acceptodds
Under review as a conference paper at ICLR 2027

Distributionally Robust Decision-Making under Reject Inference Context

Abstract

Reject inference arises when outcomes are observed only after acceptance decisions, yielding selectively labeled data that are shaped by historical policies. While existing methods mainly improve nominal learning under such one-sided feedback, they are often vulnerable to deployment shifts that alter the true conditional success probabilities. We therefore propose a distributionally robust framework for policy learning under the reject inference context. The objective is to find a robust policy that maximizes the worst‑case discounted reward, with the ambiguity set characterized as a Kullback–Leibler ball centered at the posterior beliefs. We derive a dynamic programming relation for the worst‑case discounted reward and introduce a practical backward approximation algorithm to compute the optimal robust policy. Extensive experiments on synthetic and real lending data show that the proposed method consistently improves robustness and lower-tail performance relative to competitive baselines, at only a modest cost in nominal reward.

Primary area:
optimization
Keywords:
reject inference, distribution shift, distributionally robust optimization, robust dynamic programming

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_ui1GOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies robust policy learning for the reject inference problem under the setting of distribution shift. The authors use a distributionally robust optimization framework with a Kullback–Leibler ambiguity set to learn policies that are robust to such shifts. Theoretically, the author establish a robust Bellman equation to characterize the optimal policy, contraction and truncation error bounds for the backward approximation of the value function, and estimation error of the effective discount factor. Simulation and read data application verify the performance of the proposed framework over baselines.

Critical strengths

The paper is motivated by distribution shift in the reject inference problem, which is practically relevant, particularly in sequential decision-making. In my opinion, using DRO to address this distributional uncertainty is a natural and appropriate approach. The authors frame the problem well, deriving a robust Bellman equation and develop a corresponding backward approximation algorithm within this framework, which represents a substantial technical effort.

Critical weaknesses

My major concern is about the convergence of the learn policy. Although the authors establish some theoretical guarantees, the key result of convergence to the true optimal policy is not explicitly proved. In classical DRO, one should define the true underlying distribution and show that the distribution at the center of the ambiguity set converges to it. However, the authors do not establish this convergence for their posterior-based center, as well as the convergence of the policy learnt.

Moreover, the radius should be assumed to shrink to zero at an appropriate rate as the center converges to the true distribution. In experiment, with a fixed radius, the learnt policy could still sit in the neighbour from the true policy, even as the center converges to the true distribution.

Questions for authors

Could the authors clarify what the true underlying distribution is and whether the chosen center, posterio mean, of the ambiguity set converges to it? as well as the covergen guarantee of the learnt policy.

Another common choice for constructing the ambiguity set is the Wasserstein distance. Could the authors clarify the motivation for choosing KL? Perhaps a Wasserstein-based formulation could be more computationally friendly?

I would also be interested in seeing the computational cost compared with the non-DRO case, as the additional optimization required by DRO could substantially increase the computational burden.

I am also concerned about the an overly conservative decision policy. Could the radius adapt over time based on a measure of uncertainty or variability?

1 comment
AuthorsOpenReview ↗

Response to Reviewer ui1G

We thank the reviewer for the questions. We have uploaded the revised manuscript, with all revisions highlighted in blue, to address your concerns and would greatly appreciate your reconsideration.

Q1 and W1. We have added Proposition 4 in Section 3.4 (pp. 6–7), with a complete proof in Appendix B.6 (pp. 23–29) to prove convergence of the learned policy. The main result is as follows. If the underlying environment is stationary and each state receives sufficiently many accepted observations, the posterior estimate of the success probability converges to its true value. When the ambiguity radius and estimated effective discount factor also converge, the corresponding ambiguity sets and robust value functions converge. Furthermore, provided that the optimal decision is not exactly at the acceptance threshold, we prove that the learned policy eventually makes the same decisions as the corresponding oracle policy.

We distinguish between two convergence objectives. With a fixed positive ambiguity radius, the learned policy converges to the robust oracle, which may differ from the nominal optimal policy because it continues to protect against distributional uncertainty. When the radius shrinks to zero, the learned policy instead converges to the nominal oracle based on the true success probabilities.

We also provide an explicit error bound and a possible shrinking-radius schedule in Corollary 9, together with numerical approximation conditions in Corollary 10. Finally, we emphasize that sufficient feedback is essential: if a state is persistently rejected, its posterior may not converge. Our consistency guarantee therefore holds under sufficient feedback rather than unconditionally for the original policy.

Q2. We use KL divergence because it provides a natural way to describe uncertainty in Bernoulli success probabilities, allowing the range of possible deviations to depend on the current estimated probability. Nevertheless, we agree that Wasserstein distance is a reasonable alternative and can be computationally simpler. To address this question, we have added Appendix B.7 (p. 29), which explains how Wasserstein ambiguity can be incorporated into our robust dynamic programming framework. In particular, the corresponding worst-case probabilities can be computed directly without the numerical root-finding required by KL divergence. We have also added an experimental comparison in Appendix C.7 (Table 14). The results show that Wasserstein achieves slightly higher nominal reward (96.55 versus 95.03), whereas KL performs slightly better under a severe distribution shift (44.26 versus 43.87). These findings suggest that our proposed framework is not restricted to KL divergence, while different ambiguity measures may lead to different trade-offs between nominal performance and robustness.

Q3. We have added a direct comparison between DRRI and the nominal dynamic programming method SL-OPT in Appendix C.11 (Table 18, pp. 46–47). The results show that DRRI requires only slightly more computation than SL-OPT. Specifically, DRRI takes 0.896 seconds per replication, compared with 0.864 seconds for SL-OPT, representing a 3.7% increase in runtime. Their memory usage is also similar (197.2 versus 194.8 MiB). This relatively small difference is partly because our implementation precomputes the worst-case probability bounds, avoiding repeated optimization during policy evaluation. We also investigate how runtime and memory usage change with the number of states, the belief-grid size, and the truncation depth in Table 17. These results show that computational costs increase with problem size.

Q4. We agree that a fixed ambiguity radius may lead to overly conservative decisions, particularly as more observations become available and uncertainty about the success probabilities decreases. To address this concern, we have added an uncertainty-adaptive ambiguity radius in Appendix B.8, which gradually decreases as more accepted observations are collected while maintaining a minimum level of protection against potential distribution shifts. The adaptive formulation can be incorporated into our existing robust dynamic programming framework without changing its fundamental structure.

We have also added experiments comparing fixed and adaptive ambiguity radii in Appendix C.8 (Table 15). The results show that adaptive radii can improve nominal performance while retaining much of the robustness of the fixed-radius approach. For example, an adaptive radius with a minimum value of 0.005 increases nominal reward from 95.03 to 98.15, while achieving comparable performance under severe shifts (42.81 versus 44.26). However, allowing the radius to decay toward zero improves nominal reward further (103.88) but substantially reduces performance under severe shifts (37.92). These results demonstrate that adaptive radii offer a flexible way to balance nominal performance and robustness, rather than uniformly improving both.

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

This paper studies distributionally robust sequential decision-making under selective labels, where outcomes are observed only after acceptance. The authors maintain Beta–Bernoulli belief states and introduce KL ambiguity sets to account for deviations in conditional success probabilities. They derive a robust Bellman equation, develop a backward approximation algorithm, and establish bounds for truncation and effective-discount estimation errors. Experiments on synthetic environments and FICO data evaluate the trade-off between nominal utility and downside performance.

Critical strengths
  1. The paper addresses a practically relevant problem with a meaningful robust extension.

Selective feedback makes current decisions affect both immediate rewards and future information acquisition. The additional uncertainty from deployment shifts is particularly relevant in applications such as lending. The proposed formulation incorporates these effects into a sequential robust optimization problem, where the adverse success probability governs both the immediate reward and subsequent belief transitions. The resulting policy accounts for information acquisition and stopping decisions under distributional uncertainty.

  1. The proposed framework is computationally tractable and supported by extensive experiments.

The paper develops a state-dependent robust Bellman recursion and a practical backward approximation scheme, together with truncation and effective-discount estimation guarantees. The experimental evaluation includes the direct nominal baseline SL-OPT, several structured adverse shifts, and sensitivity analyses covering the robustness radius, cost, discount factor, truncation depth, and terminal condition. The FICO experiments also report improvements in lower-tail returns across different preprocessing configurations. These results provide encouraging evidence for the practical value of the proposed approach.

Critical weaknesses
  1. The adversarial-probability characterization in Theorem 1 contains an inconsistency.

Expanding the robust acceptance objective gives

q\left[1+\bar\gamma(V^+-V^-)\right].$$ Under the sign convention used in the paper's exponential-tilting formula, the corresponding coefficient should be $$D=\bar\gamma(V^- -V^+)-1.$$ However, Theorem 1 (p. 3) defines $D=\bar\gamma(V^+-V^-)-1$, which may lead to an incorrect selection of the KL endpoint. 2. The contribution of the robust sequential extension and the scope of the experimental evidence could be clarified. The framework builds on the selective-label dynamic programming formulation of Wei (2021) and established results in robust dynamic programming. Its main contribution lies in incorporating KL-based distributional uncertainty into this setting. The practical significance of this extension is therefore an important aspect of the evaluation. The experiments provide useful evidence under adverse conditions, but the connection between the theoretical robust objective and the evaluation protocols deserves further explanation. Equation (2) considers an adaptive adversary whose ambiguity set depends on the evolving nominal belief. Section 4.3 evaluates fixed perturbations of the underlying success probabilities, while the FICO experiments primarily examine variations induced by random replay ordering. Clarifying the robustness properties assessed by each experiment would help establish the scope of the empirical conclusions. The FICO evaluation protocol also requires clarification. Appendix C.3 indicates that test-set performance is used during feature selection and discretization, but the relationship between these records and the final replay data is unspecified. A clear description of the data splits and selection procedure is needed to assess the independence of the reported evaluation. Finally, the experiments could provide more direct evidence of the value of sequential information acquisition within the robust policy. A comparison with a simple KL-pessimistic acceptance rule, or an analysis of decisions driven by continuation value, would help quantify the contribution of the full dynamic programming approach.
Questions for authors
  1. Could the authors reconcile the definition of DD and the KL endpoint-selection rules in Theorem 1, Appendix B.3, and Algorithms 1–2? Please also confirm which implementation was used to produce the reported experimental results.

  2. Could the authors clarify the data splits used for FICO preprocessing, feature and bin selection, robustness-radius selection, and final online evaluation? In particular, were the final replay records independent of the data used for label-dependent model selection?

  3. Could the authors clarify the relationship between the theoretical adaptive KL formulation and the robustness evaluated in the experiments? It would also be helpful to quantify the benefit of the full robust DP over a simple pessimistic policy such as 1qL(μ;η)>c\mathbf{1}{q_L(\mu;\eta)>c}, using matched priors and warm starts or an analysis of continuation-driven decisions.

3 comments
AuthorsOpenReview ↗

Response to Reviewer KKXi

We thank the reviewer for the careful feedback. We have uploaded the revised manuscript, with all revisions highlighted in blue, to address your concerns and would greatly appreciate your reconsideration.

1. Robustness motivation and technical contribution.

Our primary objective differs from Wei (2021). While Wei (2021) studies optimal online decision-making under selective labels in a nominal environment, our work focuses on robustness against distributional shifts. Such shifts affect both immediate rewards and future information acquisition, while selective feedback limits the ability to detect and correct model misspecification. The revised Introduction (pp. 1–2) emphasizes this distinct objective. We acknowledge that the Bayesian updates, finite-domain decomposition, and backward approximation build on Wei (2021).

Our contributions address three technical challenges. First, Theorem 1/Appendix B.3 characterizes a robust Bellman recursion in which the adversarial probability jointly determines immediate rewards and successor-belief transitions, rather than merely adjusting immediate rewards. Second, Propositions 2–3/Appendices B.4–B.5 establish approximation and estimation guarantees for the robust formulation, including their implications for decision stability. Third, new Section 3.4/Proposition 4 and Appendix B.6 establish conditional posterior, ambiguity-set, value, and policy consistency under selective feedback, requiring sufficient feedback and a strict decision margin. Appendix A.3 further clarifies the relationship to existing robust MDP/POMDP literature.

2. Sign correction and implementation.

We thank the reviewer for identifying these inconsistencies. We have corrected the sign convention and reconciled the definition and KL endpoint-selection rules in Theorem 1, Appendix B.3, and Algorithms 1–2. As explained in Appendix D, Algorithm 1 requires solving the KL-constrained optimization through a Lagrange multiplier, whereas Algorithm 2 provides an equivalent but more efficient implementation by precomputing the KL interval endpoints. We use Algorithm 2 in our experiments to reduce computational costs.

3. FICO splits and radius selection.

New Appendix C.9 (pp. 44–45) specifies disjoint development, radius-validation, and final replay samples. Feature ranking/elimination and bin selection use development data and its internal hold-out; AUCs of 0.790 and 0.777 are selection diagnostics, not independent final-performance estimates. A separate 30% validation sample selects ηtrain=0.0005\eta_{\mathrm{train}}=0.0005. Final replay records are excluded from these label-dependent choices. The representation and radius are frozen before evaluation; only accepted outcomes update beliefs. The 1,000 replays vary orderings of one cohort, not independent borrower samples.

4. Adaptive theory versus experimental robustness.

Equation (2) considers a posterior-adaptive adversary, where the success probability jointly affects immediate rewards and future belief transitions. Our experiments examine the resulting robust policy under complementary evaluation settings. In Section 4.3, the success probabilities are shifted to a fixed KL-based adverse environment. Although the shift is not adaptively selected at each belief state, it still affects both realized rewards and the observed success/failure outcomes that drive subsequent posterior updates, thereby preserving the central interaction between distributional uncertainty and information acquisition. Section 4.4 further examines global, boundary, and subgroup perturbations to assess robustness under differently structured shifts. These experiments evaluate the policy's resilience to adverse deployment conditions rather than directly computing the adaptive minimax value in Equation (2). The FICO study provides complementary evidence on downside performance across online replay trajectories, without imposing explicit repayment-probability shifts. We have clarified the respective objectives and scope of these evaluations in Sections 4.3–4.4 and Appendix C.9.

5. Full robust DP versus KL-LCB.

We added KL-LCB, 1{qL(μ;ηtrain)>c}\mathbf{1}\{q_L(\mu;\eta_{\mathrm{train}})>c\}, to Appendix C.2 and Tables 1–3/Figure 1, with matched priors, radius, posterior updates, and warm starts. It removes the continuation term while retaining KL pessimism. DRRI can accept despite nonpositive immediate worst-case reward when future information justifies the cost. At B0=0B_0=0, Table 1 gives 151.17 versus 119.49 for DRRI and KL-LCB. Table 2 gives 95.03 versus 74.14 nominally and 51.68 versus 43.52 at ηeval=0.06\eta_{\mathrm{eval}}=0.06; FICO mean rewards are 59.353 versus 49.853.

Reviewer_KKXiOpenReview ↗

Response to Authors

Thank you for the detailed response and the revised manuscript. The corrected sign convention, the clarification of the relationship to Wei (2021) and robust dynamic programming, and the additional KL-LCB comparison address my main concerns. In particular, the new comparison provides useful evidence for the value of retaining continuation effects under the tested settings. I also appreciate the clearer distinction between the adaptive theoretical objective and the synthetic and FICO evaluations. Overall, I view the work as a useful incremental contribution, and I maintain my rating of 3 (weak acceptance).

AuthorsOpenReview ↗

Response to Reviewer KKXi

Thanks for your prompt reply and appreciation.

Reviewer_e7CWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies online accept/reject decisions under selective labels, where outcomes are observed only for accepted individuals. Building on the Bayesian dynamic-programming formulation of Wei (2021), it introduces robustness through a Bernoulli KL ambiguity set around the current belief mean. At each step, an adversary selects the success probability within this set, affecting both immediate reward and subsequent belief transitions.

For finite covariate spaces, the problem decomposes into state-wise dynamic programs with effective discount factors. The authors approximate the solution using a truncated backward recursion and provide approximation-error bounds. Empirically, DRRI sacrifices nominal reward but performs better under sufficiently large downward shifts in success probabilities. On FICO, it achieves the second-highest mean reward and the highest reported lower-tail quantiles across random arrival orders.

Critical strengths

S1. Relevance and practicality.

Robust decision-making under selective labels is relevant in applications such as lending. The proposed method provides a tractable backward-recursion procedure for the finite state spaces considered.

S2. Extensive empirical evaluation.

The experiments cover a wide range of baselines and evaluate robustness under both KL-based shifts and several structured logit perturbations.

Critical weaknesses

W1. Limited novelty.

The Bayesian formulation, belief updates, finite-domain decomposition, and backward approximation build directly on Wei (2021). The main addition is the per-step KL worst case over the success probability, and I believe this is incremental.

W2. Modest gains on real data.

On FICO, DRRI achieves the second-highest mean reward. Although it achieves the best reported lower-tail quantiles, the improvements over the strongest competing baseline at each quantile are around ~ 0.6–2.3%. These gains are modest. Moreover, the FICO replay varies arrival order without introducing an explicit shift in repayment probabilities, so it does not directly evaluate robustness to such shifts.

Questions for authors

Q1. Could the authors clarify the main technical contribution beyond introducing a per-step KL worst case into Wei (2021)’s dynamic program? Which parts of the theoretical analysis require new arguments?

Q2. How were the training ηtrain\eta_{\mathrm{train}} selected for the synthetic and FICO experiments? Is it tuned? I am asking because it seems very important and can erase or reverse DRRI’s advantage as seen by comparing Table 5 with Table 2 (performance goes below the best performing baseline once eta exceeds a certain value).

1 comment
AuthorsOpenReview ↗

Response to Reviewer e7CW

We thank the reviewer for the constructive feedback. We have uploaded the revised manuscript, with all revisions highlighted in blue, to address your concerns and would greatly appreciate your reconsideration.

Q1 and W1 Our objective differs fundamentally from Wei (2021). While Wei focuses on optimal online decision-making under selective labels, balancing immediate rewards against future information, our work addresses the robustness of such sequential decisions against distributional shifts. This distinction is important in many real applications, such as lending, where changes in economic conditions may cause actual repayment probabilities to deviate substantially from historical beliefs. Such deviations affect not only immediate lending losses but also the reliability of information acquired for future decisions. Moreover, rejected applicants provide no feedback, limiting the decision-maker's ability to detect and correct model misspecification. Thus, policies optimized for nominal online learning may become vulnerable under adverse deployment conditions, motivating our focus on maximizing worst-case discounted reward. The revised Introduction (pp. 1–2) explicitly emphasizes this distinct research objective. Addressing it presents three key technical challenges.

First, robustness changes the valuation of information. The adversarial success probability simultaneously controls immediate rewards and the probabilities of successor beliefs. The acceptance objective is inf⁡q∈U(μ;η){q−c+γˉ(x)[qV++(1−q)V−]}\inf_{q\in\mathcal U(\mu;\eta)}\{q-c+\bar\gamma(x)[qV^++(1-q)V^-]\}, where V+V^+ and V−V^- are success/failure continuation values. A reward-only adjustment retaining nominal transitions misses uncertainty in future feedback. Theorem 1/Appendix B.3 characterizes the coupled minimax recursion and KL-constrained adversarial response. Setting η=0\eta=0 recovers Wei's nominal formulation.

Second, robust planning requires error control. The unbounded belief count and unknown arrival frequencies introduce computational and estimation errors into this minimax problem. Proposition 2 extends, rather than originates, backward-truncation analysis to the robust operator. Proposition 3/Appendix B.5 quantifies effective-discount estimation error, its propagation to robust values, and action stability under a decision margin. These guarantees assess whether numerical and statistical errors can alter robust decisions.

Third, selective feedback obstructs statistical consistency. New Section 3.4/Proposition 4 and Appendix B.6 connect adaptively acquired labels to posterior, ambiguity-set, value, and policy convergence. The arguments use accepted-label martingales, joint continuity of KL endpoints under varying centers/radii, and vanishing belief-update effects. Acceptance-score convergence, not value convergence alone, establishes eventual agreement under a strict margin. Sufficient feedback is required; positive limiting radii target a robust oracle, while vanishing radii recover the nominal oracle.

W2. We agree that DRRI does not achieve the highest mean reward on the FICO dataset. However, we would like to clarify that the primary objective of distributionally robust decision-making is not to maximize nominal expected reward, but to protect decision quality against adverse deviations from the nominal environment. This distinction is particularly important in lending applications, where policies that perform well on average may still be vulnerable to unfavorable outcomes when the underlying repayment probabilities deteriorate. We therefore do not expect DRRI to uniformly outperform nominal methods in terms of average reward, especially when the deployment environment closely resembles the training environment. This trade-off is explicitly demonstrated in our simulation experiments. As shown in Table 2, when the distribution shift is absent or relatively mild (ηeval≤0.02\eta_{\mathrm{eval}}\leq 0.02), nominal methods such as CL, Thompson Sampling, and SL-OPT outperform DRRI. This is consistent with their objectives, which emphasize nominal reward optimization. However, as the deviation from the training environment becomes more severe, their performance deteriorates more rapidly, whereas DRRI exhibits greater stability.

Q2. New Appendix C.10/Table 16 tests coarser quantile bins, supervised tree bins, and an alternative feature pair; DRRI retains higher lower-tail quantiles than R-osl across all three. Appendix C.3 uses a synthetic reference radius of 0.01, not a claimed independently validated optimum. New Appendix C.9 reports FICO selection of 0.0005 on a separate 30% validation sample excluded from replay. Radii remain fixed across test shifts. We agree that Table 5 exposes excessive conservatism and performance reversals: robustness is not improved by arbitrarily enlarging the radius. Appendix C.5.1 documents this limitation, and new Appendix C.8/Table 15 evaluates adaptive alternatives without claiming uniform dominance.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Responses to Reviewers

Dear Reviewers and Area Chair,

Thank you for your careful reading and constructive feedback. We have revised the manuscript to strengthen the theoretical analysis, add the requested comparisons, and clarify the experimental protocol and scope of our claims. For ease of reference, we summarize the main changes below.

  1. Contribution and technical corrections. We revised the Introduction to clarify how distributional uncertainty affects both immediate rewards and future posterior transitions, and how DRRI accounts for this coupling in sequential decision-making under selective feedback. We also corrected the sign convention in Theorem 1 and the corresponding endpoint-selection explanation in Appendix B.3.

  2. Consistency guarantees. Section 3.4 (Proposition 4) and Appendix B.6 now establish posterior, robust-value, and policy consistency under explicit assumptions, including a stationary reference model and sufficient accepted-label feedback. Under a strict decision margin, the policy eventually agrees with its limiting oracle. We distinguish the robust oracle for a positive limiting radius from the nominal oracle when the radius vanishes, and explicitly acknowledge that the original policy does not automatically guarantee sufficient feedback. Additional results address numerical approximation and diminishing forced exploration.

  3. Comparison with pessimistic myopic decisions. We added KL-LCB, which uses a KL-pessimistic success probability but omits continuation value (Appendix C.2). DRRI improves upon this comparator in the reported nominal, shifted-environment, and FICO evaluations (Tables 1–3). This comparison directly examines the benefit of sequential planning beyond pessimistic immediate rewards.

  4. Alternative ambiguity sets and adaptive radii. Appendices B.7–B.8 and C.7–C.8 provide formulations and experiments for Wasserstein ambiguity and uncertainty-adaptive radii. KL and Wasserstein yield comparable performance in these experiments. The adaptive-radius results clarify the trade-off between nominal utility and protection against severe shifts, including the role of a positive radius floor as posterior uncertainty decreases.

  5. FICO evaluation and robustness-radius selection. Appendix C.9 clarifies the separation of development, radius-validation, and final evaluation data, with preprocessing and radius choices frozen before replay. We also added sensitivity analyses for feature selection and discretization (Appendix C.10), which preserve the observed lower-tail advantage over R-osl. We explicitly clarify that the FICO results characterize replay performance conditional on the evaluation cohort, while the synthetic experiments examine imposed changes in success probabilities.

  6. Computational overhead. We added a direct comparison with nominal SL-OPT (Appendix C.11, Table 18). In the reported synthetic benchmark, fixed-radius DRRI requires 0.896 seconds versus 0.864 seconds for SL-OPT, corresponding to approximately 3.7% additional runtime, with comparable memory usage.

We hope these revisions make the contribution, supporting evidence, and limitations clearer. If your schedule permits, we would be grateful for a brief follow-up on whether the revisions address your main concerns and whether any points would benefit from further clarification. We remain available to respond promptly during the discussion. Thank you again for your time and thoughtful consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.