acceptodds
Under review as a conference paper at ICLR 2027

Modeling and Debiasing Temporal Popularity in Recommendation with Counting Processes

Abstract

Implicit-feedback recommender systems are trained on logged interactions in which observed choices conflate user-specific preference with time-varying item popularity. We formulate timestamped user-item interactions as counting processes and decompose the conditional interaction intensity into user activity, dynamic item popularity, and user-specific utility. This formulation reveals an identifiability issue: without additional structure, item-time effects shared across users can be absorbed into the utility function. We show that centering the utility across users resolves this ambiguity up to ranking-equivalent shifts. We further establish that unmodeled popularity can induce a shared direction in user-context representations, leading to geometric homogenization and degraded uniformity. Motivated by this analysis, we propose a general counting process framework that jointly addresses the challenging tasks of debiasing temporal item popularity and identifying user–item-specific effects. Experiments on synthetic and real-world datasets support the theoretical analysis and show consistent improvements across diverse recommendation backbones.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Recommender Systems, Temporal Popularity Bias, Counting Process, Representation Learning

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_ndQyOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper models timestamped user–item interactions as a counting process, decomposing item choice logits into time-varying popularity and user-specific utility. Theoretically, it proves that unmodeled popularity induces a shared directional shift across users that collapses representation uniformity and that zero-mean centering across users restores identifiability. Methodologically, it jointly trains a Hawkes popularity process and any recommender backbone via a 15-negative importance-sampled softmax loss plus a U-statistic centering penalty. While effective on synthetic and head items, unregularized Hawkes rates over-concentrate on KuaiRand dataset.

Critical strengths

By modeling timestamped implicit feedback via conditional intensity and conditioning on user interaction arrival times (Proposition 2), the paper cleanly eliminates unknown user activity rates and reduces temporal popularity confounding to an additive logit ambiguity. Lemma 1, Theorem 1, and Corollary 1 then provide a transparent geometric proof of why unmodeled temporal popularity causes representation collapse: projecting log⁡π0(v∣t)\log \pi_0 (v|t) onto item embeddings induces a shared shift across all user-context vectors at time 𝑡, contracting pairwise distances on the unit sphere and driving the hyperspherical uniformity loss to 0. This theoretical mechanism is directly verified in the controlled 30-seed synthetic experiment

Rather than hiding or glossing over the sharp drop in Tail Recall@10 on KuaiRand in Table 1, the authors dedicate Appendix A.7 to a rigorous root-cause investigation. By comparing utility-only inference against default 𝜂=0.5 (Table 4), measuring normalized entropy 𝐻_norm and top-5% probability mass of the fitted Hawkes distribution vs. empirical 3-day popularity (Table 5), and plotting learned background rates across item frequency percentiles, the authors clearly separate the effectiveness of the learned utility from the over-concentration of the unregularized Hawkes background rate.

Critical weaknesses

Table 1 Reporting Errors:Duplicated SASRec/MF Row: In Table 1 on Micro-Video, the entire row for Vanilla SASRec (0.0121±0.0009, 0.0061±0.0006, 0.0168±0.0017, 0.0089±0.0009, 0.0112±0.0005, 0.0054±0.0005) is identical to Vanilla MF. Might be some copy/paste problem

Missing Loss-Function Control Standard sequential recommenders (SASRec, TiSASRec, GRU, etc.) are typically trained with binary cross-entropy (BCE) or BPR using 1 negative sample, whereas +(Ours) trains the backbone with an importance-corrected sampled softmax choice loss using M=15 negative samples and adds a separate 128-dimensional item embedding table. In collaborative filtering and sequential recommendation, replacing 1-negative BCE/BPR with multi-negative sampled softmax is well known to produce large gains in both Recall/NDCG and representation uniformity (since softmax cross-entropy directly optimizes an alignment-and-uniformity objective). Without an ablation training the vanilla backbone under ℓ_choice with uniform popularity (𝜆_𝑣 ≡ 1, 𝛾=0) it is impossible to tell how much of the large jump in Table 1, Table 2, and Figure 2 comes from temporal popularity modeling vs. simply switching the training loss to 15-negative sampled softmax.

Disconnect Between Theory and Practice on the Centering Penalty (𝛾) and Unregularized Hawkes Estimation (λ_v)

Negligible Empirical Impact of Centering: While Proposition 1 relies on user-centering to establish identifiability, Figure 2 shows that setting γ=0 (no centering penalty) already achieves essentially the same representation uniformity and Recall@10 as the best γ>0 across all six backbones and three datasets. This suggests that architectural separation alone already prevents most popularity absorption in practice.

Unregularized Joint Estimation of 𝜆_𝑣(𝑡;Φ): Inside ℓ_choice because 𝜆_𝑣(𝑡;Φ) is trained solely through the 15-negative importance-sampled softmax choice loss ℓ_choice rather than a temporal point-process likelihood, the background rate is poorly calibrated on KuaiRand, concentrating 99.4% of probability mass onto the top 5% of items (Table 5). When combined at inference via Eq. (3) using the validation-selected 𝜂=0.5 (chosen on overall Recall@10, which favors Head items), it destroys Tail Recall@10 by 85%–87% across GRU, SASRec, TiSASRec, and FEARec.

Questions for authors

see above weakness summary

8 comments

We sincerely thank Reviewer ndQy for the careful examination of our experimental methodology and results, which has motivated additional experiments and strengthened our empirical evaluation.


W1. Correction of the Reporting Error in Table 1

We thank the reviewer for carefully identifying this error. The duplicated values in Table 1 resulted from a copy-and-paste error in the Vanilla SASRec row for Micro-Video. We have corrected the results in the revised manuscript, but the overall interpretation of the results remains unchanged. Please refer to Table 1 below or line 390 of the revised manuscript.

Table 1: Micro Video - SASRec Performance

MethodAll Recall@10All NDCG@10Head Recall@10Head NDCG@10Tail Recall@10Tail NDCG@10
SASRec0.1402±0.00300.0661±0.00140.1743±0.00640.0798±0.00310.1241±0.00150.0596±0.0007
SASRec +(Ours)0.1632±0.00390.0803±0.00220.2017±0.00480.0983±0.00430.1449±0.00400.0718±0.0014


W2. Clarification of Training Controls and Additional Loss-Controlled Experiments

We first clarify that the Vanilla backbones were trained using BCE with 15 negative samples per positive interaction. Thus, the number of negative samples was controlled when comparing Vanilla and Ours. We thank the reviewer for pointing out the need for clarification and have added this detail to the revised manuscript (revised manuscript, lines 1005-1007).

Following the reviewer's suggestion, we conducted an additional experiment using the same sampled-softmax loss as Ours, but with uniform popularity (λv(t;Φ)=1\lambda_v(t;\Phi) = 1) and no centering penalty (γ=0\gamma=0). As shown in Tables 2-1, 2-2, 2-3, and 3 below (see also lines 1134–1201 of the revised manuscript), replacing BCE with the sampled-softmax loss substantially improves recommendation performance and representation uniformity. Nevertheless, Ours achieves further improvements, outperforming this loss-controlled baseline in 16 of 18 settings for overall Recall@10, all 18 settings for overall NDCG@10, and 12 of 18 settings for representation uniformity.

These results demonstrate that the improvements cannot be attributed solely to the change in the training objective, supporting the additional benefits of explicitly modeling temporal popularity and separating it from user-specific utility.



W3. Clarification Request Regarding the Theory–Practice Disconnect

We would appreciate further clarification on the specific disconnect between the theory and its practical implementation.



W4. Understanding the Role of the Centering Penalty in Practice

We thank the reviewer for this observation. To explain the strong performance at γ=0\gamma=0, we additionally examined whether the learned utility approximately satisfies the centering condition even without explicit regularization. However, we have not yet established why this centering phenomenon occurs. What is clear is that identifiability is not guaranteed unless the centering constraint is satisfied. Therefore, we believe that introducing the centering penalty is necessary to prevent cases in which the observed centering phenomenon does not arise, thereby encouraging the proposed model to satisfy the condition required for identifiability. We have added this analysis and the corresponding results (revised manuscripts, lines 469--494).



W5.

We are currently conducting additional experiments to address weakness 5. We will post our response as soon as the experiments are completed.

Table 2-1: Micro Video

MethodAll Recall@10All NDCG@10Head Recall@10Head NDCG@10Tail Recall@10Tail NDCG@10
MF0.0245±0.00090.0122±0.00030.0403±0.00260.0203±0.00090.0169±0.00080.0085±0.0005
+(Ours)0.0637±0.00030.0318±0.00020.0928±0.00080.0460±0.00040.0500±0.00030.0250±0.0005
GRU0.0752±0.00290.0405±0.00070.1157±0.00390.0623±0.00170.0560±0.00330.0303±0.0011
+(Ours)0.1004±0.00130.0609±0.00020.1609±0.00280.0924±0.00300.0718±0.00300.0461±0.0015
SASRec0.1492±0.00310.0712±0.00160.1792±0.00330.0850±0.00230.1350±0.00310.0646±0.0013
+(Ours)0.1632±0.00390.0803±0.00220.2017±0.00480.0983±0.00430.1449±0.00400.0718±0.0014
TiSASRec0.1545±0.00080.0734±0.00080.1872±0.00470.0884±0.00170.1391±0.00200.0663±0.0007
+(Ours)0.1695±0.00140.0845±0.00070.1998±0.00350.1005±0.00150.1552±0.00100.0768±0.0006
FEARec0.1570±0.00130.0741±0.00050.1893±0.00170.0891±0.00140.1417±0.00240.0670±0.0010
+(Ours)0.1742±0.00070.0855±0.00040.2124±0.00100.1046±0.00130.1561±0.00070.0765±0.0003
BSARec0.1613±0.00280.0755±0.00100.1933±0.00330.0906±0.00120.1455±0.00250.0680±0.0009
+(Ours)0.1715±0.00090.0848±0.00030.2105±0.00390.1045±0.00140.1530±0.00120.0755±0.0004

Table 2-2: MovieLens

MethodAll Recall@10All NDCG@10Head Recall@10Head NDCG@10Tail Recall@10Tail NDCG@10
MF0.0294±0.00090.0139±0.00040.0435±0.00220.0202±0.00110.0234±0.00040.0112±0.0003
+(Ours)0.0375±0.00040.0177±0.00010.0728±0.00030.0341±0.00020.0225±0.00050.0108±0.0002
GRU0.0803±0.00080.0388±0.00050.0970±0.00130.0471±0.00080.0732±0.00090.0353±0.0005
+(Ours)0.0803±0.00090.0389±0.00030.1061±0.00050.0533±0.00040.0694±0.00120.0328±0.0006
SASRec0.0839±0.00070.0359±0.00010.0966±0.00290.0416±0.00200.0785±0.00190.0335±0.0010
+(Ours)0.1069±0.00070.0515±0.00030.1346±0.00080.0666±0.00050.0951±0.00090.0451±0.0005
TiSASRec0.0921±0.00110.0386±0.00060.1059±0.00200.0447±0.00080.0863±0.00080.0361±0.0005
+(Ours)0.1080±0.00090.0521±0.00050.1313±0.00240.0649±0.00130.0981±0.00100.0467±0.0004
FEARec0.0861±0.00080.0361±0.00030.1014±0.00120.0427±0.00060.0797±0.00120.0333±0.0005
+(Ours)0.0810±0.00300.0385±0.00160.0918±0.00530.0436±0.00280.0764±0.00210.0363±0.0011
BSARec0.0727±0.00130.0303±0.00050.0827±0.00150.0348±0.00070.0696±0.00240.0287±0.0010
+(Ours)0.0882±0.00070.0418±0.00020.1046±0.00140.0505±0.00080.0812±0.00060.0381±0.0001

Table 2-3: KuaiRand

MethodAll Recall@10All NDCG@10Head Recall@10Head NDCG@10Tail Recall@10Tail NDCG@10
MF0.0393±0.00070.0190±0.00040.0468±0.00070.0227±0.00040.0114±0.00060.0050±0.0003
+(Ours)0.0672±0.00020.0324±0.00010.0794±0.00010.0386±0.00020.0209±0.00110.0094±0.0003
GRU0.0532±0.00060.0254±0.00040.0644±0.00100.0308±0.00060.0112±0.00100.0050±0.0005
+(Ours)0.0684±0.00090.0336±0.00060.0860±0.00100.0424±0.00070.0017±0.00040.0006±0.0001
SASRec0.0536±0.00090.0228±0.00030.0618±0.00130.0262±0.00050.0225±0.00160.0099±0.0008
+(Ours)0.0735±0.00060.0358±0.00050.0921±0.00080.0449±0.00060.0036±0.00010.0013±0.0000
TiSASRec0.0593±0.00100.0250±0.00040.0688±0.00130.0289±0.00050.0236±0.00040.0104±0.0001
+(Ours)0.0749±0.00030.0363±0.00020.0939±0.00040.0456±0.00030.0036±0.00040.0013±0.0001
FEARec0.0576±0.00160.0247±0.00060.0667±0.00240.0285±0.00110.0234±0.00180.0105±0.0011
+(Ours)0.0611±0.00120.0305±0.00060.0770±0.00160.0385±0.00070.0008±0.00040.0003±0.0001
BSARec0.0414±0.00050.0169±0.00020.0483±0.00120.0195±0.00040.0202±0.00090.0086±0.0004
+(Ours)0.0679±0.00040.0333±0.00030.0854±0.00040.0421±0.00040.0019±0.00030.0007±0.0001

Table 3: Uniformity

BackboneMicro Video (Vanilla)Micro Video (Ours)MovieLens-1M (Vanilla)MovieLens-1M (Ours)KuaiRand (Vanilla)KuaiRand (Ours)
MF−3.9336±0.0025−3.8220±0.0035−3.6788±0.0014−3.7804±0.0027−3.7638±0.0037−3.7123±0.0052
GRU−3.0772±0.0286−3.8029±0.0031−2.6302±0.0183−3.5521±0.0185−2.5160±0.0451−3.1821±0.0073
SASRec−3.7316±0.0253−3.5652±0.0171−2.9970±0.0266−3.0466±0.0066−2.8904±0.0223−3.0597±0.0185
TiSASRec−3.6769±0.0101−3.6116±0.0111−2.7048±0.0217−2.9562±0.0246−2.5738±0.0286−2.8812±0.0194
FEARec−3.8309±0.0048−3.6014±0.0247−2.8215±0.0198−3.1607±0.0504−2.4473±0.0645−2.7468±0.0397
BSARec−3.8369±0.0093−3.6024±0.0136−2.9827±0.0257−3.2671±0.0137−2.4649±0.0506−3.0626±0.0120
Reviewer_gMVmOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper decomposes user‑item interaction intensity into user activity, time‑varying item popularity, and user‑specific utility. It identifies an ambiguity between popularity and utility and introduces a centering constraint to achieve ranking‑equivalent identifiability. It further provides a theoretical analysis connecting unmodeled popularity effects to representation homogenization. Building on these insights, this work proposes a framework combining Hawkes‑based popularity modeling, importance‑corrected choice estimation, and stochastic centering.

Critical strengths

The paper addresses the important problem of temporal popularity bias in recommendation systems. The countingprocess formulation provides a coherent probabilistic perspective for separating dynamic item popularity from userspecific utility. The identifiability and representationgeometry analyses offer useful theoretical motivation for the proposed framework.

Critical weaknesses

The identifiability guarantee hinges on the centeredutility assumption, which might not hold for realworld user preferences. Enforcing this constraint picks out one specific decomposition, yet it cannot guarantee recovery of groundtruth intrinsic preferences under model misspecification.

Questions for authors

1.What are the fundamental methodological differences between the proposed framework and TCHN, particularly regarding Hawkes-based modeling and preference-conformity disentanglement? Could the authors provide a more detailed comparison? 2.What happens when user preferences contain systematic item-level effects that violate this assumption?

2 comments

We sincerely thank Reviewer gMVm for the insightful questions regarding identifiability and related work, which have helped us better articulate the contributions and scope of our study.


Q1. Clarification of Our Distinct Research Focus and Methodological Differences from TCHN.

We thank the reviewer for highlighting this relevant work. Although both TCHN and our framework employ Hawkes processes, they address fundamentally different phenomena. TCHN models behavioral conformity arising from inter-user influences through mutually exciting Hawkes processes. In contrast, our work investigates representation homogenization caused by model misspecification, where unmodeled temporal item popularity is incorrectly absorbed into user-specific utility, distorting learned preferences. We further establish the identifiability of popularity and utility under a centering condition.

These approaches are complementary. An interesting direction for future work is to extend our framework to dependent counting processes that explicitly model behavioral conformity, potentially incorporating the ideas developed in TCHN.


W1 & Q2. Clarification of Identifiability and Interpretation under Violations of the Centering Assumption.

We thank the reviewer for raising this important question. Suppose the true utility contains a systematic item-level effect, such that f∗(xu(t),v)=g(xu(t),v)+b(v,t),f^*(x_u(t),v)=g(x_u(t),v)+b(v,t),

where gg satisfies the centering condition and b(v,t)b(v,t) represents a shared item-level effect. Substituting this into Eq. (1), we obtain

\propto \pi_0(v\mid t)\exp\{f^*(x_u(t),v)\} =\pi_0(v\mid t)\exp\{b(v,t)\} \exp\{g(x_u(t),v)\} \propto \widetilde\pi_0(v\mid t) \exp\{g(x_u(t),v)\},$$ where $$\widetilde\pi_0(v\mid t) =\frac{\pi_0(v\mid t)\exp\{b(v, t)\}} {\sum_k\pi_0(k\mid t)\exp\{b(k, t)\}}.$$ Thus, the same conditional choice distribution can be represented by a centered utility $g$ and a redefined item-level component $\widetilde\pi_0$. **Regarding Assumption 1** Reviewer Tguc raised a similar concern regarding this assumption. We would like to clarify the following point. The centering assumption does not imply that the true average preference for each item is zero. Rather, it defines the remaining user-specific effect after removing the common item-level effect relative to a given population. This is a standard identification constraint used to distinguish user–item-specific interaction effects from item-level main effects, analogous to the conventional constraints employed in two-way ANOVA. To avoid further misunderstanding, we will explicitly clarify that Proposition 1 establishes the identifiability of centered user-specific relative utility, rather than the recovery of population-invariant intrinsic preferences.

We appreciate Reviewer gMVm's careful evaluation and valuable suggestions, which helped us better articulate the scope and contributions of our work. Although we are unsure whether you have seen our responses, we hope they have helped resolve your questions and concerns.

Thank you once again for your time and consideration.

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

This paper proposes a counting-process framework for modeling and debiasing temporal item popularity in recommendation. It aims to separate time-varying item popularity from user-specific utility and address the identifiability problem between them. The method combines Hawkes-based popularity modeling, an importance-corrected choice loss, and a centering penalty, and provides a theoretical analysis of popularity-induced representation homogenization. Experiments on synthetic data and three real-world datasets with six recommendation backbones show improvements in recommendation performance and representation uniformity.

Critical strengths
  1. The paper provides a theoretical analysis of the identifiability problem between temporal popularity and user-specific utility and establishes a connection between popularity confounding and representation homogenization.
  2. The experiments cover three real-world datasets and six recommendation backbones, demonstrating improvements in recommendation performance and representation uniformity across different architectures.
Critical weaknesses
  1. Insufficient justification of the centering assumption. The main theoretical contribution relies on Assumption 1, which requires the average utility of each item across all users to be zero. However, this assumption may not hold for real user preferences. For example, if we remove half of the users who dislike a particular item, the average preference for that item will increase. To satisfy the centering constraint, the model may absorb this shared preference into the popularity component, even though the item's underlying popularity has not changed. This suggests that the learned popularity and utility may depend on the sampled user population rather than reflecting the true popularity and user preferences. The paper does not sufficiently discuss this issue or evaluate the robustness of its decomposition under different user populations. Moreover, Figure 2 shows that most of the performance improvement is achieved without centering ((γ=0\gamma=0)), raising further questions about the practical necessity of this constraint.
  2. Insufficient justification for the counting-process formulation. The paper models user–item interactions as counting processes. However, in many recommendation settings, each user–item pair has at most one observed interaction. The synthetic experiments also explicitly exclude previously interacted items, meaning that each user item counting process can only increase from 0 to 1. In such cases, the necessity of the counting-process formulation is unclear. Therefore, the paper does not sufficiently explain what benefits the counting-process formulation provides in these recommendation settings.
Questions for authors
  1. How would removing users who dislike an item affect the estimated popularity and utility, given that the centering constraint depends on the sampled user population?
  2. Given that many recommendation datasets contain only one interaction per user–item pair, what is the necessity of the counting-process formulation, and what additional benefits does it provide over directly modeling temporal item popularity?
3 comments

We sincerely thank Reviewer Tguc for the thoughtful comments on the centering assumption and counting-process formulation, which have helped us clarify the theoretical foundations of our framework.


W1 & Q1. Justification of the Centering Assumption.

  • We would like to clarify that Assumption 1 does not imply that intrinsic user preferences average to zero. Rather, it defines user-specific utility as the remaining effect after removing item-level effects shared across the reference user population. This is a standard identification constraint, analogous to the sum-to-zero constraint in two-way ANOVA, used to distinguish shared effects from user–item-specific interactions. To avoid misunderstanding, we will explicitly clarify that Proposition 1 identifies population-centered, user-specific relative utility, rather than recovering population-invariant intrinsic preferences (revised manuscript, lines 168--187).

  • We also appreciate the reviewer's valid concern regarding the practical necessity of the centering penalty. We further examined the learned utility scores and found that their user-averaged values were already concentrated near zero even without explicit centering regularization (revised manuscript, lines 469–494). However, since this behavior is not theoretically guaranteed in general, we retain the explicit centering penalty to enforce the identification constraint.


W2 & Q2. Justification for the Counting-Process Formulation.

  • Hawkes processes have been widely used to model the diffusion of information and popularity. It is well established that such diffusion dynamics can be described either through a self-exciting process, in which past interactions with an item influence the occurrence of future interactions, or through a branching-process interpretation, in which each consumption event may trigger subsequent consumption events [1].

  • In our study, we also found that Hawkes processes generally provide a good empirical fit to the item-level intensity functions in the datasets considered. This observation motivated our choice to model item-level interaction intensities using Hawkes processes. However, we would like to emphasize that the central contribution of our paper does not lie in the use of Hawkes processes themselves, but rather in providing a systematic characterization of popularity in recommendation systems through a counting-process framework.

  • The reviewer raises an important question regarding the advantages of formulating recommendation models within a time-dependent counting-process framework. Indeed, when time is fixed, the preference model derived from our counting-process formulation takes the same mathematical form as existing time-aware recommendation models. In particular, Eq. (1) can be expressed as a time-varying Luce choice model and admits an equivalent interpretation in terms of Gumbel random utility.

  • However, we believe that this equivalence does not diminish the value of our proposed theoretical framework. Rather, it demonstrates that the framework provides a unified probabilistic foundation that encompasses existing choice-theoretic formulations.

    • First, counting processes provide a general stochastic framework for modeling the temporal occurrence of interaction events. Within this framework, time-varying recommendation models can be naturally incorporated and interpreted in terms of event intensities and conditional choice probabilities.

    • Second, our study characterizes the popularity effect through the counting processes governing user-item interactions. This formulation allows us to identify an inherent identifiability issue in separating the effects underlying observed interactions. Furthermore, we demonstrate how model misspecification can induce systematic bias in the estimated interaction effects, leading to unintended distortions in the inferred user-item preferences.

    • Finally, the proposed framework provides a foundation for further methodological and theoretical developments. In particular, it can potentially accommodate more complex observation mechanisms, including censoring and truncation. Moreover, the martingale structure associated with counting processes offers a natural route toward establishing asymptotic properties of the resulting estimators using Martingale-based asymptotic theory. As suggested by the reviewer, we will explicitly discuss and clarify the advantages of using a counting-process framework for modeling recommendation systems in the conclusion of the revised manuscript.

[1] Rizoiu, M. A., Xie, L., Sanner, S., Cebrian, M., Yu, H., & Van Hentenryck, P. (2017, April). Expecting to be hip: Hawkes intensity processes for social media popularity. In Proceedings of the 26th international conference on world wide web (pp. 735-744).

Thank you for your detailed responses. I have no further questions and have decided to increase my rating.

We sincerely thank Reviewer Tguc for the thoughtful review and constructive discussion. Your insightful comments helped us clarify the theoretical foundations of our work. We also greatly appreciate your careful consideration of our responses and your willingness to reconsider your assessment.

Thank you again for your time and valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.