acceptodds
Under review as a conference paper at ICLR 2027

Likelihood Ratio Aware Data Selection for Direct Preference Learning

Abstract

Direct Preference Optimization (DPO) is a widely used offline preference alignment method for large language models (LLMs), and its performance is highly dependent on the quality of the preference data. Conventional selection methods prioritize the separability between paired responses, but we argue that this overlooks the inherent value of individual samples related to the true data distribution. To address this, we propose a likelihood ratio aware data selection framework that identifies high value samples by integrating distributional reliability with data separability. Our method enhances training stability by filtering out low likelihood ratio regions whose gradients are large and directionally incoherent, while achieving superior alignment performance even with a small subset of data that surpasses full dataset training. Extensive experiments across multiple large language models demonstrate that our approach consistently outperforms standard DPO, its variants, and existing selection baselines, and generalizes well to other preference learning losses.

Primary area:
foundation or frontier models, including LLMs
Keywords:
DPO, LLMs, Likelihood ratio, data selection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ySCYOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a preference-data selection criterion combining the winner–loser implicit-reward margin with their summed log likelihood ratios. Cross-fitted DPO proxies score the data, and the selected subset is used for preference training. Experiments evaluate alignment performance, selection ratios, proxy settings, computational costs, and transfer across models and datasets.

Critical strengths
  1. The method is simple and well motivated, combining individual response scores with the preference margin to capture information overlooked by margin-only selection. Its lightweight scoring rule is easy to integrate into existing preference-training pipelines without changing the training objective.
  2. The method is supported by substantial performance gains over existing data-selection baselines. Targeted ablations, evaluations across multiple model families, and repeated LLaMA runs further strengthen the empirical evidence for its effectiveness.
Critical weaknesses
  1. The proxy settings behind the headline results are insufficiently specified. Page 7, line 366 states that experiments mainly use three proxy-estimation rounds. However, Table 2's LLaMA 50% LC/WR of 27.7/28.8 matches Table 4's one-round result of 27.69/28.84, while its 10% result of 19.2/24.1 matches the two-round result of 19.23/24.05. The three-round 10% LC is 16.12, below full-data DPO's 18.2. If different configurations are used within the same table, the configuration associated with each result should be clearly specified.
  2. The interpretation of proxy-based scores could be better supported. High LLR may partly reflect responses that the DPO proxy learns readily, rather than higher intrinsic data quality. Conversely, low LLR may indicate either poor-quality data or valid examples that the proxy has not yet learned. The current analyses do not clearly distinguish these explanations, leaving some uncertainty about what the criterion identifies beyond its demonstrated performance gains.
Questions for authors
  1. Should the parameter domain in Proposition 4.3 be restricted, for example by requiring both original parameters to lie in [−1, 1]? The stated condition alone does not ensure that the reduction preserves the ranking. Please clarify the applicable range.
  2. Which proxy-estimation settings were used in Tables 2, 4, 6, 11, and 12? Please clarify the matching results across Tables 2 and 4 and specify which configuration the runtime comparison measures.
2 comments

Thank Reviewers for their sincere efforts! We will answer to the questions one by one as follows. Here, W means Weaknesses and Q means Questions.


[W1. Clarification of Proxy Estimation Settings]

We thank the reviewer for carefully cross-checking the results, and we agree that the proxy estimation settings should have been stated explicitly.

  • LLaMA: We evaluated m∈1,2,3m \in {1,2,3}, and Table 2 reports the best-performing mm for each selection ratio: m=2m=2 at 10% and m=1m=1 at 50%.

  • Mistral and Qwen: We did not perform a sweep over mm and instead used the default setting of m=3m=3. Thus, their reported results were not tuned with respect to this hyperparameter.

  • Robustness across proxy settings: Table 4 already reports the results for all three values of mm. At the 50% selection ratio, our method outperforms both full-data DPO (18.2/15.5) and the strongest selection baseline, Loss (21.1/18.0), for every mm. Even the weakest configuration (m=3m=3, 25.1/24.1) maintains this advantage.

We will clarify the detailed proxy estimation settings in the appendix and upload the revised manuscript as soon as possible.


[W2. Interpretation of Proxy LLRs and Data Quality]

We agree that a proxy-based score does not by itself establish intrinsic data quality, and we will temper this language in the revised manuscript.

One clarification is important, however. Because of cross-fitting (Algorithm 1), each sample's likelihood ratio is computed by a proxy that never trained on that sample. A high LLR therefore does not simply mean that the proxy fitted this example easily.

To examine what LLR reflects, we compared it with two independent reward models (see our response to Reviewer MDV7, Q1). Proxy LLRs correlate positively with external reward scores, including within reward-margin strata, but the highest-LLR pairs also show higher label disagreement. LLR thus tracks the overall reward level of a pair rather than label reliability, while the margin term in our score preserves the preference signal.

Several existing results further support this distinction:

  • Prioritizing low-LLR samples does not help. If low-LLR samples were mostly valid but unlearned, prioritizing them (λ<0\lambda<0) should help, as in active learning. Instead, it underperforms (Figs. 4 and 10).

  • Low-LLR gradients are mutually near-orthogonal. These gradients (Table 1) do not point in a shared direction that the model has yet to learn.

  • Performance of the selected subsets is robust to the proxy architecture (Fig. 19), which argues against a purely proxy-specific notion of "easiness."


[Q1. Clarification of the Parameter Domain in Proposition 4.3]

We thank the reviewer for the question. Ranking preservation is ensured by the restriction λ,λ1≥−1\lambda, \lambda_1 \geq −1 imposed in the proof (Appendix A.3), which keeps the common ratio in Eq. 17 positive. This restriction was stated in the proof but omitted from the main statement, and we will make it explicit.

Under this restriction, the applicable range is λ1≥−1\lambda_1\geq-1 and λ2−λ1<2\lambda_2-\lambda_1 < 2. This is the region where 1+λ11+\lambda_1 exceeds λ2−1\lambda_2−1. The boundary λ2−λ1=2\lambda_2-\lambda_1=2 corresponds to selection by the summed likelihood ratio, recovered as λ→∞\lambda\rightarrow\infty (case (4) in Appendix A.3). Outside it, the score would reward a pair more for the likelihood of its losing response than of its winning response, contradicting the preference signal that the criterion is designed to preserve. Such weightings are therefore outside the intended scope of the selection criterion. The domain suggested by the reviewer, [−1,1]2[−1, 1]^2, is contained in this range except for the degenerate corner (−1,1)(−1, 1), where the score is identically zero, and is thus a valid sufficient condition.

In the revised manuscript, we will explicitly state the condition λ2−λ1<2\lambda_2-\lambda_1 < 2 in Proposition 4.3 and add the corresponding one-line check to case (1) of the proof.


[Q2. Clarification of Proxy Estimation Settings Across Tables]

Please see our response to W1 for a detailed explanation. For clarity, we summarize the proxy-estimation settings used in each table:

  • Table 2: LLaMA uses m=2m=2 at 10% and m=1m=1 at 50% selection ratios. Mistral and Qwen use m=3m=3.
  • Table 4: All three settings, m∈1,2,3m \in {1,2,3}, are reported. The LLaMA results in Table 2 match the corresponding settings in Table 4.
  • Table 6: The runtime comparison uses m=1m=1.
  • Table 11: The nonlinear score combination experiments use m=3m=3.
  • Table 12: The same settings as Table 2 are used: m=2m=2 at 10% and m=1m=1 at 50%.

The additional clarifications and analyses are helpful. I retain my rating.

Reviewer_JfcFOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2(+1)
Presentation
1
Soundness
2(+1)
Read review
Paper summary

This paper proposes a data selection method for Direct Preference Optimization (DPO). The method uses the absolute values of the implicit rewards, rather than only the relative reward difference between the preferred and rejected responses, to select higher-scoring preference pairs before training. The authors report that models trained on selected subsets can outperform models trained on the full dataset, while also reducing training cost.

Critical strengths
  • The proposed selection method improves downstream alignment performance in several reported settings. In particular, some models trained on selected subsets outperform models trained on the full dataset.
  • The method reduces the amount of data used for the final training stage and can therefore reduce the total training time.
Critical weaknesses
  • The main experimental results are incomplete and their sources are unclear. A large portion of the results in Table 2 is directly taken from other papers, but the paper does not clearly state whether the base models evaluation settings are consistent with those used for the authors' own method. For Qwen-2.5-7B-SFT, many entries are left blank because the corresponding results were not reported by previous work. This substantially weakens the completeness and professionalism of the main comparison.
  • The paper lacks sufficient ablation studies. The paper does not analyze the contribution of each component of the proposed score. For example, the effects of the preferred response (y_w) and rejected response (y_l) should be evaluated separately to determine which part is responsible for the reported gains.
  • The claim about training failure in Section 5.3 is insufficiently supported. The paper states, "As shown in Figure 12 (in Appendix C.4), the training accuracy stagnates near 50%, indicating a failure to learn." However, the paper should also report the corresponding loss curves and other training statistics. Training accuracy alone is not sufficient to establish that the model failed to learn.
Questions for authors

Please address the weaknesses above. Additionally, improve the paper's layout and add a flowchart to make the overall procedure easier to follow.

3 comments

Thank Reviewers for their sincere efforts! We will answer to the questions one by one as follows. Here, W means Weaknesses and Q means Questions.


[W1. Clarification of Experimental Settings]

The source of each entry is stated in the caption of Table 2. The General DPO variations for LLaMA and Mistral are taken from SimPO, and other rows are rerun within our pipeline. The base models and evaluation settings are consistent with SimPO’s. We use the same SFT checkpoints as SimPO (Sec. 5.1, footnotes 4–5) and follow its evaluation protocol (Appendix D.3). For Qwen-2.5, we filled in DPO, IPO, SimPO, soft preference learning methods, and data selection baselines. We did not rerun the remaining General DPO variations, for two reasons. They are not primary baselines for a data selection method, and SimPO is the strongest among them on both LLaMA and Mistral (Table 2).


[W2. Ablation Analysis of Winning and Losing Response]

The separate effects of ywy_w and yly_l are already covered by the λ\lambda sweep. By Proposition 4.3, combinations of (λ1,λ2)(\lambda_1, \lambda_2) within its applicable range is reducible to a single λ\lambda. Moreover, Eq. 6 reads sλ=(λ+1)βlog⁡lr(yw)+(λ−1)βlog⁡lr(yl)s^{\lambda} = (\lambda+1)\beta \log lr(y_w) + (\lambda−1)\beta \log lr(y_l). Hence,

  • λ=1\lambda = 1 selects by ywy_w’s likelihood ratio alone.
  • λ=−1\lambda = −1 selects by yly_l’s likelihood ratio alone, favoring pairs whose losing response has a low likelihood ratio.
  • λ=0\lambda = 0 recovers margin-only selection. All three settings are reported in Fig. 4, Fig. 10, and Tables 19–20.

On the alignment benchmarks (Figs. 4 and 10), performance peaks at 0<λ<10 < \lambda < 1, outperforming both the winner-only (λ=1\lambda=1) and loser-only (λ=−1\lambda=-1) criteria. According to Eq. 6, this range assigns a positive weight to the winning response's LLR and a negative weight to the losing response's LLR, with greater emphasis on the winning response. These results suggest that appropriately balancing the contributions of both responses is more effective than relying on either response alone.


[W3. Additional Evidence of Training Failure]

Training accuracy here is the fraction of training pairs whose implicit reward margin is positive, which is the quantity the DPO objective directly optimizes. Accuracy near 50% therefore means the policy does not separate chosen from rejected responses, i.e., it fails to learn the preference signal. We agree that loss curves make this clearer, and we provide the loss and reward-margin curves of the IS and IS+selection runs on LLaMA-3 at this link. Specifically, 1) the loss is numerically unstable and 2) the reward margin remains mostly negative throughout training. Together with the training accuracy, these curves confirm that the IS variants fail to learn. We will add these curves to Appendix D.4 and upload the revised manuscript as soon as possible.


[Q1. Improving Pipeline Clarity]

The overall procedure is already summarized in Algorithm 1 (p. 6). To improve readability, we will refine the paper's layout and add a flowchart illustrating the three-stage pipeline: likelihood ratio estimation, sample selection, and policy training. We will incorporate these changes into the revised manuscript and upload it as soon as possible.

Dear Reviewer JfcF,

The authors have provided the reponse. Could you please read the response and give the feedback whether the rebuttal has addressed your concern?

Thanks.

Thank you for the authors' clarification. I have improved my rating accordingly.

Reviewer_MDV7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a likelihood-ratio-aware data selection method for DPO. The authors argue that reward-margin-based selection overlooks the individual quality of preferred and rejected responses. They establish a theoretical lower bound connecting likelihood ratios to true rewards under the optimal policy. They also analyze how low-likelihood-ratio samples relate to larger and less consistent gradients. Based on these findings, the method combines reward margins with likelihood-ratio information through a weighted scoring function. Cross-fitted DPO proxy models are then used to estimate likelihood ratios and select preference pairs.

Critical strengths
  • The paper addresses a relevant problem in preference data selection, and the proposed method is simple to implement without modifying the DPO training objective.

  • The experimental results are encouraging, with consistent improvements across several model families and evaluation benchmarks.

  • The analysis of likelihood ratios and gradient behavior provides useful insights into the characteristics of selected preference data.

Critical weaknesses
  • The proposed method mainly adds a weighted likelihood-ratio term to the existing DPO reward margin. Since both signals are already derived from DPO implicit rewards, the methodological contribution appears relatively incremental.

  • The theoretical results motivate the use of likelihood ratios but do not directly justify the proposed linear scoring function or explain why it should lead to better data selection.

  • The main weighting parameter (λ\lambda) is chosen through grid search and then fixed across models and samples. It is unclear whether this global weighting is appropriate for different preference distributions or sample characteristics.

  • The theoretical support for data quality estimation is limited and unclear. The reward lower bound holds for the optimal policy and does not guarantee that higher likelihood ratios correspond to higher-quality responses (Theorem 3.1). It neither guarantees that likelihood ratios preserve response quality rankings nor applies directly to the cross-fitted DPO proxies used for selection.

Questions for authors
  • Can the authors validate whether proxy-based likelihood ratios correlate with response quality measured by an independent reward model or human judgments?

  • How many training runs were needed for (λ\lambda) in the grid search? Would new models or datasets require repeating this search, and how would that affect the overall cost?

  • Would sample-adaptive weighting improve data selection?

3 comments

Thank Reviewers for their sincere efforts! We will answer to the questions one by one as follows. Here, W means Weaknesses and Q means Questions.

[W1. Methodological Novelty ]

We respectfully disagree that the contribution is incremental. What makes DPO valuable is that it performs RL-style alignment while minimizing additional model training and inference. Our method preserves this spirit while additionally accounting for response quality. It only uses quantities that DPO already defines and requires no external model.

The contribution is not a new signal source, but the observation that margin-based selection is blind to the quality of each individual response (Fig. 1a). The Pearson correlation between LLR and the DPO loss, which is monotone in the margin, is only 0.21 on the training set (Fig. 2). Accounting for LLR therefore changes which samples are selected (Table 5) and improves alignment. Our external reward model analysis (response to Q1) further shows that, within each reward-margin decile, higher-LLR pairs receive higher external rewards.

As requested in Q1, we report the correlation with an independent reward model. However, we do not consider re-introducing an external reward model for selection appropriate for DPO-based work. Our claim is not that a DPO proxy is a better selector than an external model. Rather, it is that selection should consider the quality of each response itself, not only the margin between responses.


[W2. Theoretical Motivation and Justification of the Linear Scoring Function]

We would like to clarify the role of the theory. Theorem 3.1 motivates the likelihood ratio as a selection signal (Sec. 3.2, Remark 3.2). It was not presented as a justification of the specific linear form.

The linear score is the simplest instantiation of jointly considering the reward margin and the likelihood ratio, and it has a direct per-response interpretation (Eq. 6).

As stated on p. 5, nonlinear combinations are also examined empirically in Appendix D.4 (Table 14). Adding a multiplicative interaction term yields no consistent performance improvement across evaluation metrics and selection ratios. Therefore, we retain the simpler and more interpretable linear form.


[W3. Generality of Global Weighting]

We agree that datasets with different preference distributions may have different optimal λ\lambda. Across the three model families we evaluate, however, λ=2/3\lambda = 2/3 tuned on LLaMA-3 transfers without retuning. Performance peaks in 0<λ<10 < \lambda < 1 for all three models (Fig. 10). Regarding sample characteristics, sample adaptive weighting is possible when $\lambda $ is defined as a fixed function of sample features. One simple instance is already examined in Appendix C.4. Letting λ\lambda depend linearly on the reward margin, λ(x)=λ+γ⋅RMθ\lambda(x) = \lambda + \gamma\cdot RM_{\theta}, is equivalent to adding the multiplicative term γ⋅RMθ⋅LLRθ\gamma\cdot RM_{\theta} \cdot LLR_{\theta} in Table 14, which gave no consistent gain. More principled forms of sample-dependent λ\lambda require careful design and validation, and we will add this direction to Future Work.


[W4. Theoretical Support for Data Quality Estimation]

We would like to clarify several points regarding the theoretical justification of our method.

  • Theorem 3.1 is motivational. We state it explicitly as a lower bound on the centered reward under the optimal policy, and as motivation for the criterion rather than a bound on the proxy-based score (Remark 3.2; Limitations).

  • The cross-fitted proxy follows prior work (Gao et al., 2025) [1]. Its effect is examined empirically: the number of proxy models (Table 4) and the proxy architecture (Fig. 19).

  • Evidence on sample quality is likewise empirical, through perplexity (Table 5) and alignment performance (Tables 2–3), and correlation with external reward models (Tables A–C, response to Q1).

  • Under the optimal policy and for a fixed prompt xx, Eq. 3 makes βlog⁡lrθ∗(y)\beta \log lr_{\theta^{\ast}}(y) an affine function of r∗(x,y)r^{\ast}(x, y), since the offset βlog⁡Z(x)\beta \log Z(x) depends only on xx. Rankings among responses to the same prompt are therefore preserved exactly. Across prompts the offset varies, and Theorem 3.1 gives only a lower bound on the centered reward, so we do not claim that likelihood ratios preserve a global ranking. Data selection does not require one; whether low-scoring pairs are less valuable is examined empirically, as above.

[1] Gao, Chengqian, et al. "Principled Data Selection for Alignment: The Hidden Risks of Difficult Examples", ICML 2025.

[Q1. Empirical Validation of Proxy LLRs as a Response Quality Signal]

As requested, we compared the cross-fitted proxy log likelihood ratios (LLRs) with scores from two independent reward models, ArmoRM-Llama3-8B and Skywork-Reward-8B, on 60,680 UltraFeedback preference pairs. These reward models are themselves learned estimates of response quality rather than ground truth, so we present this analysis as an external consistency check. Collecting human judgments was not feasible within the rebuttal period.

  1. Correlation with external reward scores. Proxy LLRs are positively correlated with external reward scores for both chosen and rejected responses. The correlations are moderate (0.31–0.38 for ArmoRM and 0.51–0.58 for Skywork) and similar across the number of proxy runs mm.

Table A. Spearman correlations between proxy LLRs and external reward scores.

ProxyArmoRM (Chosen / Rejected)Skywork (Chosen / Rejected)
m=10.346 / 0.3810.508 / 0.579
m=20.330 / 0.3630.518 / 0.579
m=3 (default)0.310 / 0.3490.514 / 0.574
  1. Response quality across LLR deciles. We divide pairs into deciles by total LLR, T=Lw+LlT = L_w + L_l (m=3m=3). Compared with the lowest decile, pairs in the highest decile receive higher ArmoRM rewards for both chosen and rejected responses. However, they also show higher preference-label disagreement with ArmoRM (32.8% vs. 22.0%). Total LLR thus reflects the overall reward level of a pair, but not the reliability of its preference label.

Table B. Mean ArmoRM rewards and label disagreement in the lowest and highest total-LLR deciles (m=3m=3). Label disagreement is the fraction of pairs in which ArmoRM scores the rejected response higher than the chosen one.

MetricLowest 10%Highest 10%
Chosen reward0.1060.148
Rejected reward0.0730.138
Label disagreement22.0%32.8%
  1. Association beyond the reward margin. Within each of 10 reward-margin deciles, we compare the top and bottom thirds by total LLR. The high-LLR group receives higher ArmoRM rewards for both chosen and rejected responses in all 10 strata, for every mm. The association between LLR and external reward is therefore not explained by the margin alone.

Table C. Margin-stratified comparison of ArmoRM rewards (averaged over 10 strata). Positive strata counts the strata in which the high-LLR group has the higher reward.

ProxyChosen (Low → High)Rejected (Low → High)Positive strata (C / R)
m=10.1159 → 0.1485 (+0.0326)0.0863 → 0.1248 (+0.0385)10/10, 10/10
m=20.1170 → 0.1473 (+0.0303)0.0875 → 0.1234 (+0.0360)10/10, 10/10
m=3 (default)0.1182 → 0.1457 (+0.0276)0.0889 → 0.1218 (+0.0329)10/10, 10/10

Proxy LLRs are positively associated with external reward scores, including within reward-margin strata, but they do not by themselves indicate label reliability (Table B). This is consistent with our selection results. Performance peaks at 0<λ<10 < \lambda < 1, where the score still penalizes losing responses that the policy finds likely, and drops at λ=1\lambda = 1, where this penalty vanishes and selection moves toward likelihood alone (Fig. 4, Fig. 10). We therefore do not claim that proxy LLRs directly measure response quality, only that they provide an inexpensive signal correlated with it, which our score combines with the margin.


[Q2. Hyperparameter Search Cost]

We searched λ∈{1,2/3,1/3,0,−1/3,−2/3,−9/11,−1}\lambda \in \{1, 2/3, 1/3, 0, −1/3, −2/3, −9/11, −1 \} on LLaMA-3 at a 50% selection ratio, i.e., eight training runs, and λ=2/3\lambda = 2/3 performed best. All other experiments use this setting unless specified. A new model or dataset may require a new search. However, the same λ\lambda transfers to Mistral and Qwen without retuning (Table 2, Appendix D.4). For the overall cost, proxy estimation (Step 1) does not depend on λ\lambda and is computed once, and re-scoring (Step 2) takes seconds (19s, Table 6). Each grid point therefore costs only one DPO training run on the selected subset.


[Q3. Effectiveness of Sample-Adaptive Weighting]

As discussed in W3, we examined a simple form of sample-adaptive weighting in which λ depends linearly on the reward margin (Appendix D.4, Table 14). This is equivalent to adding a multiplicative interaction term between the reward margin and LLR, which yielded no consistent improvement across evaluation metrics and selection ratios. While this does not rule out potential benefits from more sophisticated adaptive weighting schemes, we leave their design and validation to future work.

Thank you for the detailed response and additional analyses. The added experiments partly address my concerns. However, my main concern about the theoretical support remains. As the authors acknowledge, Theorem 3.1 is only motivational: it holds for the optimal policy, does not justify the linear score, and does not apply to the cross-fitted proxies used for selection. The method therefore rests primarily on empirical evidence, and stronger theoretical support is needed. Novelty is also a concern. The response frames the contribution as the observation that margin-based selection overlooks individual response quality. However, this observation is fairly intuitive, the use of likelihood ratios is not new, the cross-fitted proxy follows prior work, and both quantities are already defined by DPO's implicit reward. The methodological contribution therefore remains relatively incremental. Overall, I inclined to maintain my original rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

Dear Reviewers and AC,

We sincerely thank all reviewers for their constructive feedback. We have revised the manuscript to address the major concerns, with all changes highlighted in blue. The key revisions are summarized below.

  • Theoretical correction: We clarified the parameter domain of Proposition 4.3 and revised the proof of ranking preservation (Sec. 4, Appendix A.3).
  • External validation of proxy LLRs: We added analyses using two external reward models on 60,680 UltraFeedback preference pairs, examining the relationship between LLRs and response quality, as well as its limitations (Appendix D.4, Tables 11–13).
  • Experimental transparency: We explicitly documented the proxy estimation settings across models and selection ratios to improve reproducibility (Appendix D.2).
  • Additional evidence and presentation improvements: We added training-loss and reward-margin curves to substantiate the observed importance-sampling instability (Figs. 14–15), along with a three-stage pipeline flowchart (Fig. 8).

We have uploaded the revised PDF and kindly ask the reviewers to examine the updated manuscript. We also updated our previous rebuttal responses solely to reflect changes in table, figure, and section numbering following the manuscript revision.

We hope these revisions address the reviewers' concerns and strengthen the theoretical rigor, empirical support, and reproducibility of our work. Please feel free to ask any further questions.

Thank you for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.