acceptodds
Under review as a conference paper at ICLR 2027

Efficient Policy-Space Regression via Reparameterized Optimal Targets

Abstract

Supervised ranking-based methods such as Direct Preference Optimization (DPO) have shown that bypassing complex RL dynamics can improve the stability and efficiency of post-training. However, these methods impose supervision primarily in the reward space, and their guarantees do not generally ensure alignment of the model policy with the optimal policy in policy space, potentially introducing bias. To address this issue, we propose Minimum Cross-entropy Policy Optimization (MCPO), which performs regression directly in the policy space. MCPO first reparameterizes the optimal policy into a more tractable target, and then regresses the policy toward it using a low-variance cross-entropy-style objective. Theoretically, we show that the optimal policy remains a stationary point of the MCPO objective, preserving the desired optimum under practical variance control. Moreover, under idealized tabular softmax models, MCPO guarantees monotonic progress toward the optimal policy when clipping is non-degenerate; without clipping, it further yields the closest single finite-step update to the optimal policy among all objectives. Empirically, MCPO consistently outperforms all baselines across all evaluated settings. In preference alignment, it achieves better performance using only 50% of the training steps, establishing an efficient paradigm for LLM post-training.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, reinforcement learning, reasoning, alignment

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_DY2cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

PPO and GRPO rely on local policy gradient updates, which may require many steps to approach the optimal policy, leading to inefficient training. This paper proposes a different approach by leveraging the closed-form optimal policy in KL-regularized RL. Since the optimal policy cannot be computed directly, the authors use reparameterization to construct a tractable training target from rollout rewards. Instead of following local policy gradients, the method trains the policy to match this target using cross-entropy loss, aiming to achieve faster and more efficient convergence.

Critical strengths
  1. The motivation is clear and well justified, as local policy gradient updates in PPO/GRPO may not efficiently move the policy toward the optimum. Policy-space regression is a natural and intuitive way to address this limitation.

  2. The authors use the closed-form optimal policy in KL-regularized RL and reparameterize it into a tractable training target without enumerating the entire action space. This offers an interesting alternative to standard policy gradient methods.

  3. MCPO achieves better final performance, faster convergence, and more stable training than GRPO in the reported experiments. The faster reward improvement is consistent with the paper’s motivation, though it does not directly establish faster convergence to the optimal policy.

Critical weaknesses

I like the paper’s motivation, but it is not yet clear whether MCPO’s empirical gains come from the core mechanism of policy-space regression. Lemma 1, Theorem 2, and Figure 1 support more direct progress toward the optimal policy in an idealized tabular setting, while the LLM experiments only show faster reward improvement. More evidence is needed to connect the theoretical advantage to the gains observed in LLM training.

Questions for authors

Q1. The theory and Figure 1 suggest that MCPO approaches the optimal policy faster, but the LLM experiments only show faster reward improvement. Could you compare DKL(π∗∥πθ)D_{\mathrm{KL}}(\pi^{*} \parallel \pi_{\theta}) for MCPO and PPO/GRPO in a neural network setting where the optimal policy is known?

Q2. MCPO changes several components at once, including the objective, reward weighting, baseline, and clipping, while the current ablations mainly examine the clipping threshold and �0��0�. Could you ablate these components using the same rollouts and training budget to clarify what drives the gains?

Q3. MCPO treats a full response as one action, while actions in multi-turn tasks affect future states and rewards. How would you construct the optimal policy target in this setting? A discussion or a small multi-turn experiment would help clarify its applicability.

1 comment
AuthorsOpenReview ↗

Response to Reviewer DY2c

Thank you for your suggestion regarding the source of the gains! Our response is as follows.

Weakness & Q1

Thank you for pointing out the need for a more direct evaluation metric. Since estimating $ D_ {KL}(\pi^ *|| \pi) $ requires sampling from $ \pi^ * $ , which is generally intractable to compute accurately, we initially did not include this metric and instead reported only the rewards. We have now supplemented our experiments with $ D_{KL}(\pi^ *| | \pi) $ comparisons in the neural network setting, as well as an indirect comparison of $ D_{KL}(\pi^ * || \pi) $ in our LLM experiments.

1. Comparison of $ D_{KL}(\pi^*||\pi) $ in the Neural Network

Thank you for suggesting experiments in a "neural network setting where the optimal policy is known."

We constructed state features using standard Gaussian distributions, with 4 states and 2 dimensions per state. We used a two-layer MLP with ReLU activations as the policy network, where the final layer outputs 2 numbers that are passed through a softmax function to obtain the policy. Rewards were sampled from a Bernoulli distribution, the reference policy was uniform, and $ \beta $ was set to 0.1.

Under this setting, we trained GRPO and MCPO from the same initialization, using the same learning rate of $ 1\mathrm{e}{-2} $ for 1,000 iterations.

The results are as follows:

  • GRPO: The reward increased from 1.80 to 2.19, while $ D_{KL}(\pi^*||\pi) $ decreased from 0.86 to 0.74.
  • MCPO: The reward increased from 1.80 to 2.69, while $ D_{KL}(\pi^*||\pi) $ decreased from 0.86 to 0.48.

We repeated the experiments with multiple random seeds and observed consistent results. We believe these results demonstrate the advantages of MCPO, at least in the neural network setting. We will also incorporate these experimental results into the next version of the paper.

2. indirect comparison in LLM

Although $ D_{KL}(\pi^*||\pi) $ is difficult to measure directly, we can decompose it as follows:

$ D_ {KL}(\pi^ *||\pi) = \sum_ a \pi^ *(a) \log \frac{\pi^ *}{\pi} = \frac{1}{\mathbb{E}_ {\pi _ {ref}}[e^ {\frac{r}{\beta}}]} \left( \mathbb{E} _ {\pi _ {ref}} [e^ {\frac{r}{\beta}} \log \pi _ {ref} + \frac{r}{\beta} e^ {\frac{r}{\beta}} ] \right) - \log \mathbb{E} _ {\pi _ {ref}}[e^ {\frac{r}{\beta}}] - \frac{1}{\mathbb{E}_ {\pi _ {ref}}} \mathbb{E} _ {\pi _ {ref}}[e^ {\frac{r}{\beta}} \log \pi] $

where the only term that differs across different policies π\pi is Eπref[erβlog⁡π]\mathbb{E} _{\pi _{ref}}[e^{\frac{r}{\beta}} \log \pi], which can be estimated by sampling from πref\pi _{ref}.

Therefore, we calculated Eπref[erβlog⁡π]\mathbb{E} _{\pi _{ref}}[e^{\frac{r}{\beta}} \log \pi] for the Llama-8B model trained with different methods on ArenaHard. The results for GRPO, BNPO, DPO, and MCPO are -1791749, -1518047, -1062195, and -2042607, respectively. MCPO achieves the smallest value, indicating that its Eπref[erβlog⁡π]\mathbb{E} _{\pi _{ref}}[e^{\frac{r}{\beta}} \log \pi] is also the smallest.

Q2: Ablation Study

Since MCPO introduces a new objective function, we conducted ablation studies on the ArenaHard v2 benchmark using Llama-8B to investigate the contributions of its other components. Specifically, we successively removed the baseline and clipping mechanisms, and replaced reward weighting with advantage weighting.

MethodWR2LC2
MCPO56.1071.22
w/o baseline54.8364.21
w/o clipping45.0549.82
w/o reward weighting50.9857.22

We observe that removing the clipping mechanism leads to the most significant performance degradation, possibly due to increased variance. All these components contribute to MCPO, with the full MCPO achieving the best overall performance.

Q3: actions in multi-turn tasks

Thank you for considering the multi-turn setting. In fact, we previously attempted to extend MCPO to the token level, where the weight of each token is computed independently, which is also similar to the multi-turn setting.

However, we were unable to obtain theoretically satisfactory results. This is because estimating the target requires independently estimating the exponential of the expected future return for each token, which is generally difficult to do without bias. An exception is when we have perfectly accurate process rewards, in which case each token only needs to maximize its own process reward. We can then treat each token as an action and re-derive Eq. 8 accordingly.

However, perfectly accurate process rewards are unavailable in most cases. We also experimented with directly using the return of the entire sequence as the reward for each token and applying Eq. 8 directly. The resulting performance was very similar to that of sequence-level MCPO, with a difference of only approximately one point. Nevertheless, since sequence-level MCPO has a more theoretically sound formulation, we ultimately chose to present the sequence-level version.

We will incorporate the above discussion into the revised version of the paper.

Reviewer_bK8AOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MCPO, which regresses the policy toward a reparameterized optimal target using a clipped cross-entropy objective. It shows that π* is a stationary point of the objective and gives tabular-softmax convergence results. It reports gains over RL and reward-space regression baselines on RLHF, RLVR and offline RL. The main weaknesses are in motivation and theory. The central claim, that reward-space regression cannot guarantee policy-space alignment, ignores prior work that does give such guarantees, so the related-work coverage is insufficient and the motivation is weakened.

Critical strengths
  1. The proposed method is simple. The target replaces the intractable partition function with a per-prompt batch-mean reward, so it is computable, and the change is a small modification to an existing GRPO-style trainer.
  2. The theory is reasonable and consistent although it relies on the tabular model assumption. The paper shows that π* is a stationary point of the clipped objective (Thm 1) and gives a tabular-softmax progress analysis (Thm 2, Lemmas 3–4), which addresses the reward-space bias of DPO-style methods.
  3. The evaluation is broad. It covers RLHF, RLVR and offline RL, with several model families and scales, ablations on clipping and β, and training-curve and gradient-variance analyses.
Critical weaknesses
  1. Policy optimization with optimal-policy convergence guarantee is well studied over the years since DPO. The paper clearly does not have enough coverage of the related work. Most importantly, the argument that "reward-space regression makes alignment in the policy space difficult to guarantee" that the proposed method is motivated from is more of a lack of thorough literature review than an established fact. There are many methods that optimize the policy in the reward-space with guaranteed alignment in the policy space, including but not limited to ([1] for online/offline training and [2] for offline training): [1] Zhang et al., GVPO: Group Variance Policy Optimization for Large Language Model Post-Training, Neurips 2025 [2] Chen et al., Noise Contrastive Alignment of Language Models with Explicit Rewards, Neurips 2024

  2. The theory relies on the tabular model assumption, which oversimplifies the essence that LLM is a generative model with "infinite" instead of "finite" possible candidate responses. Instead, [1][2]'s theories work on the LLM setting natively.

Questions for authors

To address my concerns raised in the weaknesses, can the authors conduct additional experiments on the methods used in [1] and [2] as stronger baselines?

3 comments
AuthorsOpenReview ↗

Response to Reviewer bK8A

Thank you for suggesting additional related work. We have supplemented our discussion and experimental comparisons with these works.

W1 & Q: Related Work

We appreciate the reviewer's suggestions regarding related work. We have read the two recommended papers and added both discussions and experimental comparisons.

1. Discussion

We have carefully studied both papers. Although both claim to guarantee convergence in the policy space to the optimal policy, we identified potential gaps in the rigor of their proofs.

Both papers adopt a similar proof strategy: they establish that the optimal policy is the unique global minimizer of their respective objective functions. However, since LLMs are generally optimized using gradient-based methods, and policy gradients have the unique property that $ \mathbb{E}[\nabla_\theta\log\pi_\theta]=0, $

if the errors in these two methods are constant across all actions, the policy gradient will vanish. This means that although minimizing their objective functions theoretically guarantees unbiased convergence at the objective level, the actual optimization process may stop at a policy where the errors are equal across all actions, rather than at the global optimum, due to the properties of policy gradients.

From this perspective, these methods do not strictly guarantee unbiased convergence of the policy. In contrast, our method directly analyzes policy convergence at the gradient level, providing a more rigorous guarantee.

2. Experiments

We have added experimental results for GVPO [1] and NCA [2] on both of our main experimental settings.

Llama3.1-8B-Instruct:

MethodAlpacaEval-WR2AlpacaEval-LC2AlpacaEval-WR1ArenaHard-GPTArenaHard-Gemini
GVPO50.9860.8497.5259.2029.14
NCA46.3656.2894.1757.6021.59
MCPO56.1071.2299.0288.0059.22

Mistral-7B-Instruct:

MethodAlpacaEval-WR2AlpacaEval-LC2AlpacaEval-WR1ArenaHard-GPTArenaHard-Gemini
GVPO30.5038.3493.2736.6015.20
NCA24.4331.4691.6726.0013.90
MCPO57.0250.3595.4148.2021.26

MCPO outperforms both methods in these experiments.

Summary

Once again, we thank the reviewer for pointing us to these related works. We will incorporate the above discussion and experimental results into the related work and experimental sections of the revised paper, and explicitly cite both papers.

W2: theory relies on the tabular

Thank you for considering the conditions of our theorem. In fact, the convergence-to-optimal-policy guarantee of MCPO (Theorem 1) does not require the tabular assumption. Its assumptions are the same as those in [1–2], making it applicable to LLMs. Only the subsequent results concerning convergence rates require the tabular assumption.

We will further clarify this distinction in the revised version.

Thank you for adding the experiments on the two mentioned related baselines. Although I am not able to validate the results myself, I believe the reported discrepancies between them and the MCPO is significant enough for an empirical justification. Plus, your clarification of the tabular assumption addresses my concern. However, the rest of the theoretical part is where I am still less convinced. If I am not mistaken, MCPO also suffers a discrepancy between theory and practice when we are limited by finite sampling (as reviewer RjBw also mentioned). Therefore, I do not think that [1-2]'s theoretical claims are compromised just because there can be "equal-across-action errors" due to practical limitations. From my perspective, MCPO is actually in the same category as [1-2] in principle, except that it is a better performing one. That said, I am now leaning toward a weak accept given the additional content.

AuthorsOpenReview ↗

Response to Reviewer bK8A

Thank you for the positive assessment! We agree that MCPO is also subject to issues such as finite-sample effects, and we will acknowledge these limitations in the revised version.

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

This paper proposes Minimum Cross-entropy Policy Optimization (MCPO) for LLM post-training. It constructs an unnormalized target from the reference policy and centered rewards, then fits the policy using a cross-entropy-style objective with detached, clipped weights. The paper claims preservation of the optimal policy as a stationary point, monotonic improvement under an idealized tabular model, and optimal one-step progress without clipping. Validation includes theoretical derivations and experiments on online preference alignment, verifiable-reward reasoning, and offline preference learning. The reported results show improved performance across these settings and comparable preference-alignment performance with fewer training steps.

Critical strengths

Critical Strengths

  1. A concrete policy-space regression formulation. MCPO provides an intuitive way to construct training targets from a reference policy and rewards, with a specific clipped update objective. This formulation is potentially useful for developing practical alternatives for LLM policy optimization, even if the theoretical guarantees require revision.
  2. Promising empirical coverage. The experiments span online preference alignment, verifiable-reward reasoning, and offline preference learning across several model families. The reported improvements, particularly in online preference alignment, provide evidence of potential practical value beyond a single task or model.
Critical weaknesses
  1. The stationary-point guarantee does not cover the same-batch estimator described in the paper.

    Theorem 1 treats ρs=Z(s)/Z^(s)\rho_s=Z(s)/\hat{Z}(s) as fixed when taking the action expectation. However, Eq. (4), Algorithm 1, and Appendix B (p. 14, lines 729–736) use the same generated responses to estimate both the normalization factor and the gradient. These quantities are statistically dependent; stopping gradients does not remove that dependence.

    Consider one state and two actions with πref=(1/2,1/2)\pi_{\mathrm{ref}}=(1/2,1/2), rewards (0,log⁡3)(0,\log 3), and β=1\beta=1, giving π⋆=(1/4,3/4)\pi^{\star}=(1/4,3/4). Set π=πold=π⋆\pi=\pi_{\mathrm{old}}=\pi^{\star}, G=3G=3, and clipping thresholds [−0.28,0.30][-0.28,0.30]. For K∼Binomial(3,3/4)K\sim\mathrm{Binomial}(3,3/4), the common batch weight is

    w(K)=clip⁡(2⋅3−K/3−1,−0.28,0.30).w(K)=\operatorname{clip}(2\cdot 3^{-K/3}-1,-0.28,0.30).

    The expected negative loss gradient with respect to the higher-reward action's logit, with the other logit fixed, is

    \approx -0.04927147 \ne 0.$$ Thus, the expected update is nonzero at the true optimum under the stated procedure. The fixed-target population calculation can still hold, but it does not establish the claimed optimality preservation for this estimator. This affects a central justification for MCPO and may require changing the estimator or substantially revising the guarantee.
  2. The policy-improvement claims are not established by the supplied proofs.

    Lemma 1 claims optimal one-step progress among all objectives. Yet Eqs. (46)–(47) discard O(η2)O(\eta^2) terms and introduce a fixed weighted-norm constraint, Eπ[g2]=τ\mathbb{E}_\pi[g^2]=\tau, absent from the lemma statement. This establishes at most a constrained first-order result, rather than exact finite-step optimality. Without an explicit normalization convention, rescaling an objective also changes its effective update magnitude at the same learning rate.

    Separately, Eq. (21) on p. 16 incorrectly expands a log moment-generating function using central moments rather than cumulants. For π=(1/2,1/2)\pi=(1/2,1/2), π⋆=(3/4,1/4)\pi^{\star}=(3/4,1/4), and ρ=1\rho=1, the exact improvement is

    =\eta/4-\eta^2/8+\eta^4/192+O(\eta^6),$$ whereas Eq. (21) gives the fourth-order term $-\eta^4/384$. The downstream derivations relying on this identity therefore require correction. These observations do not rule out appropriately restricted results or alternative proofs, but they leave central theoretical contributions unsupported in their current form.
Questions for authors
  1. Estimator clarification: Does the implementation reuse the same responses for target normalization and gradient estimation, as stated in Appendix B? If so, how is the nonzero expected gradient above reconciled with the stationary-point claim? If a different estimator is used, what is its precise definition and which guarantee applies to it?

  2. Scope and validity of the theoretical results: Is Lemma 1 intended as a first-order statement under a fixed weighted-norm constraint? Can the authors provide corrected statements and derivations, including a replacement for Eq. (21), and identify which policy-improvement and convergence guarantees remain valid?

1 comment
AuthorsOpenReview ↗

Response to Reviewer RjBw

We sincerely thank the reviewer for the careful review. We address the questions raised below.

W1 & Q1: The stationary-point guarantee

Thank you for pointing out the lack of independence between normalization and gradient estimation. We acknowledge that the dependence arises because the same samples are shared by both procedures, which introduces bias into the gradient estimates and, consequently, leads to instability at the fixed point.

This sample reuse is essentially analogous to the practice in GRPO, where a single reward sample is used both to estimate the expected reward and to compute the policy gradient. Although this introduces bias, it improves sample efficiency. In practice, the performance gains resulting from variance reduction outweigh the bias introduced by sample reuse. Nevertheless, we acknowledge that this design leads to biased gradient estimates, and we will explicitly highlight this theoretical-practical discrepancy in the revised version of the paper.

We have also added experiments comparing MCPO with "a different estimator." Specifically, we adopt a leave-one-out strategy, in which the normalization for each sample within a group is estimated using the remaining rollouts in that group, while keeping all other aspects unchanged from MCPO. The experimental results are presented below:

MethodLlama-WR2Llama-LC2Mistral-WR2Mistral-LC2
MCPO56.1071.2257.0250.35
Leave-one-out53.4957.9646.3644.08

As shown in the table, the leave-one-out estimator does not lead to improved performance. We conjecture that this is due to the increased variance introduced by using fewer samples for normalization.

Nevertheless, we will explicitly acknowledge the bias introduced by sample reuse in the gradient estimator as a limitation of our method in the revised version. We will also incorporate the above discussion and the experimental comparison with the unbiased estimator into the Discussion section.

W2 & Q2: Scope and validity of the theoretical results

Thank you for pointing out issues with the details of our derivation.

1. Lemma 1

Thank you for pointing out the lack of rigor in the statement. We agree that Lemma 1 is intended as a first-order statement under a fixed weighted-norm constraint. We will accordingly weaken the claims in the revised version of the paper to better reflect the scope of this result.

2. Eq. 21

Thank you for identifying the erroneous part of the derivation. In fact, the purpose of Eq. 21 is to establish the one-step monotonic convergence of MCPO in the tabular setting. We can circumvent the need for a full expansion and establish the result through an alternative approach, as follows:

  1. Assume that $ \frac{\pi^*(a)}{\pi(a)} \leq \tau$. It follows that $-1 \leq g(a) \leq \tau - 1 $ .

  2. Let M = \max\left\\{1, \tau - 1\right\\}. Then, we have ∣g∣≤M|g| \leq M.

  3. By Bernstein's exponential moment inequality, we obtain:

    $ \log \mathbb{E}_ {\pi_ k} \left[ \exp \left( \tilde{\eta}_ k \left( \frac{\pi^ *(a)}{\pi_ k(a)} - 1 \right) \right) \right] \leq \frac{\tilde{\eta}_ k^ 2 \upsilon_ 1(\pi^ *)}{2 \left( 1 - \frac{\tilde{\eta}_ k}{3} \max_ a \left| \frac{\pi^ *(a)}{\pi_ k(a)} - 1 \right| \right)} $

    This inequality holds under the condition η~kmax⁡a∣π∗(a)πk(a)−1∣<3\tilde{\eta}_k \max_a \left| \frac{\pi^*(a)}{\pi_k(a)} - 1 \right| < 3, which can be satisfied by choosing a sufficiently small step size.

  4. Finally, we obtain:

    $ \Delta(\pi _ {k+1}) \geq \tilde{\eta}_ k \upsilon_ 1(\pi^ *) \left(1 - \frac{\tilde{\eta}_ k}{2 \left(1 - \frac{\tilde{\eta}_ k}{3} \max_ a \left| \frac{\pi^ *(a)}{\pi_ k(a)} - 1 \right| \right)}\right)$

    Therefore, choosing a sufficiently small step size guarantees monotonic improvement.

However, since Lemma 3 is not used in the proof of the clipped version of MCPO, the impact of this modification is limited to Lemma 3. We will incorporate this modification into the revised version of the paper. The policy-improvement and convergence guarantees remain unchanged: for the tabular version of MCPO, monotonic improvement can still be guaranteed with a sufficiently small step size.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Additional Experiments for the Chairs

Dear Area Chairs, Senior Area Chairs, and Program Chairs,

Thank you for the time and effort you have devoted to coordinating and evaluating our submission. We have provided point-by-point responses to all weaknesses and questions raised in the three reviews, along with additional theoretical analyses and experiments to address the reviewers' key concerns. We appreciate the opportunity to present this additional evidence for your consideration.

Our work is motivated by the idea of direct regression in policy space, with several practical techniques incorporated to improve its effectiveness. To better bridge the gap between theory and practice and further strengthen the empirical evidence, we have focused on the following aspects in our rebuttal:

  1. Gradient estimation bias (Reviewer RjBw): We acknowledged the bias introduced by our gradient estimator and clarified that this is a practical design choice aimed at improving data efficiency. We also provided comparative experiments against a theoretically unbiased estimator and committed to weakening our stationary-point guarantee with respect to the practical objective function.

  2. Overstated theoretical claims and proof rigor (Reviewer RjBw): We acknowledged that some of our original theorems were stated too strongly and that certain parts of the reasoning lacked sufficient rigor. We have committed to appropriately weakening these claims and provided revised proof strategies that preserve the original conclusions.

  3. Related work (Reviewer bK8A): We supplemented our submission with additional experiments addressing the relevant prior work and will explicitly cite and discuss these works in the revised manuscript.

  4. Direct regression metrics for KL divergence (Reviewer DY2c): We added direct measurements in neural network settings, as well as indirect evaluation metrics in LLM scenarios. We further provided evidence demonstrating the validity of these indirect metrics.

  5. Ablation studies (Reviewer DY2c): We conducted additional ablation experiments following the reviewer's suggestions to further investigate the contributions of different components of our method.

  6. Multi-turn task scenarios (Reviewer DY2c): We added further discussion of the applicability and implications of our method in multi-turn task settings.

Overall, we believe that the additional theoretical analyses and experiments address the major concerns regarding the effectiveness of our method. We respectfully ask the chairs to evaluate our submission in light of this expanded theoretical and experimental evidence, together with our comprehensive responses to all reviewer comments.

Thank you again 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.