acceptodds
Under review as a conference paper at ICLR 2027

On Advantage Estimates for Max@K Policy Gradients

Abstract

Reinforcement learning with verifiable rewards is widely used for post-training reasoning models, but sparse outcome rewards make exploration difficult. A complementary approach is to optimize inference-time objectives such as pass@ and max@ directly, yet existing policy-gradient estimators for these objectives use different signals, baselines, and normalizations, making their relationships unclear. We study this issue through baseline design and advantage centering. Starting from the advantage estimator of a leading method in the field, we show that it is policy-gradient unbiased but yields a non-centered advantage. We then introduce a Leave-Two-Out baseline that preserves policy-gradient unbiasedness while making realized batch advantages exactly centered. The resulting method, MaxPO, has an efficient quadratic-time implementation and integrates naturally into group-based RL for LLM post-training. We further derive the canonical finite-batch advantage for max@, providing a unified view of existing advantage estimators. Empirically, we verify that the L2O baseline reduces gradient variance and outperforms non-centered alternatives.

Primary area:
reinforcement learning
Keywords:
Reinforcement Learning, Policy Optimization, Large Language Models, Variance Reduction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_rGJ3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MaxPO introduces a leave-two-out baseline for max@K policy gradients, producing an unbiased estimator with exactly centered advantages. It derives connections to existing estimators and evaluates variance in controlled environments and performance on mathematical reasoning and code generation.

Critical strengths
  1. The analysis clarifies which existing max@K estimators share the same centered signal and how standard-deviation normalization changes their unbiasedness properties.
Critical weaknesses
  1. Average gains are modest. Averaging the task-average scores across both models in Table 2, improvements over the strongest baseline at each model and sampling budget are approximately 2.1, 2.0, and 2.5 percentage points at pass@8, pass@32, and pass@256. The table does not report uncertainty across training runs.

  2. The benefit over closely related prior methods is untested. Appendices F.5 and F.6 connect MaxPO to specific unnormalized or mean-centered forms of [1] and [2], but neither method is directly evaluated. It remains unclear whether MaxPO's estimator offers practical gains over their existing formulations.

    [1] Chen et al. Pass@k Training for Adaptively Balancing Exploration and Exploitation of Large Reasoning Models. arXiv:2508.10751, 2025. [2] Bagirov et al. The Best of N Worlds: Aligning Reinforcement Learning with Best-of-N Sampling via max@k Optimisation. arXiv:2510.23393, 2025.

Questions for authors
  1. Do the average accuracy gains persist across independent training seeds?
  2. Under matched training and tuning budgets, how does MaxPO compare with [1] and [2], and how much does the normalization choice affect performance?
3 comments
AuthorsOpenReview ↗

Thank you for the review

Thank you for the review and for carefully reading our work.

W1. Average gains are modest. Averaging the task-average scores across both models in Table 2, improvements over the strongest baseline at each model and sampling budget are approximately 2.1, 2.0, and 2.5 percentage points at pass@8, pass@32, and pass@256.

The results on coding tasks in the appendix show greater gains (Table 6, Figure 11). Compared to GRPO: 2.6, 4.9, 5.3 at pass@(8,32,256) respectively. Also, to elaborate on your claim regarding Table 2, the improvements on the Qwen-7B model are up to around 4-5% points, and up to around 2% points for the Llama-3B model. So the performance may be model-dependent, and on some models the improvements appear significant.

Q1. Do the average accuracy gains persist across independent training seeds?

We agree that checking persistence is important, and to address the concern we ran 2 extra seeds on the Qwen 7B model with math tasks (so a total of 3 seeds; the extra experiments cost around 2000USD).

Methodk=1248163264128256
GRPO41.7 ± 0.547.0 ± 0.851.4 ± 1.155.1 ± 1.458.7 ± 1.762.4 ± 1.866.1 ± 2.069.7 ± 2.173.4 ± 2.1
Entropy-Adv41.0 ± 1.346.3 ± 1.351.0 ± 1.455.1 ± 1.459.0 ± 1.462.7 ± 1.366.6 ± 1.270.3 ± 1.373.8 ± 1.4
PKPO39.0 ± 0.944.1 ± 0.848.6 ± 0.752.7 ± 0.756.5 ± 0.860.2 ± 0.664.0 ± 0.667.9 ± 0.671.8 ± 0.6
MaxPO40.5 ± 0.446.8 ± 0.452.0 ± 0.556.6 ± 0.560.7 ± 0.564.6 ± 0.668.4 ± 0.672.1 ± 0.775.7 ± 0.9

Values are mean ± standard error. We added the table to App J, and will replace the results in the paper when accepted.

The results show that gains remain consistent for higher k values, but were perhaps overestimated in the previous experiments. An important observation to us was the comparison with PKPO. Namely, the improvements over PKPO remain consistent across all 3 seeds, and MaxPO always outperformed it. This provides supporting evidence to our main claim that the centered baseline outperforms the one from PKPO.

In practice, it is quite difficult to achieve fully statistically significant results in LLM training due to the cost. We believe the additional corroborative evidence (lowered gradient variance, toy experiments) provide a strong case that our main claim of improving over the non-centered PKPO is valid.

Q2. and W2. Under matched training and tuning budgets, how does MaxPO compare with [1] and [2], and how much does the normalization choice affect performance?

As a preliminary we note that Bagirov’s method was developed concurrently to ours, and that Chen’s work is limited to binary rewards (and their derivation does not extend to continuous rewards). Both are currently only available as arxiv papers and the ICLR reviewer policy notes that lack of comparisons to arxiv papers cannot be a basis for rejection, but authors are strongly encouraged to discuss such works (https://iclr.cc/Conferences/2027/ReviewerGuidelines). We believe the current discussion fits within these guidelines.

A bit more background:

Compared to Bagirov: While one of the baseline techniques that they use ends up proportional to ours, the magnitude of their estimator remains biased w.r.t. the correct magnitude. Moreover, they do not perform experiments with this variant. Their main experimental method divides the advantages with the std of the output advantages. However, this biases the estimator. For example, if the rewards are binary, then the normalization actually makes the normalized advantages identical to the advantages of GRPO (except that when there are sufficiently many correct responses in the batch, then the advantage will be 0 for Bagirov, but not for GRPO). Thus the estimator is not theoretically unbiased and it remains unclear how related to the max@k objective it is.

Chen’s work similarly has an std normalization, and the connection to max@k optimization in that case is not clear. Furthermore, Chen’s work is algorithmically limited to binary rewards, whereas ours works with continuous rewards. We also have coding experiments with non-binary pass rate rewards where Chen’s method is not applicable (Table 6, Figure 11). In practice, we believe the performance on binary reward tasks will be similar.

The main comparison algorithm is PKPO that is also unbiased, but their baseline technique is uncentered. Our experiments examined both the gradient variance and the performance, and provided convincing evidence that MaxPO outperforms PKPO and the gradient variance is reduced.

Again, thank you and do not hesitate to ask if there are any remaining concerns or questions.

Thanks for the additional experiments and for clarifying the relationship to prior work. Both increased my confidence in the results, and I’ve raised my score to 3.

(And section 5.1 describes variance reduction as a theoretical property. Are there conditions under which the leave-two-out baseline provably reduces policy-gradient variance relative to PKPO, or should I read the reduction as an empirical finding? My understanding is that Proposition 3.1 establishes exact centering, which does not by itself imply lower gradient variance.

AuthorsOpenReview ↗

Thank you for the response

Thank you for the follow-up and for increasing the score. The other reviewers also asked about this point.

A precise sufficient condition is E[∥g∥2A~i]≥0,\mathbb{E}[\lVert g\rVert^2\tilde{A}_i] \ge 0, where g=∇log⁡π(x)g = \nabla \log \pi(x) and A~i\tilde{A}_i is the centered advantage of MaxPO.

The optimal variance reduction baseline (https://arxiv.org/abs/1301.2315) is given by E[(∇log⁡π(x))2r(x)]/E[(∇log⁡π(x))2]\mathbb{E}[(\nabla \log \pi(x))^2 r(x)] / \mathbb{E}[(\nabla \log \pi(x))^2], i.e., it is a weighted average of the different rewards, where the weighting is proportional to the square of the log-derivative. If we are agnostic to the magnitude of the log-derivative (e.g., assume that it is constant or independent of the reward), then the centered expected reward gives the minimum variance (i.e., our centered baseline gives the minimum variance). The main comparison method, PKPO, also has a baseline that is agnostic to the log-derivative terms. While in practice, it can happen that log-derivative terms align effectively so that PKPO has lower variance, in a typical setting, our approach will be better. Note also that this optimal baseline technique is rarely used in practice because it has not been shown to work better than the centered baseline in practice. Our experiments also showed that MaxPO has a lower gradient variance than PKPO both in toy experiments, and in the LLM experiments (Fig. 1,3,4,5,7)

A simple intuition for why centering helps is the following:

Consider advantages that are offset from the center by bb so that Ai=A~i+bA_i = \tilde{A}_i + b, where A~i\tilde{A}_i are centered advantages.

Let g=∇θlog⁡π(x)g=\nabla_\theta\log\pi(x). It is known that E[g]=0\mathbb{E}[g]=0 and assume E[A~i]=0\mathbb{E}[\tilde{A}_i]=0, and that ∥g∥2\lVert g\rVert^2 and A~i\tilde{A}_i are uncorrelated.

V[gAi]=V[g(A~i+b)]=V[gA~i]+2bE[∥g∥2A~i]+b2V[g].\mathbb{V}[g A_i] = \mathbb{V}[g(\tilde{A}_i+b)] = \mathbb{V}[g\tilde{A}_i] + 2b\mathbb{E}[\lVert g\rVert^2\tilde{A}_i] + b^2\mathbb{V}[g].

By uncorrelation and centering, the cross term vanishes:

E[∥g∥2A~i]=E[∥g∥2]E[A~i]=0.\mathbb{E}[\lVert g\rVert^2\tilde{A}_i] = \mathbb{E}[\lVert g\rVert^2]\mathbb{E}[\tilde{A}_i] = 0.

Therefore,

V[gAi]=V[gA~i]+b2V[g].\mathbb{V}[g A_i] = \mathbb{V}[g\tilde{A}_i] + b^2\mathbb{V}[g].

The offset thus adds an extra variance of b2V[g]b^2\mathbb{V}[g]. A centered baseline removes this offset.

More generally, because the advantages of PKPO are always bigger than the centered advantage, we have b≥0b \ge 0, and thus we do not need the cross term to vanish, but a sufficient condition is E[∥g∥2A~i]≥0\mathbb{E}[\lVert g\rVert^2\tilde{A}_i] \ge 0.

We added a paragraph about this at the end of Section 2.1. We agree that this improves the clarity and thank you for the suggestion.

Reviewer_nK4iOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MaxPO, a max@K policy optimization method that introduces a Leave-Two-Out (L2O) baseline to obtain unbiased policy gradients with exactly centered advantages, reducing gradient variance and improving LLM reasoning performance.

Critical strengths
  • The paper tackles an important problem of reducing gradient variance in max@K policy optimization.

  • The paper provides a theoretical analysis of advantage estimation and introduces a principled L2O baseline with unbiasedness and exact centering guarantees.

Critical weaknesses
  • The paper treats exact batch-level advantage centering as a key principle for variance reduction. However, centering advantages does not necessarily minimize the variance of the resulting policy-gradient estimator, which also depends on score-function magnitudes and cross-sample covariance. Can the authors establish sufficient conditions under which L2O provably reduces gradient variance relative to EI-only? Alternatively, can they construct settings where L2O increases variance?

  • Table 1 establishes that several existing pass@K/max@K estimators recover the same canonical advantage up to constant scaling or normalization. If these methods already induce equivalent policy-gradient directions, what is the substantive algorithmic novelty of MaxPO beyond an alternative derivation? More importantly, why are these theoretically equivalent estimators excluded from the main empirical comparisons? Without such comparisons, how can the claimed empirical improvements be attributed to a genuinely new estimator rather than simply correcting the known non-centered EI baseline?

  • The unbiasedness guarantee is derived for an on-policy sequence-level policy-gradient estimator, whereas the actual LLM experiments employ token-level PPO-style clipping, importance ratios, and KL regularization. Under these modifications, the implemented update is no longer an unbiased estimator of the original max@K policy gradient in general. Can the authors characterize the bias introduced by the practical training objective and demonstrate whether the theoretical advantage of L2O persists under clipping? In particular, could the observed performance gains arise from interactions with the surrogate objective rather than from unbiased advantage estimation itself?

Questions for authors

The authors should address the points listed in the weakness section.

5 comments
AuthorsOpenReview ↗

Thank you for the review 1/2

Thank you for the review and for carefully reading our work.

W1. The paper treats exact batch-level advantage centering as a key principle for variance reduction. However, centering advantages does not necessarily minimize the variance of the resulting policy-gradient estimator, which also depends on score-function magnitudes and cross-sample covariance. Can the authors establish sufficient conditions under which L2O provably reduces gradient variance relative to EI-only? Alternatively, can they construct settings where L2O increases variance?

The optimal variance reduction baseline (https://arxiv.org/abs/1301.2315) is given by E[(∇log⁡π(x))2r(x)]/E[(∇log⁡π(x))2]\mathbb{E}[(\nabla \log \pi(x))^2 r(x)] / \mathbb{E}[(\nabla \log \pi(x))^2], i.e., it is a weighted average of the different rewards, where the weighting is proportional to the square of the log-derivative. If we are agnostic to the magnitude of the log-derivative (e.g., assume that it is constant or independent of the reward), then the centered expected reward gives the minimum variance (i.e., our centered baseline gives the minimum variance). The main comparison method, PKPO, also has a baseline that is agnostic to the log-derivative terms. While in practice, it can happen that log-derivative terms align effectively so that PKPO has lower variance, in a typical setting, our approach will be better. Note also that this optimal baseline technique is rarely used in practice because it has not been shown to work better than the centered baseline in practice. Our experiments also showed that MaxPO has a lower gradient variance than PKPO both in toy experiments, and in the LLM experiments (Fig. 1,3,4,5,7)

A simple intuition for why centering helps is the following:

Consider advantages that are offset from the center by bb so that Ai=A~i+bA_i = \tilde{A}_i + b, where A~i\tilde{A}_i are centered advantages.

Let g=∇θlog⁡π(x)g=\nabla_\theta\log\pi(x). It is known that E[g]=0\mathbb{E}[g]=0 and assume E[A~i]=0\mathbb{E}[\tilde{A}_i]=0, and that ∥g∥2\lVert g\rVert^2 and A~i\tilde{A}_i are uncorrelated.

V[gAi]=V[g(A~i+b)]=V[gA~i]+2bE[∥g∥2A~i]+b2V[g].\mathbb{V}[g A_i] = \mathbb{V}[g(\tilde{A}_i+b)] = \mathbb{V}[g\tilde{A}_i] + 2b\mathbb{E}[\lVert g\rVert^2\tilde{A}_i] + b^2\mathbb{V}[g].

By uncorrelation and centering, the cross term vanishes:

E[∥g∥2A~i]=E[∥g∥2]E[A~i]=0.\mathbb{E}[\lVert g\rVert^2\tilde{A}_i] = \mathbb{E}[\lVert g\rVert^2]\mathbb{E}[\tilde{A}_i] = 0.

Therefore,

V[gAi]=V[gA~i]+b2V[g].\mathbb{V}[g A_i] = \mathbb{V}[g\tilde{A}_i] + b^2\mathbb{V}[g].

The offset thus adds an extra variance of b2V[g]b^2\mathbb{V}[g]. A centered baseline removes this offset.

We added a paragraph about this at the end of Section 2.1. We agree that this improves the clarity and thank you for the suggestion.

AuthorsOpenReview ↗

Thank you for the review 2/2

W2. Table 1 establishes that several existing pass@K/max@K estimators recover the same canonical advantage up to constant scaling or normalization. If these methods already induce equivalent policy-gradient directions, what is the substantive algorithmic novelty of MaxPO beyond an alternative derivation? More importantly, why are these theoretically equivalent estimators excluded from the main empirical comparisons?

As a preliminary note that Bagirov’s method was developed concurrently to ours, and that Chen’s work is limited to binary rewards (and their derivation does not extend to continuous rewards). Both are currently only available as arxiv papers and the ICLR reviewer policy notes that lack of comparisons to arxiv papers cannot be a basis for rejection, but authors are strongly encouraged to discuss such works (https://iclr.cc/Conferences/2027/ReviewerGuidelines). We believe the current discussion fits within these guidelines.

A bit more background:

Compared to Bagirov: While one of the baseline techniques that they use ends up proportional to ours, the magnitude of their estimator remains biased w.r.t. the correct magnitude. Moreover, they do not perform experiments with this variant. Their main experimental method divides the advantages with the std of the output advantages. However, this biases the estimator, and may be theoretically unsound. For example, if the rewards are binary, then the normalization actually makes the normalized advantages identical to the advantages of GRPO (except when there are sufficient correct responses in the batch, then the advantage is 0 for Bagirov, but not GRPO). Thus the estimator is not theoretically unbiased and it remains unclear how related to the max@k objective it is.

Chen’s work similarly has an std normalization, and the connection to max@k optimization in that case is not clear. Furthermore, Chen’s work is algorithmically limited to binary rewards, whereas ours works with continuous rewards. We also have coding experiments with non-binary pass rate rewards where Chen’s method is not applicable (Table 6, Figure 11).

The main comparison algorithm is PKPO that is also unbiased, but their baseline technique is uncentered. Our experiments examined both the gradient variance and the performance, and provided convincing evidence that MaxPO outperforms PKPO and the gradient variance is reduced.

how can the claimed empirical improvements be attributed to a genuinely new estimator rather than simply correcting the known non-centered EI baseline

Our main claim is correcting the non-centering. Past works similarly build on a base estimator with improvements. The relationship to previous works is: Koyamada et al. (2022) created the base estimator for max@k objectives and applied it in RL -> Tang et al. (ICML2025) applied it to LLM tasks and proposed a new baseline that uses the k-1 other samples to construct a baseline b(x−i)b(x_{-i}) -> Walder et al. (NeurIPS2025) created PKPO that uses a similar baseline as Tang, but allows having a batch size N > k, and the estimator is computed by averaging Tang’s method over all subsets of size k. This coincides with the expected improvement formulation considered in our paper. Our work builds on this by adding the L2O baseline to center this estimator and reduce the gradient variance. An important observation is that whatever baseline function you plug into Tang’s method b(x−i)b(x_{-i}), by applying the L2O correction to this baseline, the resulting estimator ends up being the same canonical form of estimator given by us. Thus, this estimator constructs a fundamental centered advantage.

W3 The unbiasedness guarantee is derived for an on-policy sequence-level policy-gradient estimator, whereas the actual LLM experiments employ token-level PPO-style clipping, importance ratios, and KL regularization. Under these modifications, the implemented update is no longer an unbiased estimator of the original max@K policy gradient in general. Can the authors characterize the bias introduced by the practical training objective and demonstrate whether the theoretical advantage of L2O persists under clipping? In particular, could the observed performance gains arise from interactions with the surrogate objective rather than from unbiased advantage estimation itself?

This is an important point to consider. To address it, we checked the logs, and clipping is almost never activated (clip fraction <0.1%). Moreover, there was no KL regularization. Our view is that the comparison with PKPO is a matched comparison that only changes the baseline. The results show greater performance and reduced gradient variance (Fig. 3). Additionally, the toy experiments (Fig. 1, 4, 5, 7) are based on experiments with the simple REINFORCE method that do not have the implementation details that you were concerned about, and there is a clear advantage to MaxPO over PKPO in terms of both learning performance and gradient variance.

Thank you.

Thank you again for your thoughtful review. As the discussion period is nearing its end, we would appreciate it if you could review our responses and additional results. We hope they address your concerns, and we would be grateful if you could reconsider your assessment.

Reviewer_jU2ZOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies advantage estimation for max@K policy gradients and proposes MaxPO, which augments an expected-improvement estimator with a Leave-Two-Out baseline. The construction preserves policy-gradient unbiasedness while ensuring that advantages sum to zero within every sampled group. The paper also derives a canonical finite-batch advantage and uses it to clarify relationships among existing estimators. The main contributions are: (1) An L2O baseline that avoids the implicit sample dependence introduced by applying ordinary Leave-One-Out directly to EI scores. (2) A unified interpretation of max@K advantages that connects several existing methods through baseline choices, constant scaling, and normalization. (3) An (O(B^2)) implementation and empirical evaluation on bandits, maze environments, mathematical reasoning, and code generation with continuous rewards.

Critical strengths
  1. clear conceptual distinctions. It carefully separates policy-gradient unbiasedness, zero-mean advantages, and exact within-group centering.

  2. practical integration. MaxPO replaces the advantage computation within existing group-based RL pipelines without requiring an additional learned value model.

  3. good empirical performance. On five mathematical reasoning benchmarks, MaxPO improves average pass@256 over PKPO from 70.6% to 74.3% on Qwen2.5-Math-7B and from 63.5% to 65.0% on Llama-3.2-3B-Instruct.

Critical weaknesses
  1. Exact centering does not guarantee lower gradient variance. the paper should characterize when variance reduction is expected.
Questions for authors

When does centering reduce gradient variance? Can you characterize conditions under which L2O reduces variance and qualify the corresponding claims?

2 comments
AuthorsOpenReview ↗

Thank you for the review

Thank you for the review and for carefully reading our work.

W1. Exact centering does not guarantee lower gradient variance. the paper should characterize when variance reduction is expected.

Q1. When does centering reduce gradient variance? Can you characterize conditions under which L2O reduces variance and qualify the corresponding claims?

The optimal variance reduction baseline (https://arxiv.org/abs/1301.2315) is given by E[(∇log⁡π(x))2r(x)]/E[(∇log⁡π(x))2]\mathbb{E}[(\nabla \log \pi(x))^2 r(x)] / \mathbb{E}[(\nabla \log \pi(x))^2], i.e., it is a weighted average of the different rewards, where the weighting is proportional to the square of the log-derivative. If we are agnostic to the magnitude of the log-derivative (e.g., assume that it is constant or independent of the reward), then the centered expected reward gives the minimum variance (i.e., our centered baseline gives the minimum variance). The main comparison method, PKPO, also has a baseline that is agnostic to the log-derivative terms. While in practice, it can happen that log-derivative terms align effectively so that PKPO has lower variance, in a typical setting, our approach will be better. Note also that this optimal baseline technique is rarely used in practice because it has not been shown to work better than the centered baseline in practice. Our experiments also showed that MaxPO has a lower gradient variance than PKPO both in toy experiments, and in the LLM experiments (Fig. 1,3,4,5,7)

A simple intuition for why centering helps is the following:

Consider advantages that are offset from the center by bb so that Ai=A~i+bA_i = \tilde{A}_i + b, where A~i\tilde{A}_i are centered advantages.

Let g=∇θlog⁡π(x)g=\nabla_\theta\log\pi(x). It is known that E[g]=0\mathbb{E}[g]=0 and assume E[A~i]=0\mathbb{E}[\tilde{A}_i]=0, and that ∥g∥2\lVert g\rVert^2 and A~i\tilde{A}_i are uncorrelated.

V[gAi]=V[g(A~i+b)]=V[gA~i]+2bE[∥g∥2A~i]+b2V[g].\mathbb{V}[g A_i] = \mathbb{V}[g(\tilde{A}_i+b)] = \mathbb{V}[g\tilde{A}_i] + 2b\mathbb{E}[\lVert g\rVert^2\tilde{A}_i] + b^2\mathbb{V}[g].

By uncorrelation and centering, the cross term vanishes:

E[∥g∥2A~i]=E[∥g∥2]E[A~i]=0.\mathbb{E}[\lVert g\rVert^2\tilde{A}_i] = \mathbb{E}[\lVert g\rVert^2]\mathbb{E}[\tilde{A}_i] = 0.

Therefore,

V[gAi]=V[gA~i]+b2V[g].\mathbb{V}[g A_i] = \mathbb{V}[g\tilde{A}_i] + b^2\mathbb{V}[g].

The offset thus adds an extra variance of b2V[g]b^2\mathbb{V}[g]. A centered baseline removes this offset.

We added a paragraph about this at the end of Section 2.1. We agree that this improves the clarity and thank you for the suggestion.

Do not hesitate to ask if you have any remaining concerns or questions.

Thank you again for your thoughtful review. As the discussion period is nearing its end, we would appreciate it if you could review our responses and additional results. We hope they address your concerns, and we would be grateful if you could reconsider your assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.