UFO-RL: Uncertainty-Focused Optimization for Efficient Reinforcement Learning Data Selection

Thank you very much for your constructive and insightful review. We agree with the key weaknesses you identified and have conducted a series of new supplementary experiments to address your concerns. We believe these revisions significantly strengthen our paper. Our responses are as follows:

1. Regarding Algorithm Dependency: Focused on GRPO, Applicable to PPO We thank the reviewer for this critical point. We wish to emphasize that while our work is primarily focused on and validated within the GRPO framework, the underlying principle of our method—selecting data from the Zone of Proximal Development —is algorithm-agnostic.

To prove its broader applicability, we additionally tested UFO-RL on the mainstream Proximal Policy Optimization (PPO) algorithm.

• New Experimental Results (Qwen2.5-7B on GSM8K):

  Algorithm	Data Strategy	Accuracy (%)
  PPO	Full Data (100%)	90.48
  PPO	Random (10%)	89.51
  PPO	UFO-RL (10%)	90.77

• Conclusion: This experiment confirms that UFO-RL's effectiveness is not confined to GRPO and generalizes successfully to PPO. We will clarify in the revised manuscript that GRPO is our primary framework while including this PPO validation to demonstrate the generalizability of our approach.

2. Regarding Lack of Clarity in Table 1: We apologize for the confusion regarding Table 1. Your observation is correct, and it highlights a lack of clarity in our presentation, which we will fix.

The core purpose of this table was to illustrate that a significant portion of the data is polarized at the extremes (i.e., consistently wrong or consistently correct). To emphasize this point, the original table intentionally presented only a subset of the data distribution—specifically the percentages for data points near the extremes. This is why the percentages shown did not sum to 100%, which understandably caused confusion.

You correctly noted that for stronger models like Qwen2.5-7B, the (85, 100] bin contains a very large fraction of the data. This actually reinforces our central argument: for capable models, the multi-sampling accuracy metric becomes less useful as it classifies most data as "easy," failing to provide the fine-grained difficulty assessment needed for efficient learning. This is precisely the problem our continuous confidence score is designed to solve.

To resolve this, we will revise Table 1 in the paper to show a complete, granular, and mutually exclusive distribution.

• Proposed New Format for Table 1: | Accuracy Bin | Description | Qwen2.5-7B (%) | Llama3.1-8B (%) | | :--- | :--- | :--- | :--- | | Acc = 0% | Consistently Wrong | 1.36 | 0.90 | | (0, 15%) | Mostly Wrong | 1.85 | 1.50 | | [15%, 85%] | Intermediate | 21.08 | 43.86 | | (85%, 100%)| Mostly Correct | 26.64 | 32.21 | | Acc = 100% | Consistently Correct | 49.07 | 21.53 | | Total | | 100.00 | 100.00 |

This revised format will transparently show the entire data distribution and more effectively highlight the polarization issue that motivates our work.

3. Regarding Fairness of the Computational Cost Analysis: Your query regarding the end-to-end cost analysis is entirely correct. To provide a fully transparent and fair comparison, we have updated the end-to-end cost analysis to cover all Qwen model sizes. As we stated in Appendix A.5 of our original manuscript, all training was conducted on a computing cluster equipped with 8 NVIDIA A100 GPUs. This powerful hardware configuration explains the fast training speeds on smaller data subsets. We will state this configuration more explicitly in the main text of the paper.

• New End-to-End Performance and Cost Comparison (GSM8K, 8x A100 GPU): | Model | Method | Data Selection Time (s) | RL Training Time (s) | Total Time (s) | Total Speedup (vs. Full) | GSM8K Acc (%) | Math500 Acc (%) | | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- | | Qwen2.5-0.5B | UFO-RL (Ours) | 2 | 140 | 142 | 12.8x | 46.27 | 30.60 | | | Full Data | 0 | 1815 | 1815 | 1.0x | 52.66 | 30.80 | | Qwen2.5-1.5B | UFO-RL (Ours) | 6 | 407 | 413 | 13.8x | 76.63 | 55.40 | | | Full Data | 0 | 5694 | 5694 | 1.0x | 76.78 | 53.00 | | Qwen2.5-3B | UFO-RL (Ours) | 11 | 739 | 750 | 13.6x | 84.37 | 67.40 | | | Full Data | 0 | 10224 | 10224 | 1.0x | 83.33 | 66.20 | | Qwen2.5-7B | UFO-RL (Ours) | 22 | 1154 | 1176 | 11.0x | 92.03 | 76.40 | | | Full Data | 0 | 12959 | 12959 | 1.0x | 91.88 | 75.00 |

• Conclusion: The comprehensive data show that this efficiency advantage holds across all model sizes. The one-time screening cost of UFO-RL is negligible compared to the massive savings in RL training time, ultimately achieving a 10-14x end-to-end training speedup across the entire model family while maintaining or even surpassing the performance of full-data training.

Responses to Specific Questions:

1. What is the similarity metric in Table 2? It is the Spearman's rank correlation coefficient. We chose it because we are more concerned with the ranking consistency of data difficulty between the two methods rather than a linear relationship. We will state this explicitly in the revised text.

2. Regarding the claim that "larger models are more robust"? Your observation is correct; the original statement was too absolute. We will revise it to a more nuanced "a tendency for the performance peak to shift to more difficult data," which is more precise and better reflects the trends shown in Figure 1.

3. Why was 10% of the data chosen? This is an excellent question. We conducted an ablation study on the data selection percentage (k) to determine the optimal trade-off between cost and performance.

• New Experimental Results (UFO-RL, Qwen2.5-7B): | Data Percentage (k%) | 5% | 10% | 20% | 30% | 100% (Full) | | :--- | :--- | :--- | :--- | :--- | :--- | | Accuracy (%) | 90.47 | 92.03 | 92.71 | 92.41 | 91.88 |

• Conclusion: Performance peaks around 20%.We chose 10% as our main reported number because it achieves the best trade-off between computational cost (half that of the 20% run) and model performance (over 99% of the peak performance), demonstrating the extreme efficiency of our method.

Summary

Thank you again for your valuable feedback. We believe these new experiments and revisions substantially strengthen the rigor and generalizability of our work and have fully addressed your concerns. We hope to earn your support for our paper.

We sincerely thank you for your invaluable feedback and insightful comments on our paper. Your review is highly constructive and has been instrumental in helping us refine our work. We are encouraged that you recognize the significant advantages of our method in terms of computational efficiency and breadth of evaluation. In response to the concerns and questions you raised, we have conducted further analysis and supplementary experiments, which we address point-by-point below.

Regarding Main Concerns (Weaknesses)

1. On the Theoretical Basis and Heuristic Nature of the Method Reviewer's Comment: "The whole method is based on ZPD hypothesis, which has no theoretical basis. The paper did include experiments to verify the hypothesis, but in Fig.1, red line may overstate the clarity of the trend, the raw data (black dots) shows inconsistencies, especially for smaller or weaker models."

Our Response: We appreciate this insightful point. We wish to clarify that the motivation for our work is grounded in the intrinsic mechanisms and efficiency bottlenecks of Reinforcement Learning for LLMs, particularly algorithms like GRPO.

In RL algorithms such as GRPO, policy optimization relies on multi-sampling from the same prompt and calculating gradients based on the collective reward signal from the resulting generation group. Through this process, we identified a critical inefficiency when the data is either too easy or too difficult:

If a problem is too easy, two problematic scenarios arise. First, if all sampled generations are correct, the uniform reward of 1 provides no reward variance, leading to a vanishing gradient[1]. Second, if the group contains mostly correct answers and only a few incorrect ones, the group-relative optimization will assign a negative reward signal to the correct answers, incorrectly penalizing good generations.

Conversely, if a problem is too difficult, similar issues occur. If all generations are incorrect, the uniform reward of 0 again results in a vanishing gradient[1]. If the group consists of mostly incorrect answers and only a few correct ones, the group-relative optimization will not only heavily reward the rare correct answers but may also assign a positive signal to "less incorrect" answers, causing the model to learn a suboptimal or flawed policy.

In both of these extreme cases, the lack of reward variance or the skewed reward signals cause the learning process to become ineffective or unstable. This observation from algorithmic practice forms the direct motivation for our research: to achieve efficient RL training, it is imperative to select data that can produce a mix of correct and incorrect outcomes.

It is precisely from this algorithm-level requirement that we introduce the Zone of Proximal Development (ZPD) theory. We employ it not as a strict mathematical theorem, but as a powerful conceptual framework to explain and guide how we can systematically identify and select this "intermediate difficulty" (or "fuzzy") data, which provides the most effective learning signals for algorithms like GRPO.

Regarding the volatility in Figure 1, we agree with your observation and consider it expected for smaller models. However, the crucial finding is that a clear non-monotonic trend consistently emerges across all six models. This cross-model consistency strongly demonstrates that focusing on intermediate-difficulty data is a universally effective optimization strategy that directly addresses the vanishing-gradient problem inherent in the RL algorithm itself.

2. On the Novelty and Reliability of the Confidence Score Reviewer's Comment: "Token-level confidence estimation is not a new concept, and it may not serve as a reliable proxy for semantic uncertainty. The similarity between confidence and semantic uncertainty is sensitive to prompt design; however, the paper does not include any experiments analyzing prompt effects on similarity in Table 2."

Our Response: We thank the reviewer for this close examination, which allows us to further clarify our core innovation.

As mentioned, methods like GRPO require identifying "non-extreme" data, but the conventional approach—evaluating data via actual multi-sampling—imposes a prohibitive computational cost, which is the central problem we aim to solve.

We fully agree that "token-level confidence" is not a new concept. Our core innovation lies in its application: We are the first to propose using the confidence score from a single forward pass as a computationally efficient proxy to predict a sample's "fuzziness" under multi-sampling. In other words, we use it to rapidly identify data most likely to provide a useful learning signal for GRPO, thereby completely bypassing the expensive multi-sampling evaluation process. Our contribution is a practical engineering framework (UFO-RL) designed to enhance RL training efficiency by directly tackling this algorithmic bottleneck.

Regarding its reliability as a proxy, we followed your suggestion and conducted a prompt-sensitivity ablation study. The results (table below) show that despite variations in prompt style, the similarity between our confidence score and multi-sample accuracy remains high (>0.8), demonstrating its robustness as an efficient proxy.

Supplemental Experiment: Effect of Prompt Design on Confidence-Accuracy Similarity

We will add this experiment to the appendix to further strengthen the case for our method's reliability.

Regarding Specific Questions

1. On the Justification for the Quadratic Form of the "Fuzziness Score" Reviewer's Comment: "Could you provide further justification for the specific quadratic form $Score(s_i)=1−(s_i−\mu)^2$? Have you experimented with other functions to isolate the 'intermediate uncertainty' region?"

Our Response: This is an excellent question. We chose the quadratic function because it is the most mathematically simple and computationally efficient way to model our goal of finding "intermediate difficulty" data—that is, to assign the highest score to samples with confidence near the mean, which are most likely to produce the mixed-reward signals GRPO requires.

To prove that our method's success is not contingent on this specific function, we conducted a new ablation study. In this experiment, we used a scoring function based on absolute distance $Score = 1 - |s_i - \mu|$, which also selects for data near the mean but is simpler and has no additional hyperparameters. The results show that this function performs almost identically to the original quadratic function. This more strongly demonstrates that the effectiveness of UFO-RL stems from the core principle of selecting intermediate-difficulty data, not from a specific mathematical formula.

Supplemental Experiment: Performance Comparison of Different "Fuzziness Score" Functions

2. On a Dynamic Data Selection Ablation Study Reviewer's Comment: "Have you considered a dynamic implementation where confidence scores and the selected data subset are periodically re-calculated during training, may act as ablation study?"

Our Response: This is a very insightful suggestion. As the model learns via the GRPO algorithm, its capabilities change, and thus the range of "effective data" should also be dynamic. We therefore conducted a new ablation study to explore a curriculum-based dynamic strategy.

In this dynamic implementation, we first selected the 5% "fuzziest" data from the entire training set and trained the model on this subset for the first half of the training process. Then, using the updated model, we re-evaluated the remaining 95% of the data and selected the new 5% "fuzziest" data for the second half of training. This ensures the model is always focusing on the most informative samples relative to its current state.

The results (as shown in the newly added table below) indicate that this dynamic strategy does yield a small but consistent performance improvement (~0.3-0.4%), but at the cost of an additional data selection step. Given that a core goal of our work is to maximize efficiency while maintaining performance, the current static method achieves a better trade-off between performance and cost. Nevertheless, we fully recognize the potential of dynamic selection and will discuss it in our future work.

Supplemental Experiment: Static vs. Dynamic Data Selection Strategy

Once again, we sincerely thank you for your detailed review and insightful feedback. We are confident that the revisions and additions made in response to your comments will significantly strengthen the quality and impact of our paper. We look forward to your further feedback on our revised manuscript.

[1]Yu Q, Zhang Z, Zhu R, et al. Dapo: An open-source llm reinforcement learning system at scale[J]. arXiv preprint arXiv:2503.14476, 2025.

Dear Reviewer y4mU,

Thank you again for your valuable advice on our work. We recognize this is a busy time for all reviewers and truly appreciate you making time for this discussion.

We have submitted our rebuttal and additional experiments, which we hope have thoroughly addressed your concerns. We would greatly value the opportunity to continue our dialogue, especially as the deadline approaches.

Could you please let us know if your concerns have been adequately addressed? If they have, we would be very grateful if you would consider updating your score.

Thank you for your consideration.

Best regards,

The Authors of Paper 27330

Dear Reviewer y4mU,

Thank you once again for your insightful and constructive review. We understand you are extremely busy during this period, and we truly appreciate the time and expertise you've dedicated to our paper.

As the discussion period is nearing its end, we would be very grateful for a final moment of your attention. In our rebuttal, we sought to thoroughly address your primary concerns by not only clarifying our points but also conducting three new targeted experiments.

To briefly summarize how we addressed your key points:

• On Theoretical Grounding: We clarified that our method's core motivation stems from the practical, algorithmic need to mitigate vanishing gradients in RL, and ran a new prompt-sensitivity analysis to prove our confidence metric's robustness as a proxy.
• On Design Choices: We conducted two new ablation studies—one on the "fuzziness score" function and another on a dynamic selection strategy—to validate our approach and explore the trade-offs you astutely pointed out.

We are confident that these clarifications and new empirical results have directly resolved the initial weaknesses you identified and have substantially strengthened our work.

We would be deeply grateful to know if our response has sufficiently addressed your concerns. Your final feedback would be invaluable to us.

Thank you for your time and consideration.

Sincerely, Authors of Paper 27330

Dear Reviewer y4mU,

We hope this message finds you well. We know this is an incredibly demanding time for all reviewers, and we are sincerely grateful for the considerable time and insightful feedback you have already dedicated to our paper (27330).

With the discussion period concluding in approximately 30 hours, we wanted to make one final, respectful check-in. Your review was instrumental, prompting us to conduct three new, targeted experiments to directly address the weaknesses and questions you raised. We believe this new empirical data has substantially strengthened our work.

To briefly recap, based directly on your feedback, we have:

Validated our confidence metric's robustness by running a new prompt-sensitivity analysis, addressing your concern about it being a potentially unreliable proxy.

Strengthened the justification for our "fuzziness score" with a new ablation study on the scoring function, answering your question about the quadratic form.

Explored a dynamic data selection strategy in another new ablation study, as per your excellent suggestion.

We feel these additions, spurred directly by your guidance, have thoroughly addressed the initial concerns you outlined. We would be deeply appreciative if you might have a final moment to let us know whether this new evidence has resolved your concerns. Your final assessment is incredibly valuable to us, and we hope our diligent response might merit a re-evaluation.

Thank you once again for your time and expertise.

Sincerely, The Authors of Paper 27330

We sincerely thank you for your valuable time and constructive feedback. We agree that a solid theoretical foundation is crucial. Your insightful comments have been instrumental in helping us to improve the rigor and clarity of our work.

Before addressing your specific concerns, we would like to emphasize that our original manuscript had already laid the groundwork for our method's machine learning basis in several key sections:

1. Problem Definition (Introduction): We explicitly identified the core challenge in RL fine-tuning as the need for "robust policy gradient estimation" and noted the lack of mechanisms to prioritize instances yielding the most impactful "learning signal".
2. Sampling-based Difficulty Analysis (Section 3): We quantified data difficulty using "sampling accuracy," a statistical estimate based on empirical rewards, and our experiments provided initial evidence that data of intermediate difficulty is most effective for learning.
3. Confidence-based Efficiency Optimization (Section 4): We proposed an efficient confidence estimation method based on the model's intrinsic generation probabilities, requiring only a "single forward pass" as a more efficient machine learning metric for uncertainty.

Your review has prompted us to connect these analyses more deeply and explicitly with the intrinsic mechanism of the GRPO (Group-Relative Policy Optimization) algorithm used in our experiments.

Concern 1: Deepening the Theoretical Foundation — Why Extreme Data is Harmful for GRPO

We wish to clarify that our method's efficacy stems from its precise alignment with the intrinsic mechanism of the GRPO algorithm, specifically by avoiding the pitfalls that arise when GRPO processes data of extreme difficulty.

GRPO's core strategy is to learn from the contrast between "good" and "bad" responses generated for the same prompt. This contrastive mechanism fails—and can even produce harmful signals—when the data is not well-calibrated:

• For "easy" data, where the model's generated responses are almost all correct, the GRPO algorithm loses its basis for effective contrast. In this scenario, its optimization objective becomes ambiguous. Instead of learning the fundamental difference between 'right' and 'wrong', it might be misled into distinguishing between subtle, non-essential variations (e.g., phrasing) among multiple correct answers. This can generate a perverse learning signal, which erroneously penalizes a perfectly valid solution simply because it differs slightly from another correct one in the same group.

• Similarly, for "hard" data, where all responses are incorrect, GRPO lacks a correct 'positive example' to learn from. To perform its relative optimization, the algorithm may be forced to reward the "least incorrect" response among a group of failures. This creates a misleading learning signal, as it is not teaching the model how to be correct but is instead reinforcing a flawed reasoning path, merely because it is marginally better than other flawed paths.

Therefore, the data that allows the GRPO algorithm to function most effectively and safely is precisely the "intermediate difficulty" or "fuzzy" data identified by UFO-RL. These instances reliably provide a mix of correct and incorrect answers, offering the clean and true contrast necessary for GRPO to generate a high-quality policy gradient. This demonstrates that UFO-RL is not just a tool for efficiency, but also a necessary safeguard for the training stability and correctness of contrast-based algorithms like GRPO.

Concern 2: Clarification of the Algorithm

We completely agree that the algorithm's description required more clarity. The process is deterministic and reproducible as follows:

1. Confidence Calculation: For each sample, a single forward pass computes the average log-probability of the generated sequence, which serves as its confidence score.
2. "Fuzziness" Score Calculation: A "fuzziness" score, $Score(s_{i})=1-(s_{i}-\mu)^{2}$, is calculated to identify samples within the ZPD, where $s_i$ is the normalized confidence and $\mu$ is the dataset's mean confidence. This prioritizes samples with near-average confidence.
3. Ranking and Selection: Samples are ranked by their "fuzziness" score in descending order, and the top 10% are selected for the training set.

Revision Plan: To ensure maximum clarity, we will add a formal Algorithm pseudo-code box to the final version of the paper.

In Summary

Thank you once again for your insightful review. We hope this clarification clearly demonstrates that our work is built on a solid analytical foundation and that our theoretical core is strongly coupled with the intrinsic mechanism of the GRPO algorithm. Our method is designed to select data that yields the most effective learning signal while avoiding harmful, misleading updates. We are confident that by making these points more explicit, the paper will be substantially strengthened.

We sincerely hope these explanations fully address your concerns and kindly ask you to reconsider your evaluation of our work.

Dear Reviewer fb4U,

Thank you again for your valuable advice on our work. We recognize this is a busy time for all reviewers and truly appreciate you making time for this discussion.

We have submitted our rebuttal and additional experiments, which we hope have thoroughly addressed your concerns. We would greatly value the opportunity to continue our dialogue, especially as the deadline approaches.

Could you please let us know if your concerns have been adequately addressed? If they have, we would be very grateful if you would consider updating your score.

Thank you for your consideration.

Best regards,

The Authors of Paper 27330

Dear Reviewer fb4U,

Thank you again for your thoughtful feedback. Your review was instrumental in helping us identify where the clarity of our paper could be significantly improved, and we have taken your comments very seriously.

We noted your candid assessment regarding the potential for misunderstanding the central parts of our work. With this in mind, our rebuttal was specifically crafted to provide a clearer and more rigorous explanation for the two main points of confusion you raised.

As the discussion period draws to a close, we hoped we might briefly summarize our clarifications:

• On the Theoretical Basis: We have now grounded our method in the intrinsic mechanism of the GRPO algorithm, explaining precisely how selecting intermediate-difficulty data is essential to avoid the harmful and misleading learning signals that arise from "too easy" or "too hard" samples. This provides the solid machine learning rationale you were looking for.
• On the Algorithm's Clarity: We presented a clear, step-by-step description of our data selection algorithm and have committed to including a formal pseudo-code box in the final version to ensure it is fully transparent and reproducible.

We believe these detailed explanations directly address your concerns about the paper's theoretical foundation and methodological clarity.

Given that your initial review highlighted these specific areas, we would be particularly grateful to know if our rebuttal has successfully clarified our approach and its justification. Your final thoughts are extremely important to us.

Thank you for your time and for helping us improve our paper.

Sincerely, Authors of Paper 27330

We sincerely thank you for your comprehensive and insightful review. Your constructive feedback is invaluable for strengthening our work. We were encouraged that you found the application of ZPD theory "intuitively appealing" and our single-pass confidence estimation "a simple yet computationally efficient proxy."

We have carefully considered all your comments and conducted new experiments to address your questions. Below, we address your concerns point-by-point.

On Weaknesses and Core Questions

1. Positioning with Respect to Related Work

Thank you for encouraging us to clarify our work's positioning. Our framework provides an efficient data selection strategy for RL fine-tuning, which is fundamentally different in its goals and paradigm from existing data selection fields. We will incorporate this discussion into the camera-ready version of our paper to better delineate our unique contributions.

• vs. Active Learning: Active Learning aims to minimize human labeling costs. In contrast, UFO-RL is designed to minimize computational costs (i.e., GPU hours) in RL training, a distinct optimization dimension.
• vs. Coreset Selection: The goal of Coreset Selection is to find a small, representative subset that approximates the full dataset. Conversely, inspired by ZPD theory, UFO-RL intentionally creates a biased, non-representative subset that is most effective for the model's current learning state.
• vs. Data Distillation: Data Distillation is a model compression technique concerned with knowledge transfer between different models. UFO-RL focuses on the self-improvement of a single model by optimizing its own learning process.

2. Direct Comparison with Multi-Sampling Methods

This is a crucial point. To provide the direct comparison you requested and to demonstrate our method's consistency across various model sizes, we have run new experiments and present the complete results for the Qwen2.5 family below.

Table A: Performance & Cost Comparison on GSM8K (on a single 8xA100 node)

Key Findings:

• Performance & Efficiency: From 0.5B to 14B parameters, UFO-RL consistently achieves performance comparable to or better than the baselines while drastically reducing the total training time. This consistent trend highlights the method's robustness.
• Regarding Metric Divergence: We argue that the difference between metrics stems from the superior granularity of our method. Our continuous confidence score provides a more precise proxy for the ZPD, enabling better sample selection.

3. Robustness to Model Calibration and Sequence-Level Uncertainty

This is an excellent question. Our method is inherently robust to calibration issues. To provide empirical proof, we conducted a new experiment using temperature scaling (T=0.8 and T=1.2) to simulate different calibration states on Qwen2.5-7B. The resulting 10% data subsets selected by UFO-RL had an 85.8% overlap, confirming the method's robustness. We plan to add this analysis to the final version of the paper.

4. Generalization Ability

To address your concerns about generalization, we verified our method's effectiveness across both task types and model scales.

• Task Generalization: To demonstrate broader applicability beyond math, we evaluated UFO-RL on the AI2 Reasoning Challenge (ARC-C), a scientific QA task.

  Table B: Performance & Efficiency on ARC-Challenge (LLaMA3.1-8B)

  Method	Data %	Data Selection Time (s)	Total Training Time (s)	ARC-C Acc (%)
  Full Data	100%	0	2432	87.11
  Random	10%	~1	214	86.26
  UFO-RL (Ours)	10%	4	203	88.23

  The results confirm that UFO-RL is also effective for general reasoning tasks, outperforming the full-data baseline with ~1/12 of the training time.

• Scalability to Larger Models: The results in Table A clearly show a consistent and effective scaling path from 0.5B to 14B models, robustly demonstrating our method's effectiveness and scalability.

5. Writing Quality

Thank you for the feedback. We have thoroughly reviewed the manuscript and will further improve its clarity, precision, and conciseness in the final version.

We again thank you for your valuable time and expertise. We believe the new experiments and analyses have significantly strengthened our paper and addressed your concerns. We hope you will re-evaluate our work favorably.

Dear Reviewer 3ahk,

Thank you again for your valuable advice on our work. We recognize this is a busy time for all reviewers and truly appreciate you making time for this discussion.

We have submitted our rebuttal and additional experiments, which we hope have thoroughly addressed your concerns. We would greatly value the opportunity to continue our dialogue, especially as the deadline approaches.

Could you please let us know if your concerns have been adequately addressed? If they have, we would be very grateful if you would consider updating your score.

Thank you for your consideration.

Best regards,

The Authors of Paper 27330