Token-Scaled Logit Distillation for Ternary Weight Generative Language Models

Generative Language Models (GLMs) have shown impressive performance in tasks such as text generation, understanding, and reasoning. However, the large model size poses challenges for practical deployment. To solve this problem, Quantization-Aware Training (QAT) has become increasingly popular. However, current QAT methods for generative models have resulted in a noticeable loss of accuracy. To counteract this issue, we propose a novel knowledge distillation method specifically designed for GLMs. Our method, called token-scaled logit distillation, prevents overfitting and provides superior learning from the teacher model and ground truth. This research marks the first evaluation of ternary weight quantization-aware training of large-scale GLMs with less than 1.0 degradation in perplexity and achieves enhanced accuracy in tasks like common-sense QA and arithmetic reasoning as well as natural language understanding. Our code is available at https://github.com/aiha-lab/TSLD.

Paper

References (50)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer cozw5/10 · confidence 4/52023-06-12

Summary

The work proposes modifications for Quantization-Aware Training (QAT) that utilize knowledge distillation and takes into account token-scaled logit distillation for preventing overfitting. They show the motivation for their method through a deep analysis and present the empirical merits of this method via language modeling and mathematical reasoning tasks.

Strengths

The proposed Token-Scaled Logit Distillation (TSLD) method as well as the analysis that comes before is novel and interesting. I like the fact that the authors detected a problem with standard KD and proposed a nice way to mitigate it which replaces the uniform weighting over the tokens. I’m not sure though about the time complexity it takes to compute this loss, as well as the quadratic nature of this complexity given larger sequences, due to the nested summation. I’d suggest discussing this in a future revision. Generally the paper is well written and motivated. The authors present a series of analysis that showcase the importance of utilizing their method during quantized training.

Weaknesses

Minor comments: * L17: state that ALPACA was fine-tuned from the LLaMA 7B model. * The notations in Section 3.2 do not follow standard mathematical rules (e.g., matrix is denoted by two letters, a vector letter is capitalized, etc.) and prior works. Please check previous works describing the transformer modules and change the equations accordingly. * L119: Masked self-attention is the wrong wording here. You probably meant causal attention. * Section 3.3 is written in a vague form and too informally. Consider adding a mathematical proof for the error accumulation in causal models. Major comments * What about long-range transformer models? How will the suggested method influence them? * In L50 you state one example for a task (arithmetic reasoning). What about other tasks? There is a missing analysis/discussion why would this task specifically benefit from the proposed method, while other tasks don’t. * I know that the paper was submitted before the QLORA was published, but in case this paper is accepted, I would expect the authors to add a comparison or at least a discussion about the merits of the proposed method against QLORA and other quantized low-rank adapter techniques. I do not consider this point in the final score calculation of this paper. * An inaccurate point in Section 3.3 second part (Necessity of Ground Truth Loss) is the fact that decoder-only models use the entire vocabulary for prediction, unlike encoder models that use only the number of classes. This is also not correct, since one may consider and optimize a few tokens instead of the entire vocabulary, and thus there is a slight difference between encoders and decoders in that case for classification tasks. * Empirical results over two datasets is not enough for me, and I wonder how the performance will change given more language modeling and other NLU tasks. * There is an empirical evaluation regarding inference speed, but I wonder how fast is training time. This is crucial, since pre-training for example is costly anyway. Missing references * Jin, Jing, et al. "KDLSQ-BERT: A quantized bert combining knowledge distillation with learned step size quantization." arXiv preprint arXiv:2101.05938 (2021).

Questions

Please respond to the concerns stated under the Weaknesses section.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

I don't see any limitations section, please review and add this. One limitation can be for example the extra training time required by this method.

Reviewer QKmL6/10 · confidence 4/52023-06-30

Summary

This paper proposes to improve the quantization of generative language models via an adaptively scaled logit distillation method. With an intuitive illustration in Figure 1 and validation results with attention heatmap discrepancy, the paper identified a cumulative quantization error introduced by the masked attention, i.e., later tokens attend to more tokens before, leading to a larger quantization error. Furthermore, the paper investigated the characteristics of Layer-to-Layer (L2L) distillation (aligning output activations and attention heatmaps) and the logit distillation from teacher models. Results show these two methods differ from the inner representations and min-max dynamic ranges. This result motivates a proposal of Token-Scaled Logit Distillation, where tokens with higher confidence of teacher predictions are down-weighted in the distillation to avoid overfitting. Experimental results on GPT-2 and OPT-series show that the proposal effectively maintains the generation quality (lower perplexity compared with baseline methods) and reasoning ability of the original full precision model (better accuracy on GSM7k).

Strengths

- Exploring the quantization techniques for billion-level language models is a vital problem. - The finding of cumulative quantization errors of mask attention in Transformer models is interesting.

Weaknesses

The paper can benefit from addressing the following weaknesses: - Paper Organization: The current paper lacks clarity in its organization, making it challenging to follow the logical flow between sections. Improving the coherence and connection between sections would greatly enhance the reading experience. - Integration of Challenges: The paper identifies two challenges in quantizing LLMs, namely the cumulative quantization errors of masked attention and the necessity of ground-truth loss. However, it would be more effective to investigate these challenges independently. Additionally, the proposed method seems less relevant to the first finding, which, in my opinion, is the most crucial problem in the quantization process. - Lack of Clear Motivation: The motivation behind the proposed dynamic weighting method, aimed at addressing the potential overfitting problem, is not adequately explained. The evidence presented in Figure 3(b) is insufficient to support the claim that "(high confidence) logit distillation might mirror the effects of ground truth loss, thereby contributing to the observed overfitting." It would be beneficial to explore whether using ground truth alone leads to overfitting, as well as whether removing distillation of high-confidence tokens can mitigate the overfitting problem. Additionally, justifying the up-weighting of low-confidence tokens, which exhibit high cross-entropy loss and could potentially represent incorrect predictions from the teacher models, is crucial to convince readers of the proposal's validity.

Questions

- In Figure 2(b) Left, why is the bottom left corner (where tokens attend fewer tokens ?) of the heatmap also bright? - In Figure 2(b) Right, what is the reason for deeper layers exhibiting lower attention MSE error in the decoder? - Original QuantGPT incorporates Token-level Contrastive Distillation and Module-dependent Dynamic Scaling, did you solely utilize logit distillation?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

N/A

Reviewer aiZD6/10 · confidence 4/52023-07-06

Summary

This paper proposed the token-scaled logit distillation method to counteract accuracy loss caused by using Quantization-Aware Training(QAT) methods for generative language models . (1) The author proposed the logit distillation method to overcome the limitation in token prediction recovery by modifying the intermediate representation. (2) the author proposed an adaptive KD method that adjusts Logit KD based on token confidence to prevent the overfitting problem caused by the combination of Logit KD and ground truth loss in QAT.

Strengths

1、The idea of adjusting Logit KD based on token confidence to prevent the overfitting problem is novel. 2、The method used in this paper is generalized and applicable to models of different sizes and different types of text generation tasks, including language modeling and reasoning. 3、The experiments involves multiple LLMs with various model scale (0.1B,2.7B,6.7B), which makes the main experiments solid and convincing.

Weaknesses

1、The FP 32 baseline used in the comparison experiment is not explained. And the experiment only uses one baseline, the proof is not sufficient. 2、 In addition, the experiment selected one dataset for each tasks for validation, and the number of datasets is small. We recommended to add SVAMP and CSQA datasets for inference tasks, and SGD datasets for text generation tasks. 3、The experiments include only main results (comparing with baselines) and the ablation studies. It would be better to conduct more detailed analyses or qualitative analysis to verify the proposed methods (e.g. whether the issues proposed in the motivation are fixed).

Questions

None

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

None

Reviewer kT5b7/10 · confidence 5/52023-07-06

Summary

This research paper focuses on Generative Language Models (GLMs) and the challenges of their deployment due to their large model sizes. The authors introduce a novel approach to tackle the challenges faced by Quantization-Aware Training (QAT) in fine-tuned GLMs. **Problem**: GLMs, despite their impressive performance in text generation, understanding, and reasoning, face deployment challenges due to their large model sizes. Quantization techniques like Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT) have been used for model compression but result in significant decreases in accuracy. **Approach**: The authors propose a new approach, Token-Scaled Logit Distillation (TSLD), to enhance QAT for sub-4bit QAT inference. TSLD is a knowledge distillation technique specifically designed for GLMs. It counters the issues of quantization errors and overfitting, prevalent in existing QAT methods. The method is based on token-scaled logit distillation, which prevents overfitting and allows better learning from the teacher model and ground truth. **Results**: When TSLD was applied to a variety of GLMs, including a 6.7 billion model for the first time, it achieved comparable, if not superior, performance in text generation on ternary and 4-bit inference. On an arithmetic reasoning task, TSLD even achieved task accuracy at least on par, if not better, with other methods. Experiments were performed on two tasks: language modeling (using the PTB dataset) and a reasoning-based text-generation task (using the GSM8K dataset). The models were trained using the HuggingFace fine-tuning repository, with varying learning rates and sequence lengths depending on the model's size. They also implemented a QAT-based KD framework using PyTorch's Pipe API for pipeline parallelism. Results showed that TSLD achieved the lowest Perplexity (PPL), signifying it performed the best in both language modeling and reasoning tasks, outperforming other methods such as Post-Training Quantization (PTQ), QAT KD, and others. TSLD's performance allowed QAT models to match full-precision performance across different model capacities. Regarding sub-4 weight quantization, the researchers found that L2L KD couldn't run on larger GLMs due to memory constraints. TSLD, however, outperformed other QAT approaches across all model sizes, maintaining PPL degradation of no more than 1.0 from the baseline. For the reasoning task, TSLD was shown to effectively balance language modeling and reasoning performance, achieving excellent reasoning accuracy while maintaining competitive perplexity. Ablation study also suggested that sub-4bit quantization with TSLD considerably speeds up the inference of large models while maintaining competitive model quality. **Conclusion**: The successful results underline the potential of the proposed TSLD method in facilitating the deployment of ultra-low precision GLMs. The method allows for the practical deployment of large-scale GLMs with less than 1.0 degradation in perplexity and no loss of accuracy in a reasoning task.

Strengths

**Innovative Method**: The paper introduces the Token-wise Scaling for Low-bit Distillation (TSLD) method for Quantization-Aware Training (QAT), a novel approach that seems to substantially improve the performance of low-precision models. And this a very early work about Ternary quantization and sub-4 quantization of GLMs. **Detailed Implementation information**: The paper provides clear and comprehensive details on the experimental setup, including task and model selection, training settings, and implementation settings. This thoroughness makes the results more credible and the experiments reproducible by other researchers. **Clean and accurate formulation**: the formulation process is clean and accurate, to the best of my knowledge. **Practical Results**: The paper demonstrates practical advantages of TSLD, such as matching full-precision model performance across various model sizes and maintaining low perplexity degradation with sub-4 weight quantization. These findings suggest that the TSLD method is not only theoretically promising but also practically beneficial, paving the way for low-precision inference in real-world applications.

Weaknesses

Lack of Broader Evaluation: The paper focuses primarily on GLMs like GPT-2 and OPT for the experiments. This limited scope might not fully illustrate how the TSLD method performs with other types of GLM models. Specifically, if the evaluation tasks can be more various and abundant, it would be very helpful Specificity of Tasks: While the paper does test the proposed method on different tasks, it mainly focuses on the PTB language modeling task and GSM8K reasoning task. An evaluation of the method's effectiveness on a broader array of tasks, or in specific real-world applications, could have added to the robustness of the study. (My concern comes from: perplexity is very insensitive, and arithmetic reasoning dataset GSM8K is not large. So "less than 1.0 perplexity degradation and preserving arithmetic reasoning task accuracy" is not that convincing. it is really recommended to have more evaluation on other tasks)

Questions

1. About the TSLD methods details "Token-Scaled Logit Distillation (TSLD). Based on investigations into the probabilistic relation of token predictions and overfitting in QAT, we propose an adaptive KD method that adjusts Logit KD based on token confidence. This approach utilizes the phenomenon of confidence disparity in token predictions from the teacher model. Our method, called TSLD, de-emphasizes Logit KD for high-confidence tokens to prevent overfitting in QAT while emphasizing logit distillation for low-confidence tokens with high entropy probability distributions." I am eager to know, how much additional computing are required in the TSLD process, such as incorporating the token confidence into the distillation process? 2. Without specific designed hardware, how much advantage could TSLD powered Ternary Weight quantization have, over the 4-bit quantization of LLMs? With what kind of designed hardware, how much advantage could TSLD powered Ternary Weight quantization have, over the 4-bit quantization of LLMs? The clarity regarding this point will illustrate the practical significance of this work.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

**Hardware Restrictions**: The TSLD method uses simple ternary weight quantization, which is designed to be hardware-friendly. However, this means that the effectiveness of TSLD may be largely dependent on the specific hardware it's being implemented on. Some hardware configurations might not support ternary or lower precision computations efficiently, which could limit the usefulness or performance of TSLD in such environments.

Reviewer EGP45/10 · confidence 4/52023-07-12

Summary

The article introduced a new approach for LLM Quantization Aware Training (QAT).

Strengths

The paper is well written and the proposed idea is described clearly.

Weaknesses

I have some concerns about the evaluation in the paper. I First, the number of tasks used for accuracy/perplexity comparison between float and the proposed method is too small. The authors only evaluate the proposed method on two small datasets. I would encourage the authors to include more datasets in the evaluation, such as those Q&A tasks and SuperGlue tests datasets listed in the GPT-3 paper. This would help to provide more evidence to support the claims made in the paper: the proposed method counteract the issues of accuracy loss. Second, the authors only report accuracy and perplexity results for the int2 data type. It would be great for the readers to see the accuracy and perplexity numbers for all data types (int8, int4, and int2). The authors do provide latency benchmarks for the int8 data type, but they do not report any quality results. I would encourage the authors to report accuracy and perplexity results for all data types so that the readers can have a better understanding of the trade-offs between accuracy, latency, and memory usage. Here are some specific suggestions for how the authors could improve the evaluation: 1. Include more tasks in the evaluation, such as those Q&A tasks and SuperGlue tests datasets listed in the GPT-3 paper. 2. Report accuracy and perplexity results for all data types (int8, int4, and int2). 3. Conduct evaluation to assess the quality of instruction tuning dataets.

Questions

1. It's not clear to me whether table 1 shows the training or the test perplexity for the PTB dataset. 2. Why only results from W2A32 are shown in table 2? Can you please also include W4A32 ones?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

Need more evaluation results to justify the contributions.

Reviewer cozw2023-08-13

I want to thank the authors for their response. Unfortunately I can't see the tables with the new results. In addition, while perplexity is fine, I'd love to see the simple step of actual results on the downstream tasks.

Authorsrebuttal2023-08-13

Thank you very much for your valuable feedback. Respectfully, starting from this rebuttal, we are allowed to attach a **1-page PDF** in a **global response** section. You can kindly refer to the table in the attached PDF. I regret that I didn't mention this in the rebuttal response section. To address this, I will include the results for the NLU and CSQA downstream task accuracy results in this comment. (more detailed results can be found in the attached PDF of the global response) **Table R1**: Accuracy results for the OPT-1.3B model fine-tuned on natural language understanding GLUE tasks using different QAT-KD methods on W4A16 and W2A16. ACC metrics are averaged based on five random seed tests. | Precision | QAT KD Method | CoLA ACC (↑) | MRPC ACC (↑) | SST-2 ACC (↑) | RTE ACC (↑) | |-----------|---------------|--------------|--------------|---------------|-------------| | OPT-1.3B FP16 | - | 61.03 | 81.92 | 94.26 | 76.53 | | W4A16 PTQ | OPTQ-CW | 54.61 | 80.14 | 95.07 | 56.32 | | - | AWQ-CW | 13.63 | 66.42 | 94.26 | 54.51 | | W4A16 QAT | Logit [20] | 50.76 | 81.94 | 93.57 | 75.23 | | - | GT+Logit | 54.07 | 83.17 | 93.34 | 75.31 | | - | **TSLD** | **56.33** | **83.33** | **94.05** | **75.97** | | W2A16 QAT | Logit [20] | 48.72 | 81.62 | 93.08 | 74.15 | | - | GT+Logit | 50.10 | 82.10 | 92.77 | 73.79 | | - | **TSLD** | **54.47** | **82.20** | **93.92** | **75.31** | **Table R3**: Accuracy (ACC_norm) results of the OPT-2.7B model fine-tuned on the CSQA (Commonsense QA) with different W2A16 QAT-KD methods. | QAT KD (W2A16) | PIQA (↑) | OpenbookQA (↑) | ARC_easy (↑) | ARC_challenge (↑) | GSM8K (↑) | |:----------------|:--------------|:-------------------|:-----------------|:-----------------------|:--------------| | OPT-2.7B FP16 | 76.71 | 49.60 | 66.12 | 37.20 | 20.39 | | Logit [20] | 74.32 | 45.40 | 58.92 | 31.91 | 20.02 | | GT+Logit | 74.97 | 46.20 | 58.84 | 32.16 | 19.56 | | **TSLD** | **75.62** | **46.81** | **59.39** | **33.45** | **20.24** | **Table R4**: Perplexity and accuracy results of W2A16 QAT-KD on PTB task and text generation-based reasoning tasks for various GLM models (GPT-Neo-1.3B, LLaMA-7B) | Model (W2A16) | GPT-Neo-1.3B PPL (↓) | OPT-6.7B GSM8K (↑) | LLaMA-7B PTB (↓) | LLaMA-7B GSM8K (↑) | |:---------------|:---------------------|:-------------------|:-----------------|:-------------------| | FP16 | 17.62 | 22.52 | 9.10 | 32.02 | | Logit [20] | 21.01 | 21.08 | 10.82 | 25.47 | | **TSLD** | **19.27** | **24.49** | **10.79** | **26.23** | In the three tables presented, it is evident that the TSLD method not only improves perplexity but also significantly enhances downstream accuracy. This can be observed across NLU (Table R1), and reasoning (CSQA) tasks(Table R3/4). If there are further points to discuss or valuable feedback to share, we would always be grateful and welcoming. We would greatly appreciate it if you could consider the various downstream task accuracy results we have prepared during the rebuttal.

Reviewer kT5b2023-08-16

Thanks for addressing my questions

Thanks for addressing my questions. I keep the original rating as 7. If the code or even only model could be provided through an anonymous link, the reproducibility of this work will improve.

Authorsrebuttal2023-08-16

Thank you for your feedback. We will ensure that your insightful and valuable comments are incorporated into our final manuscript. Furthermore, based on your comments, we have prepared an anonymized code repository. The refined version will be officially released alongside our final manuscript. https://anonymous.4open.science/r/tsld_anonymized-D6D1 Once again, we deeply appreciate the time and effort you've dedicated to feedback on our manuscript.

Reviewer kT5b2023-08-17

Thanks very much for the reply

Thanks very much for the timely reply! The anonymized code link makes me more confident of the current rating at least 7. Could you share some model weights also in anonymized link if allowed? After examine the result, I might increase my rating further. Best regards

Authorsrebuttal2023-08-18

Thanks very much for your further interest

We greatly appreciate your time in reviewing the code we provided and thank you for your further interest in our approach. To share the QAT model weight where our method is applied, we have prepared an anonymous link as follows: link: https://we.tl/t-NH1wysJtqW (Accept cookies/terms of service required before downloading, expires in 7 days) **About Model Weight** Due to the capacity limits of the anonymous file link, we are sharing the weight files of the two models where **two QAT-KD methodologies (Logit+GT, TSLD)** have been applied to the **OPT-125m PTB fine-tuned model** (Table1). In the future, we plan to share model weights (through huggingface) for verification on various billion-scale fine-tuned models. **Guideline for Model Weight Examination** For the convenience of the reviewer's examination of the model weight, here's how the model performance verification proceeds: - We evaluated the QAT model's PPL performance using the latest version of huggingface (Transformers) language-modeling example code (https://github.com/huggingface/transformers/tree/main/examples/pytorch/language-modeling). We have re-confirmed that the evaluation results in the official huggingface environment match the outcomes from our experimental setup. (Logit+GT PPL: 21.91, **TSLD PPL: 18.97**) - Note that the model file we shared corresponds to the 40 Epoch QAT results conducted in the ablation study of Figure R1 (f) in the rebuttal. By increasing the number of epochs, you can obtain the same results presented in Table 1 of the manuscript. - The shared model folder's file configuration is compatible with the **huggingface AutoModel** method. Specifically, by simply placing the names of the two model folders (`opt_ptb_40_2bit_logit_gt`, `opt_ptb_40_2bit_tsld`) into the `--model_name_or_path` argument of the main code, you can easily build ternary weight OPT-125m model. **Details** - The direction of the QAT-KD in this paper involves applying QAT to *task-specific fine-tuned GLM*. Therefore, please note that if the evaluation is conducted on a task that wasn't fine-tuned, the performance might not be satisfactory. - The model weights have per-tensor ternary quantized values stored (ex. [0, 1.7, -1.7, 0, 1.7, 0, -1.7 ... ]), but they aren't actually encoded in the INT2 type. Please note this. We plan to implement the INT2 datatype in the future for the sake of saving actual model size. - In each model folder, there are files named `training_log.json` and `eval_log.json`. These contain logs of training loss, evaluation loss, and epoch-by-epoch evaluation PPL results during the QAT-KD training. You can refer to these to examine how QAT training has proceeded. - `all_results.json` contains the model's last evaluation PPL performance and its best PPL performance. By comparing the two results from the `opt_ptb_40_2bit_logit_gt` model, you can re-confirm instances of overfitting, as seen in the logit + GT QAT. We would like to express our gratitude once again for the interest the reviewer has shown in our method. We hope that the shared model weights and guides will assist in further examination. Should there be any additional queries or difficulties during the verification process, please do not hesitate to leave a comment. We welcome any feedback.

Reviewer QKmL2023-08-16

Thanks for your detailed response and experimental evidence. - Regarding Answer 1 and Answer 2: The connection between cumulative quantization errors and Logit KD looks much more evident to me. - Regarding Answer 3: The decomposition Fig. R1(f) is very helpful, making the overfitting of groud-truth training clearer and the effect of dynamically reweighting. Overall, I think the response addressed my concerns well. I am happy to raise my score to 6, and I hope these discussions and results can be integrated into the revision as well.

Authorsrebuttal2023-08-16

We are grateful to the reviewer for reviewing our responses with materials and providing us with feedback. Your constructive and insightful comments have been invaluable in guiding our efforts to improve the manuscript. We will make sure to incorporate your valuable suggestions into the revised manuscript. Once again, we deeply appreciate your invaluable review.

Area Chair zfMD2023-08-17

Reminder: the author-reviewer discussion is closed on Aug 21st 1pm EDT.

Authorsrebuttal2023-08-18

We deeply appreciate your taking the time to review the additional experimental results and materials we presented in the rebuttal. We will certainly incorporate the valuable comments and discussions provided by the reviewer in the preparation of the revised manuscript.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC