cumulative loss of the pruned model on the calibration dataset for each ratio. We then select the fusion ratio that results in the lowest cumulative loss to fuse layers.

Analysis:

As shown in the table below, “Heuristic” refers to the original heuristic fusion ratio calculation method in CoMe, while “Posterior” indicates our newly proposed posterior approach. Notably, the “Posterior” method does not require any additional hyperparameter tuning for layer fusion. Instead, it adaptively chooses the optimal fusion ratio between layers for different models. During the fusion of two layers, the “Posterior” method evaluates the loss on the calibration dataset 21 times, which is more computationally efficient than SLEB (where the number of loss evaluations of SLEB depends on the remaining layers, typically more than 22. For example, after pruning 30%, LLaMA2-7b retains 22 layers, and LLaMA2-13b retains 28 layers).

The “Posterior” method further improves CoMe’s performance, reducing the PPL of pruned models on the Wiki-2 and C4, and achieves performance comparable to the “Heuristic” method on benchmarks. The result supports our observation that the optimal fusion ratios vary across different layers. Meanwhile, the “Heuristic” method, which links the fusion ratio to the BI score, incurs only a limited performance loss. Both the “Posterior” and “Heuristic” methods outperform existing SOTA and demonstrate greater stability.

It is worth noting that there are multiple approaches to calculating layer fusion ratios, which do not affect the novelty of our work. Improved strategies could further enhance CoMe’s performance. More importantly, the main innovation of CoMe lies in its substructure concatenation-based fusion method.

Thank you again for your valuable comments and suggestions. We hope this additional explanation further addresses your concerns, and we are happy to discuss any further questions you may have.

For details, please see Figure 6 and Tables 8,9,10 (space is limited).