acceptodds
Under review as a conference paper at ICLR 2027

Optimality-Preserving Exploration Enhancement via Correctness-Conditioned Entropy

Abstract

Reinforcement Learning with Verifiable Rewards (RLVR) relies on exploration to discover high-quality solutions, yet the policy diversity of large language models (LLMs) often collapses early in training. Entropy regularization can alleviate this issue but is widely observed to reduce accuracy. Our geometric analysis attributes this tension to conflicting objectives: RLVR minimizes the projection distance to the optimal policy space, whereas entropy regularization pulls the policy toward a uniform distribution. Since the uniform distribution is generally not optimal, these two distances cannot be minimized simultaneously, implying an inherent conflict. We therefore propose Correctness-Conditioned Entropy (CCE) regularization, which regularizes the policy's projection onto the optimal policy space rather than the policy itself. Because the projection point and the projection distance are independent and can be optimized simultaneously, CCE regularization is optimality-preserving. We further show that RLVR with CCE is equivalent to variationally minimizing the Chernoff -divergence to the maximum-entropy optimal policy, providing a principled interpretation of CCE. Empirically, integrating CCE into DAPO and MaxRL yields consistent improvements across most mathematical reasoning benchmarks. When added to DAPO, CCE improves average accuracy by 1.3–2.0 points while enabling more targeted exploration, achieving simultaneous gains in exploration and optimality.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Reinforcement Learning, Reasoning, post-training

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_SBujOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  1. This paper targets at improving reinforcement learning with verifiable rewards, with emphasizes on the exploration stage during the training process.
  2. To improve the policy diversity, they propose correctness-conditioned entropy regularization, which regularizes the policy's projection onto the optimal policy space to preserve optimality.
  3. They also empirically investigate the effectiveness of the proposed method across multiple mathematical reasoning benchmarks, and demonstrate superiority compared with baseline methods such as DAPO.
Critical strengths
  1. Exploration in reinforcement learning is a critical part, this work introduce CCE to regularize this stage to enhance policy diversity is straightforward and meaningful, and also demonstrate superiority compared with previous methods.
  2. This work has provided solid theory prove for the optimality analysis of the regularization term, provides deeper understanding for readers.
Critical weaknesses
  1. One of the key weakness of this work is the novelty of the idea, entropy regularization has been widely investigated in reinforcement learning, for example an entropy regularization term is frequently used in model training.
  2. As shown in Table 1, the difference between the proposed method and previous DivPO is whether applied correctness condition.
Questions for authors

What about the effectiveness of the proposed method on models larger than 7B or MoE models, which could provide more additional insights.

6 comments
AuthorsOpenReview ↗

Response to Reviewer SBuj (1/3)

Thank you for the feedback! We have responded to all the weaknesses and questions raised in the review:

  1. Weakness 1 (novelty): Our goal is not simply to propose another method for "increasing diversity," as conventional approaches do, but rather to achieve this without compromising optimality. This perspective has not been explored by existing methods and represents a novel objective. We have also discussed and empirically compared against existing approaches, including entropy regularization.

  2. Weakness 2 (Distinction from DivPO): The fundamental difference between our method and DivPO lies in their objectives: CCE aims to increase diversity while preserving optimality, whereas methods such as DivPO inevitably compromise optimality to some extent. To achieve this objective, which existing methods cannot accomplish, we develop a different theoretical framework directly from the basic RL formulation (Section 4), without relying on existing regularization methods. Furthermore, to make CCE practical, we address several new challenges, which also lead to differences between CCE and DivPO in terms of gradient estimation methods, the placement in the loss function, and other key aspects. Therefore, CCE is an new approach rather than a variant of DivPO.

  3. Question (Larger Models): We have provided additional experimental results.

We have also incorporated the corresponding additions into the latest version of the PDF:

  • Weakness 1 & 2: Purple (not Red)
  • Questions: Pink

A Brief Overview of Our Work

Given the primarily theoretical nature of our work (i.e., it is task-agnostic, introduces a fundamental algorithmic improvement, and is theoretically motivated, as in [1–4]), its distinction from other methods should be understood from the perspective of its underlying theoretical approach. Since the reviewer's concerns mainly focus on our novelty and distinction from DivPO, we believe a brief overview of our work would be helpful. Below, we revisit the research approach underlying our work, which illustrates why it represents a new contribution rather than a simple variant of existing entropy regularization methods:

  1. Our objective and motivation are to improve diversity in RLVR without sacrificing optimality, rather than simply increasing diversity itself, which entropy regularization has already achieved (Abstract and paragraphs 1–2 of the Introduction).

    • Entropy regularization and other diversity-enhancing methods can increase model diversity, but often at the cost of accuracy.
    • Our goal is to improve diversity while preserving accuracy.
    • Achieving both diversity and optimality is a novel research objective that, to the best of our knowledge, has not been explored in prior work.
    • This also means that we cannot simply follow existing research directions.
  2. We theoretically identify why entropy regularization compromises optimality (Section 4.3). Based on this insight, we propose CCE regularization and prove that it preserves optimality (Section 4.4).

    • Because it cannot be jointly optimized with the RLVR objective.
    • CCE is also presented in Table 1. However, it cannot be directly incorporated into the loss due to its computational intractability, which motivates the subsequent development of our estimation methods.
  3. We establish that CCE regularization is equivalent to variationally encouraging the LLM to approach the optimal policy (Section 5), providing a theoretical interpretation of our method.

  4. Since CCE involves the distribution π(⋅∣A∗)\pi(\cdot|A^*), which cannot be directly computed, we derive and design an estimator for its gradient (Eq. 14).

  5. Given that RLVR relies on offline Monte Carlo estimation, we further develop an offline estimator for the CCE gradient (Eq. 16) to reuse RLVR rollouts, incorporate it into PPO-clip (Eq. 15), and provide theoretical justification for this approach (Theorem 3).

  6. Finally, we implement CCE in two algorithms, DAPO and MaxRL, to demonstrate its effectiveness (Section 7).

To summarize, we would like to emphasize two key distinctions:

  1. Our method is not simply "another diversity-enhancing method." Rather, it addresses the fundamental problem of existing diversity-enhancing methods: their tendency to compromise optimality. This represents a new objective.

  2. Our method is not merely a variant of existing entropy regularization approaches. Rather, it is a new method developed from a different objective through new theoretical analysis.

AuthorsOpenReview ↗

Response to Reviewer SBuj (2/3)

W1: novelty

Thank you for your concern regarding the novelty of our idea. However, unlike existing methods that aim simply to "increase diversity," our motivation is to achieve this without compromising optimality.

Specifically, existing methods all aim to optimize a conventional regularization term: max⁡θL(θ)+βH(θ),\max_\theta \mathcal{L}(\theta) + \beta \mathcal{H}(\theta), where the diversity metric $ \mathcal{H} $ typically prevents the optimality metric $ \mathcal{L} $ from reaching its optimum. In contrast, our objective is to maximize diversity while preserving optimality: max⁡θ:L(θ)=max⁡LH(θ).\max_{\theta:\mathcal{L}(\theta) = \max \mathcal{L}} \mathcal{H}(\theta). CCE in Table 1 is a method we identify that can achieve this objective, whereas conventional regularization methods cannot. Therefore, CCE arises from a different objective, rather than being a simple variant of existing methods.

Driven by this new objective, our method also takes a different theoretical framework from conventional methods. We did not develop our method by simply modifying certain components of DivPO or existing entropy regularization approaches. Instead, we first identified the geometric nature of regularization terms (Section 4), and then, from this geometric perspective, investigated what types of regularization are compatible with preserving optimality, which led us to find CCE. However, the conditional probabilities involved in CCE cannot be computed directly. We therefore further derived an unbiased estimator of the CCE gradient, incorporated it into an offline estimation framework for RLVR, and established error control guarantees. This process makes CCE a new method derived from a new theoretical framework for a new objective, rather than a simple variant of existing approaches.

Furthermore, we provide a variational regression perspective on CCE (Section 5), which offers an additional interpretability analysis from an new viewpoint.

Regarding conventional methods such as entropy regularization mentioned by the reviewer, we have cited them, discussed them in Section 4, and compared against them experimentally. These results demonstrate the superiority of CCE.

W2: difference with DivPO

Thank you for raising the question regarding the distinction between CCE and existing methods. However, CCE and DivPO differ fundamentally in their objectives, theoretical frameworks, underlying principles, and implementation strategies. As outlined in the brief overview of our work, we first formulate a different objective and then identify CCE through geometric analysis as a method that satisfies this objective, rather than simply modifying an existing approach.

1. Why CCE May Appear Similar to DivPO

The purpose of Table 1 is to provide an abstract representation of different methods rather than an exhaustive comparison. At this level of abstraction, the differences between various methods may appear relatively subtle. For example:

  • The difference in Table 1 between DivPO and entropy regularization is whether the expectation is conditioned;
  • The differences among MaxRL, GRPO, and Dr-GRPO mainly lie in how the advantages are normalized.

However, this does not mean that these methods are not fundamentally different from one another. They are built upon distinct theoretical frameworks, which provide the most fundamental basis for distinguishing.

2. Theoretical Differences of CCE

These are the most significant differences between our work and DivPO, and constitute the main contributions of our paper:

  1. Objective: Existing methods focus solely on increasing diversity, whereas CCE aims to increase diversity while simultaneously preserving optimality.

  2. Theoretical Framework: CCE first introduces a novel geometric perspective for characterizing different methods (Section 4), based on which we derive CCE as a regularization method that preserves optimality. This work does not follow any existing line of research.

  3. Interpretability: We prove that CCE regularization is equivalent to variationally minimizing the divergence from the optimal policy (Section 5), providing an interpretation that is absent from existing work.

  4. Estimation Reliability: Due to the intractability of directly computing the conditional probabilities, we derive an estimator for the CCE gradient and further establish an error-control method for offline estimation (Section 6). These are also novel contributions not found in existing methods.

AuthorsOpenReview ↗

Response to Reviewer SBuj (3/3)

3. Implementation Differences from DivPO

The different theoretical frameworks naturally lead to different implementation strategies:

  1. We propose directly estimating the CCE gradient, rather than the CCE term itself, through conditional sampling, which distinguishes our approach from existing methods (Eq. 14).
  2. To account for the behavior-target policy mismatch in RLVR, we develop an offline estimation method for the CCE gradient. (Eq.16) In contrast, existing methods generally overlook this aspect.
  3. To ensure the validity of offline estimation, we derive an upper bound on its estimation error (Theorem 3). Based on this analysis, we incorporate the CCE gradient into PPO-clip (Eq.15), which serves to control the offline estimation error. In contrast, other methods simply add a regularization term to the loss function.

This leads to substantial differences in their loss functions:

  • DivPO: Lppo(θ,A)+β1N∑i,t:ai∈A∗πθ(ai,t∣si,t)πold(ai,t∣si,t)log⁡πθ(ai,t∣si,t)\mathcal{L}_ \mathrm{ppo}(\theta, A) + \beta \frac{1}{N} {\textstyle \sum_ {i,t: a_i \in A^ *} \frac{\pi_ \theta(a_{i,t}|s_{i,t})}{\pi_ \text{old}(a_{i,t}|s_{i,t})} \log \pi_ \theta(a_{i,t}|s_{i,t})}
  • CCE (Eq.16):
A_ i - \beta \tfrac{N}{K_ i} \mathbb{I}_ {a_ i \in A^ *_ {s_ i}, K_ i \ge 2} \left ( \log \pi_ {\mathrm{old}}(a_ i|s_ i)- \tfrac{1}{K _ i-1} {\textstyle \sum_ {a_ j \in A^ *_ {s_i}, j \ne i, s_ i=s_ j} \log \pi_ {\mathrm{old}}(a_ j|s_ i) } \right )$$ # Question: Larger Model Thank you for the valuable suggestion to evaluate CCE on larger models. In response, we have made additional efforts to extend our experiments to Qwen2.5-14B-Instruct. The results are reported below: | Method | ProofWriter | RobustLR | RuleTaker | ToMi-NLI | Avg. | |---|---:|---:|---:|---:|---:| | DAPO | 92.30 | 71.20 | 46.90 | 12.20 | 55.65 | | DivPO | 92.20 | 69.80 | 48.00 | 13.50 | 55.88 | | **CCE** | **94.20** | **73.40** | **47.90** | **15.30** | **57.70** | As shown, CCE achieves the best average performance among all compared methods, consistently outperforming DAPO and DivPO on most benchmarks. These additional results provide further empirical evidence that CCE remains effective when scaling up the model size from our original experiments to 14B parameters, supporting its generalizability across different model scales. Due to the high computational cost of LLM training, models in the 7–8B range are already widely accepted as standard model sizes in existing research [1–4]. Therefore, we believe that our experiments provide reasonably convincing evidence regarding scalability across model sizes. Regarding MoE models, they typically have substantially more than 30B total parameters and differ fundamentally from conventional dense LLMs in terms of architecture and computational characteristics. Consequently, **policy optimization for MoE models is often investigated separately in specialized studies (e.g., [5–6]).** Our work, in contrast, focus on dense autoregressive policies under RLVR. **We will therefore include applying CCE to MoE models as directions for future works.** # Refs [1] https://proceedings.iclr.cc/paper_files/paper/2026/file/8587c39571f3e327725e2812749a9389-Paper-Conference.pdf (ICLR 2026) \ [2] https://proceedings.iclr.cc/paper_files/paper/2026/file/c073a9d002f3dab99a7af18642d74bdb-Paper-Conference.pdf (ICLR 2026) \ [3] https://proceedings.neurips.cc/paper_files/paper/2025/file/8808b4c72d5ade9bbcf9270ac6411314-Paper-Conference.pdf (NeurIPS 2025) \ [4] https://proceedings.neurips.cc/paper_files/paper/2025/file/b1ea3f93167c55f3fc2999b68a170ff3-Paper-Conference.pdf (NeurIPS 2025) \ [5] https://aclanthology.org/2026.acl-long.1165.pdf \ [6] https://aclanthology.org/2026.findings-acl.1944.pdf
Reviewer_uwUQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the inherent tension between exploration and accuracy in reinforcement learning with verifiable rewards. The authors point out that standard entropy regularization pushes the policy toward a globally uniform distribution, thereby potentially increasing the probability of incorrect responses. The paper introduces Correctness-Conditioned Entropy (CCE), which maximizes the entropy of the policy conditioned on correctness, and analyzes its optimality-preserving property and relationship to Chernoff divergence. The authors integrate CCE into DAPO and MaxRL and conduct experiments across three model series, mainly on mathematical reasoning tasks. The results show improvements in average accuracy and the diversity of correct responses.

Critical strengths

(1) The paper clearly explains why standard entropy regularization can conflict with reward maximization. Encouraging diversity only among correct answers is a reasonable exploration method because it does not directly increase the probability of incorrect answers.

(2) The experiments cover three model series, multiple reinforcement learning baselines, several mathematical benchmarks, and two policy optimization algorithms. In addition to accuracy, the paper reports diversity metrics and coefficient ablations, providing useful evidence beyond benchmark accuracy.

Critical weaknesses

Major concerns

(1) The theoretical analysis is derived in a distribution-level binary-reward setting, whereas practical training uses an offline Monte Carlo estimator. In Equation 16, the CCE term is active only when a rollout group contains at least two correct responses. The paper does not report the fraction of rollout groups that meet this condition during training or analyze how the estimation quality changes on difficult benchmarks.

(2) The coefficient ablation on Qwen2.5-7B indicates that the effect of CCE does not depend on a single coefficient. However, the main results table reports only the best coefficient for each regularization method, without clarifying whether an independent validation set was used for parameter selection. It also does not report variation across multiple training runs. The average improvement over DAPO is 1.3–2.0 percentage points, so the stability of the cross-model results remains unclear.

(3) The experimental results show increased diversity among correct answers, but they do not establish that CCE discovers new successful reasoning strategies. The gains may instead result from redistributing probability among correct trajectories already present in the rollout groups. Analyzing semantic strategy coverage would provide stronger support for the paper’s claims regarding exploration.

Minor concerns

(1) Table 5 reports the total training time but does not separately report the additional computational and memory costs of CCE relative to DAPO or MaxRL. Reporting the proportion of prompts receiving non-zero CCE updates would also help clarify its practical efficiency.

(2) Evidence beyond mathematical reasoning remains limited. The logical reasoning experiment uses only one model and shows only a small average improvement over the strongest baseline.

Questions for authors

(1) During training, what proportion of rollout groups contains at least two correct responses? How does CCE perform when this condition is rarely met?

(2) How were the regularization coefficients in the main results selected? Can the authors provide multi-seed results using an independent validation set?

(3) Can the authors further demonstrate that CCE discovers new correct reasoning patterns rather than merely reallocating probability among existing correct responses?

3 comments
AuthorsOpenReview ↗

Response to Reviewer uwUQ (1/3)

Thank you for your constructive comments! We have carefully responded to all the weaknesses and questions as follows:

  1. Major Weakness 1 & Question 1 (CCE with few correct responses): We have provided the CCE activation fractions below, which are relatively high. We have also conducted additional experiments on an extremely challenging dataset, where CCE continues to achieve the best performance. We also theoretically justify restricting CCE estimation to groups with K≥2K \geq 2. Specifically, when K=0K=0, the CCE gradient is not used in Eq.15, so estimating it has no effect. When K=1K=1, incorporating the CCE advantage is equivalent to scaling the policy gradient without changing its direction. Therefore, skipping CCE estimation for groups with K<2K<2 has little impact. This explains the estimation quality on challenging datasets.

  2. Major Weakness 2 & Question 2 (Multiple runs and independent validation set): We use an independent validation set to reselect β\beta and provide additional multi-run results.

  3. Major Weakness 3 & Question 3 (Exploration): We provide additional semantic experiments demonstrating that CCE discovers more new successful reasoning strategies. We also reformulate the CCE gradient (Eq. 14) at the token level (as shown below), demonstrating that CCE explores all prefixes for which the product of relative accuracy and rarity is sufficiently high. This explains how CCE discovers new reasoning patterns.

  4. Minor Weakness 1 (Computational costs): We report the additional computation time for CCE, total training GPU-hours, step time, and GPU memory usage, demonstrating that CCE introduces negligible computational overhead.

  5. Minor Weakness 2 (Logical reasoning experiments): We conduct additional logical reasoning experiments using Qwen2.5-14B-Instruct, where CCE continues to achieve the best performance.

We have also incorporated the corresponding additions into the latest version of the PDF:

  • Major Weakness 1 & Question 1: Brown (not Orange)
  • Major Weakness 2 & Question 2: Blue & Green
  • Major Weakness 3 & Question 3: Red (not Purple)
  • Minor Weakness 1: Brown (not Orange)
  • Minor Weakness 2: Pink

Major W1 & Q1: CCE with few correct responses

Thank you for highlighting the need to analyze the stability of Eq.16 when the number of correct responses KK is small. We have:

  • added the frequency of CCE activation in our experiments, which is typically above 50%, indicating that CCE is frequently active.
  • evaluated CCE on a more challenging dataset, where its early activation rate is only 2.73% but increases to 16.01% in later stages. CCE also achieves the best performance on this setting.
  • further theoretically analyzed the impact of K<2K<2.
    • When $ K=0$ , the CCE gradient is not used in Eq. 15, so there is no need to estimate.
    • When $ K=1$ , the CCE gradient in Eq. 15 is equivalent to scaling the gradient without CCE, without changing its direction.
    • Therefore, not estimating CCE has only a minor impact.
    • We believe this explains the estimation quality of CCE on challenging datasets.

1. CCE activation frequency

We reran the three main CCE experiments and recorded the CCE activation group-wise frequency:

ExperimentCCE active ratio at step 1CCE active ratio at the last step
Qwen7b50.40%77.32%
Qwen4b67.22%80.50%
Llama1b43.12%68.45%

where CCE remains highly active and plays an important role. Moreover, as training progresses and accuracy improves, the CCE activation frequency increases further.

2. CCE activation frequency and performance on challenging datasets

We also conducted experiments on a more challenging dataset, using the smallest model, Llama3.2-1B-Instruct, and the most difficult training dataset, dapo-17k. In the early stage of training, the accuracy was only 1.71%, and the CCE activation rate was 2.73%. By the late stage of training, these values increased to 8.30% and 16.01%, respectively. The training results are shown below:

MethodAIME24AIME25GSM8kMATH500AMC23OlympiadminervamathHMMTavg
DAPO1.870.0052.3127.6015.507.585.150.3913.80
Entropy2.000.2753.8326.4017.836.393.130.2213.76
DivPO1.470.0055.7230.4015.176.694.600.6114.33
CCE1.330.4057.6229.6019.337.583.681.2815.10

CCE regularization achieves the highest overall performance improvement. This demonstrates that CCE's effectiveness stems from precise exploration rather than indiscriminate exploration on a large scale.

3. Theoretical analysis of CCE estimation quality on challenging problems

Since the CCE gradient is not used in the policy gradient (Eq.15) when $ K=0 $ and does not change its direction when K=1K=1, restricting Eq.16 to estimating the policy gradient only when $ K \ge 2$ does not significantly affect the estimation quality.

AuthorsOpenReview ↗

Response to Reviewer uwUQ (2/3)

Case 1 ($ K=0$ ): Eq.15 does not use any positive samples, so the CCE gradient from positive samples is not used. Therefore, skipping CCE estimation has no impact. This does not introduce additional waste, since the advantages of GRPOs are also 0 in this case, meaning they do not use any of them.

Case 2 (K=1K=1): Eq.15 uses only one positive sample. In the RLVR setting, all negative samples within a group have the same reward. Therefore, without CCE, the advantages take only two possible values: A1A_1 for negative samples and A2A_2 for the positive sample.

Suppose an CCE estimator assigns a CCE advantage of vv to this positive sample. After applying CCE, the advantages remain limited to: A′=A1A'=A_1 for negative samples and A′=A2+vA'=A_2+v for the positive sample. Thus, there exist constants c=A2−A1+vA2−A1c=\frac{A_2-A_1+v}{A_2-A_1} and k=(1−c)A1k=(1-c)A_1 such that A′=cA+kA'=cA+k, i.e., CCE scales the original advantages by cc and adds a constant kk. According to REINFORCE, adding a constant to the advantages does not change the policy gradient (E[k⋅∇log⁡π]=0\mathbb{E}[k \cdot \nabla \log \pi]=0). Therefore, when $ K=1 $, CCE is equivalent to simply scaling the original gradient. Since scaling does not change the gradient direction, its impact is negligible in fully on-policy settings or when clipping is triggered infrequently (less than approximately 1% of the time in our experiments).

Major W2 & Q2: multi-seed results and independent validation set

Thank you for your concern regarding experimental stability. We have added results on independent validation sets and made every effort to include multiple runs.

1. Independent Validation Sets

We swapped the training datasets across 3 experiments (all are disjoint from the benchmarks) and removed similar samples. Next we sampled 1k examples from each to construct independent validation sets:

ExperimentTraining SetValidation Set
Qwen7BDeepMathCompetition-Math
Qwen4BDAPO-17KDeepMath
Llama1BCompetition-MathDAPO-17K

The resulting β\beta for Entropy, DivPO, and CCE remained identical (2 runs). This may because our evaluation covers 10 benchmarks, resulting in stable scores and consistent performance trends, while different values of β\beta can lead to relatively significant performance gaps.

2. Multiple-Run Results

We initially reported single-run results because our evaluation averages scores across 10 benchmarks (compared to approximately 3 in most prior works), substantially reducing variance. Indeed, we observed score differences of less than 0.4 points across random seeds, indicating that single-run results are already sufficiently stable and reliable.

Below, we report our existing and newly conducted multi-run experiments (mean ± std over 2–3 runs):

MethodQwen7BQwen4BLlama1B
DAPO38.91 ± 0.2460.74 ± 0.1414.77 ± 0.30
KL-Cov38.38 ± 0.2960.41 ± 0.2715.45 ± 0.36
Clip-Cov38.62 ± 0.1960.55 ± 0.1215.44 ± 0.23
Entropy37.66 ± 0.2660.18 ± 0.2114.97 ± 0.27
DivPO38.34 ± 0.2560.34 ± 0.1715.41 ± 0.35
CCE41.00 ± 0.2161.94 ± 0.2516.25 ± 0.19

These results show small variations across independent runs, supporting the stability of single-run evaluations. More importantly, CCE consistently maintains its advantage across runs, demonstrating robust improvements.

Given the demonstrated stability of our results and the common practice of single-run evaluations in large-scale LLM training due to computational costs, we will make every effort to include multiple runs for all settings in a future revision. We will explicitly identify the parts that cannot be completed in a future revision and include them in the Limitations section.

Major W3 & Q3: new correct reasoning patterns

Thank you for your thoughtful comments. However, the mechanism of CCE enables it to discover new reasoning strategies, and our experiments further demonstrate that CCE indeed discovers more correct reasoning paths. If the explanation below does not fully address your concerns, we recommend referring to our response to Weakness 1 raised by Reviewer uwUQ, which provides an alternative perspective on the same issue.

1. Where "new reasoning strategies" come from?

From "exploring currently unsuccessful trajectories", meaning exploring potentially correct prefixes within incorrect trajectories (a≤ta_{\le t} where a0:T∉As∗a_{0:T} \notin A^*_s) that may lead to new correct trajectories ($ a'_ {0:T} \in A^ *_ s $ where $ a'_ {\le t}=a_ {\le t}$ ). To generate trajectories that are both "new" and "correct", such prefixes must satisfy:

  1. "New": Rare (low log⁡π\log \pi).
  2. "correct": Be able to reach a correct answer ($ \pi(A_s^* | a_{\le t}, s)\ne 0 $). Otherwise, increasing their frequency is unlikely to produce correct trajectories immediately. Such prefixes can be encouraged later when their success probabilities improve.
AuthorsOpenReview ↗

Response to Reviewer uwUQ (3/3)

2. Switching to token-level expression, CCE does explore potentially correct reasoning prefixes within incorrect trajectories

We equivalently reformulate Eq.14 at the token level, revealing that:

$ \nabla \mathcal{H}_ {c} = -\sum_ t \mathbb{E}\Bigg[ \mathbf{ \frac{\pi(A^ *_ s\mid a_ {\le t},s)}{\pi(A^ *_ s\mid s)} \cdot \Big( \mathbb{E}[\log\pi(a_ {0:T}\mid s)\mid A^ *_ s,a_ {\le t},s] -\mathbb{E}[\log\pi(a_ {0:T}\mid s)\mid A^ *_ s,s] \Big)} \cdot \nabla\log\pi(a_ t\mid a_ {<t},s) \Bigg]. $

Here, CCE assigns an additional reward based on the product of (1) the prefix's success probability relative to the prior (π(As∗∣a≤t,s)π(As∣s)\frac{\pi(A_s^*| a_{\le t},s)}{\pi(A_s^\mid s)}) and (2) its associated rarity (E[log⁡π∣⋅]\mathbb{E}[\log\pi|\cdot]). Thus, CCE favors rare prefixes with meaningful success probabilities—precisely the "currently unsuccessful trajectories".

In contrast, conventional entropy regularization uses log⁡π\log\pi (Table 1) to encourage rare prefixes indiscriminately, including those unlikely to yield correct outcomes.
Consequently, CCE does not "abandon the exploration of currently unsuccessful trajectories." Instead, CCE selectively explores promising prefixes within incorrect trajectories.

3. CCE discovers more novel correct trajectories

We report the average number of new reasoning patterns / new correct reasoning patterns per rollout relative to the base model:

MethodAIMEHMMTMATH
DAPO0.5 / 0.10.5 / 0.040.3 / 0.2
+Entropy0.9 / 0.10.8 / 0.030.5 / 0.1
+CCE0.8 / 0.20.7 / 0.090.5 / 0.4

As shown, entropy regularization generally discovers more new reasoning patterns, but not more new correct ones.
In contrast, CCE discovers the most new successful reasoning paths, demonstrating more effective exploration.

We will include these details in a future revision of the paper.

Minear W1: computational overhead

Thank you for your concern regarding the computational overhead introduced by CCE. We have added the necessary overhead measurements.

1. Experimental Results: CCE Introduces Negligible Overhead

We measured the computational costs of the key methods across different experiments. For each cell, the five numbers represent, respectively, total training GPU-hours, throughput, time per training step, time spent computing the CCE term per step, and average GPU memory usage per GPU.

MethodQwen7BQwen4BLlama1B
DAPO137/1153/1351/0/70255/1071/2401/0/758/5021/306/0/70
MaxRL140/1156/1384/0/71243/1077/2379/0/748/5014/307/0/70
CCE140/1159/1370/1.4/70256/1069/2451/2.0/748/5073/308/0.29/70

As shown, CCE introduces negligible additional computational overhead. Computing the CCE term typically takes only 1–2 seconds, compared with approximately 1,000 seconds for a complete training step.

The slightly higher overall runtime of CCE mainly comes from its encouragement of diversity, which leads to longer reasoning sequences and consequently increases inference time. The overhead directly attributable to CCE computation itself is minimal. Nevertheless, since a high proportion of prompts receive non-zero CCE advantages (see our response to W1), the additional overhead remains negligible.

2.Why CCE Introduces Little Overhead

CCE only requires reusing existing outputs (log⁡π\log \pi) and computing the additional penalty reward induced by the CCE regularization term (Eq. 15). Thus, its computational overhead is limited to the additional advantage calculation, which is negligible compared with the computationally intensive stages of RLVR, e.g., rollout generation and parameter updates.

Moreover, GPU memory consumption during RLVR is primarily determined by LLM inference, particularly the KV cache reserved by inference frameworks such as vLLM. Therefore, the additional memory overhead of computing the CCE regularization term is also negligible.

Minear W2: logical experiment

Thank you for your interest in the additional experiments.

We added the Logic experiments using Qwen2.5-14B-Instruct. The results are as follows:

MethodProofwriterRobustLRRuletakertomi-nliavg
DAPO92.3071.2046.9012.2055.65
DivPO92.2069.8048.0013.5055.88
CCE94.2073.4047.9015.3057.70

where CCE still achieves the best performance.

Reviewer_RZN3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Correctness-Conditioned Entropy (CCE) regularization for reinforcement learning with verifiable rewards. Rather than maximizing entropy over all responses, CCE maximizes the entropy of the policy conditioned on correctness. The authors interpret this conditional policy as a projection onto the reward-optimal policy space and establish that the regularized objective preserves reward optimality while selecting the maximum-entropy optimal policy. Experiments report improvements in average accuracy across three models on mathematical reasoning benchmarks.

Critical strengths
  1. Simple regularizer with a clear optimality property. CCE provides an interpretable way to encourage diversity among correct responses, which is an interesting idea. The authors also provide a theoretical explanation on, for any β>0\beta > 0, how the optimal solution for the CCE objective preserves optimality with respect to correctness as well.

  2. Consistent empirical improvements. The experiments cover three models and multiple reasoning benchmarks. The reported improvements in average accuracy across the tested settings, together with the additional diversity measurements, suggest that the method could have practical value.

Critical weaknesses
  1. Oversimplified interpretation of exploration. The paper treats probability assigned to incorrect trajectories as a drawback of entropy regularization. However, exploring currently unsuccessful trajectories can be an intended part of learning. A difference between the reward-only and entropy-regularized optima does not establish that this exploration is harmful. CCE encourages diversity among correct responses, but its connection to discovering new successful reasoning paths remains unclear.

  2. Uncertainty across different runs. The main DAPO experiments report average improvements of approximately 1.3–2.0 percentage points, but variability across independent training seeds and uncertainty on benchmark scores are not clearly reported. Moreover, this result is obtained by tuning the β\beta hyperparameter. Without repeated training runs, it is difficult to assess the robustness of these improvements.

Questions for authors
  1. The optimality-preserving result holds for any β>0\beta>0. Is there theoretical or practical guidance for selecting β\beta under finite training? In particular, does the divergence interpretation in Section 5 suggest a useful selection principle or help explain the sensitivity observed in Table 4?

  2. The authors argue that CCE preserves optimality by analyzing the global optimal of the solution, and the Chernoff divergence argument relies on the same argument as well. However, that does not guarantee convergence, or intermediate updates getting closer to that global optimum, especially when considering policies parameterized by a neural network (or even a softmax policy). Could the authors comment on this?

4 comments
AuthorsOpenReview ↗

Response to Reviewer RZN3 (1/3)

Thank you for the detailed feedback! We have responded to all weaknesses and questions:

  • Weakness 1 (Exploration): In fact, CCE can explore correct reasoning processes within currently incorrect trajectories, as they may also appear in correct trajectories. By reformulating Eq.14 at the token level, we show that CCE encourages prefixes with a sufficiently high product of relative accuracy and rarity, promoting new correct trajectories. We further provide empirical evidence that CCE discovers more new correct trajectories.
  • Weakness 2 (Experiments): We added multi-run results to demonstrate the robustness of our findings.
  • Questions:
    • We recommend setting $ \beta \approx \frac{0.1}{L} $, where $ L $ is the model's average sentence log-probability. The sensitivity observed in Table 5 arises because increasing $ \beta $ promotes more exploration but also increases variance.
    • Regarding convergence, we acknowledge that strict monotonic improvement is not guaranteed. Nevertheless, our method guarantees the correct global optimum and, like widely-used variational inference, encourages monotonic improvement. This is precisely the underlying mechanism of our approach.

We have also incorporated the corresponding additions into the latest version of the PDF:

  • Weakness 1: Red (not Purple)
  • Weakness 2: Green
  • Questions: Orange (not Brown)

W1: Exploration of unsuccessful trajectories

Thank you for your thoughtful comments.
However, CCE does not abandon the exploration of unsuccessful trajectories. Instead, CCE explores potentially correct reasoning processes within incorrect trajectories if they can also occur in correct trajectories.

1. Why might CCE appear not to explore incorrect trajectories?

This concern may stem from the sequence-level expression in our paper (e.g., Eq. 14), where an action ($ a_{0:T} $) represents a complete trajectory with a final outcome. Since a completed incorrect trajectory cannot become correct (as they have finished), exploring such trajectories (not the reasoning prefixes within them) cannot lead to "new correct trajectories" and is therefore unnecessary. However, we believe the reviewer's notions of "currently unsuccessful trajectories" and "new successful reasoning paths" refer to a token-level perspective, where incomplete reasoning prefixes within completed unsuccessful trajectories ($ a_{0:t},t<T $) may eventually lead to new correct trajectories. Due to its full-sequence perspective, Eq. 14 alone does not reveal how CCE handles these prefixes.

2. What are the "currently unsuccessful trajectories"?

"Currently unsuccessful trajectories" are not completed unsuccessful trajectories, which already include unsuccessful outcomes and would therefore be considered "permanently unsuccessful trajectories". "Exploring currently unsuccessful trajectories" means exploring potentially correct prefixes within incorrect trajectories ($ a_{ \le t } $ where $ a_ {0:T} \notin { A^ *}_ s $ ) that may lead to new correct trajectories ($ {a'}_ {0:T} \in A^ *_ s $ where $ a' _ {\le t}=a_ {\le t} $). To generate trajectories that are both "new" and "correct", such prefixes must satisfy:

  1. "New": Rare (low $ \log \pi $).
  2. "correct": Be able to reach a correct answer ($ \pi(A_s^* \mid a_{\le t}, s)\ne 0 $). Otherwise, increasing their frequency is unlikely to produce correct trajectories immediately, as the current model cannot complete them correctly. Such prefixes can be encouraged later when their success probabilities improve.

3. Switching to token-level expression, CCE does explore potentially correct reasoning prefixes within incorrect trajectories

We equivalently reformulate Eq.14 at the token level, revealing that:

$ \nabla \mathcal{H}_ {c} = -\sum_ t \mathbb{E}\Bigg[ \mathbf{ \frac{\pi(A^ *_ s\mid a_ {\le t},s)}{\pi(A^ *_ s\mid s)} \cdot \Big( \mathbb{E}[\log\pi(a_ {0:T}\mid s)\mid A^ *_ s,a_ {\le t},s] -\mathbb{E}[\log\pi(a_ {0:T}\mid s)\mid A^ *_ s,s] \Big)} \cdot \nabla\log\pi(a_ t\mid a_ {<t},s) \Bigg]. $

Here, CCE assigns an additional reward based on the product of (1) the prefix's success probability relative to the prior (π(As∣a≤t,s)π(As∣s)\frac{\pi(A_s^\mid a_{\le t},s)}{\pi(A_s^\mid s)}) and (2) its associated rarity (E[log⁡π∣⋅]\mathbb{E}[\log\pi|\cdot]). Thus, CCE favors rare prefixes with meaningful success probabilities—precisely the reviewer's "currently unsuccessful trajectories".

In contrast, conventional entropy regularization uses log⁡π\log\pi (Table 1) to encourage rare prefixes indiscriminately, including those unlikely to yield correct outcomes.
Consequently, compared with entropy regularization, CCE does not "abandon the exploration of currently unsuccessful trajectories." Instead, CCE selectively explores promising prefixes within incorrect trajectories, rather than blindly encouraging all rare trajectories..

AuthorsOpenReview ↗

Response to Reviewer RZN3 (2/3)

4. The prevalence of potentially correct reasoning prefixes

We use DeepSeek-V4.1-Flash to classify reasoning patterns in rollouts from Qwen7B models trained with different algorithms on 3 benchmarks. We quantified the frequency of "prefixes within incorrect trajectories encouraged by CCE" by measuring the average number of reasoning patterns per rollout that could appear in both incorrect and correct trajectories:

AIMEHMMTMATH
0.360.170.91

This indicates that there are many reasoning patterns that can appear in both correct and incorrect trajectories. The prefixes of incorrect trajectories explored by CCE are ubiquitous.

5. CCE discovers more new correct trajectories

We report the average number of new reasoning patterns / new correct reasoning patterns per rollout relative to the base model:

MethodAIMEHMMTMATH
DAPO0.5 / 0.10.5 / 0.040.3 / 0.2
+Entropy0.9 / 0.10.8 / 0.030.5 / 0.1
+CCE0.8 / 0.20.7 / 0.090.5 / 0.4

As shown, entropy regularization generally discovers more new reasoning patterns, but not more new correct ones.
In contrast, CCE discovers the most new successful reasoning paths, demonstrating more effective exploration.

In summary

CCE avoids encouraging "completed incorrect trajectories", not "their potentially correct prefixes". Our derivation shows that CCE prefer rare prefixes with nonzero success probabilities, even within incorrect trajectories, which are precisely the "currently unsuccessful trajectories" In contrast, entropy regularization also encourages prefixes that may never lead to correct outcomes. Therefore, CCE explores currently unsuccessful trajectories more effectively than entropy regularization.

W2: Different Runs

Thank you for your comments on experimental stability. We have supplemented the key experiments with results from multiple runs (which show very low variance), and CCE remains the leading method.

We initially reported single-run results because our evaluation averages scores across 10 benchmarks (compared to ~3 in most prior works), substantially reducing variance. Indeed, we observed score differences of less than 0.4 points across random seeds, indicating that single-run results are already sufficiently stable and reliable.

Below, we report our existing and newly conducted multi-run experiments (mean ± std over 2–3 runs):

MethodQwen7BQwen4BLlama1B
DAPO38.91 ± 0.2460.74 ± 0.1414.77 ± 0.30
KL-Cov38.38 ± 0.2960.41 ± 0.2715.45 ± 0.36
Clip-Cov38.62 ± 0.1960.55 ± 0.1215.44 ± 0.23
Entropy37.66 ± 0.2660.18 ± 0.2114.97 ± 0.27
DivPO38.34 ± 0.2560.34 ± 0.1715.41 ± 0.35
CCE41.00 ± 0.2161.94 ± 0.2516.25 ± 0.19

show small variations across independent runs, supporting the stability of single-run evaluations. More importantly, CCE consistently maintains its advantage across runs. Moreover, CCE consistently outperforms DAPO across all tested β\beta values (Table 4) and remains effective under all tested MaxRL settings (Table 3), providing further evidence of its robustness across configurations.

Previous work [1] has shown that performance differences between RL algorithms are often small under large-scale RLVR training. In this context, CCE's improvements of over 1 point are substantial, especially since differences among other baselines are generally below 1 point.

Given the demonstrated stability of our results and the common practice of single-run evaluations in large-scale LLM training due to computational costs, we will make every effort to include multiple runs for all settings in a future revision. We will explicitly identify the parts that cannot be completed in a future revision and include them in the Limitations.

Q1: Selecting β\beta

Theoretical: Increasing β\beta prioritizes exploration, but also increases variance.

As shown in Section 4, CCE regularization is equivalent to jointly minimizing two distances: (1) the projection distance from the policy $ \pi$ to the optimal policy space $ \Pi^ * $ , and (2) the distance from this projection to the maximum-entropy optimal policy $ \pi ^ * $ . According to Theorem 1, the first depends solely on optimality, while the second depends solely on diversity.

However, increasing $ \beta $ places greater emphasis on diversity, potentially prioritizing exploration over accuracy improvement during early training. While greater exploration can improve performance, it also increases gradient variance (the policy involves greater uncertainty). Thus, an excessively large β\beta may destabilize training and degrade performance, explaining the results in Table 4. Therefore, a reasonable choice of β\beta should keep the CCE-induced and reward-induced gradients at comparable orders of magnitude, balancing diversity and accuracy.

AuthorsOpenReview ↗

Response to Reviewer RZN3 (3/3)

Empirically

we find that the CCE-induced policy gradient performs well when its magnitude is approximately 10% of the reward-induced policy gradient. For example:

  • Llama: Average sequence log-probability magnitude ≈ 100; $ \beta\approx10^{-3}$ .
  • Qwen7B: Average sequence log-probability magnitude ≈ 200; $ \beta\approx5\times10^{-4}$ .
  • Qwen4B: Average sequence log-probability magnitude ≈ 1000; $ \beta\approx10^{-4}$ .

Thus, with a limited experimental budget, a practical initial choice is β≈0.1/L\beta\approx0.1/L, where LL is the estimated average magnitude of the sequence log-probability under the base model.

Q2: Convergence

Thank you for raising this important question about monotonic convergence. CCE regularization does not strictly guarantee the monotonic convergence of the Chernoff divergence (we will acknowledge this limitation in a future revision). Instead, similar to widely used variational inference methods, CCE reduces an upper bound on the Chernoff divergence.

Our theoretical result establishes the global optimality of the distribution-level objective, rather than convergence of practical neural policy optimization. The Chernoff divergence interpretation provides an optimization perspective, not a convergence guarantee.

However, compared with variational inference, CCE regularization offers two additional advantages: (1) it guarantees the correct global optimum, and (2) it is more likely to promote monotonic improvement.

1. The Similarity Between CCE Regularization and Variational Inference

In Section 5, we prove that CCE regularization optimizes an upper bound $ \mathcal{M}_ \alpha $ of the Chernoff divergence $ C_ \alpha $, and that both share the same optimal solution. This is analogous to variational inference (used in VAEs and Diffusion models), which optimizes the evidence lower bound (ELBO) rather than the likelihood directly. Thus, CCE only encourage monotonic improvement of the true objective, while strict monotonicity cannot be guaranteed in general.

2. When Is Monotonic Improvement Happened?

First and foremost, the fundamental requirement is that the loss itself must decrease monotonically, that is: $ \mathcal{M}_ \alpha^{(t+1)} < \mathcal{M}_ \alpha^{(t)}. $

Recall that

$ \mathcal{M} _ \alpha ^ { (t+1)} - \mathcal{M}_ \alpha ^ {(t)} = C _ \alpha ^ {(t+1)}-C_ \alpha ^ {(t)} + \mathbb{D} _ {KL}( \pi' ^ {(t+1)}| q _ \alpha ^ { (t+1) }) - \mathbb{D} _ {KL}( \pi' ^ {(t)}|q _ \alpha ^ {(t) }) , $

monotonic improvement of $ C_\alpha $ is guaranteed whenever

$ \mathcal{M}_ \alpha^ {(t)}-\mathcal{M}_ \alpha^ {(t+1)} \ge \mathbb{D}_ {KL}(\pi'^ {(t)}|q_ \alpha^ {(t)}) -\mathbb{D}_ {KL}(\pi'^ {(t+1)}|q_ \alpha^ {(t+1)}). $

Therefore, the following cases may lead to monotonic improvement:

Case 1: when sufficiently close to the optimum.

Since the gap vanishes at the global optimum ($ \mathbb{D}_ {KL}(\pi'|q_ \alpha)=0 $ when $ \pi=\pi^ * $ ), and assuming smooth convergence, we have $ \lim_ {t \rightarrow + \infty}\mathbb{D}_ {KL}(\pi'^ {(t)}|q_ \alpha^ {(t)}) -\mathbb{D}_ {KL}(\pi'^ {(t+1)}|q_ \alpha^ {(t+1)})=0.$

Therefore, any sufficiently substantial decrease in Mα\mathcal{M}_\alpha guarantees monotonic convergence.

Case 2: When Mα\mathcal{M}_\alpha decreases sufficiently substantially Even in the worst case, the gap variation is bounded by

$ \mathbb{D}_ {KL}(\pi'^{(t+1)}|q_ \alpha^ {(t+1)}) + \mathbb{D}_ {KL}(\pi'^{(t)}|q_ \alpha^ {(t)}) \le \alpha\log|\mathcal{A}^ *|. $

Thus, a sufficiently large decrease in the surrogate loss can still guarantee monotonic improvement of the true objective, driving π\pi toward π∗\pi^*.

Although the above analysis identifies scenarios in which monotonic improvement is possible, we cannot guarantee that these conditions will necessarily hold in practice. Therefore, we acknowledge that monotonic improvement cannot be strictly guaranteed. However, the goal of Section 5 is to provide an understanding of CCE regularization from the perspective of divergence, rather than to establish its convergence advantages.

Refs

[1] arXiv:2603.19335

Thank you for your careful rebuttal. I will retain my original positive score.

Comments on the paper

2 comments

Dear Reviewers,

Thank you for your time and thoughtful reviews. Please read the authors’ response and share your follow-up comments promptly, indicating whether your concerns have been addressed. Please raise any additional questions early enough to give the authors sufficient time to respond.

Thank you.

Best,

AC

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 additional theoretical analyses and experiments primarily address three aspects that were repeatedly raised across the reviews: (1) RZN3-W1 and uwUQ-W3: We reformulated our method's original formulation for complete sequences (Eq. 14) into an equivalent formulation based on incomplete reasoning prefixes. This reformulation clarifies how CCE explores potentially correct prefixes within currently incorrect trajectories. We also provided statistics on the new reasoning patterns acquired by models trained with different algorithms. These results demonstrate that our method enables RLVR to discover and learn a broader range of previously unseen correct reasoning trajectories. (2) RZN3-W2 and uwUQ-W2: We added experiments on independent validation sets and conducted multiple runs with different random seeds to further establish the reliability and reproducibility of our results. (3) uwUQ-W1: We theoretically demonstrated that, when K≤2K \leq 2, not computing CCE in the gradient estimator defined by Eqs. 15–16 has a negligible impact on estimation quality. We also reported the activation rate of CCE and compared it with other methods on particularly challenging datasets. These results demonstrate that CCE retains substantial advantages even under extremely challenging conditions.

Beyond these core concerns, we addressed several additional questions. Specifically, we provided a selection principle for β\beta (RZN3), explained the fluctuations in Table 5 through the variance-exploration trade-off (RZN3), and clarified that the variational reformulation of Chernoff divergence serves as a geometric interpretation rather than a strict guarantee of monotonic improvement (RZN3). We also provided a detailed computational overhead comparison, demonstrating that the additional cost introduced by CCE is negligible (uwUQ). Furthermore, we added experiments with a 14B model on the Logic task, providing stronger evidence for the effectiveness of our method on larger models and Logic tasks (uwUQ, SBuj).

Finally, we directly responded reviewer SBuj's concerns regarding the novelty of our method and its distinction from DivPO. We highlighted the fundamental differences between CCE, conventional entropy regularization, and DivPO in four aspects:

  • Objective formulation (Section 4): Whether diversity is encouraged while maintaining optimality as a prerequisite. Theoretical framework (Sections 4–5): Our novel geometric analysis and the interpretability provided by the variational reformulation.
  • Derivation (Eq. 14): The use of conditional sampling to avoid the intractable conditional distribution.
  • Implementation (Eqs. 15–16): Offline gradient estimation and the handling of PPO clipping. Moreover, our experimental results show that CCE consistently outperforms these conventional approaches.

Overall, we believe that the additional theoretical analyses and experiments address the major concerns regarding the effectiveness of our method. Both our theoretical framework and empirical results further establish the novelty and advantages of our approach over existing methods. 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.