acceptodds
Withdrawn from ICLR 2027

GASPOL : Dual Gating Sub-Networks for Continual Multi-Agent Reinforcement Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Continuous coordination under dynamic environments remains a challenge for a cooperative multi-agent system. The existing study on Continual Multi-agent Reinforcement Learning (MARL) focuses on task discrimination that requires trajectory buffers from previous tasks, which could be inapplicable in real-world applications. In addition, the existing study requires feedback or numerous trajectory data to select the best available action layer during the inference phase. Addressing the mentioned issues, this paper proposes a study on rehearsal-free continual multi-agent reinforcement learning and proposes a new method named Dual Gating Sub-networks Policy (GASPOL). GASPOL introduces a dual expert system, i.e., global experts and local experts, as a part of the trained policy network. Global experts preserve general knowledge (stability) of all learned tasks, while local experts obtain task-specific knowledge (plasticity). GASPOL leverages a new actor advantage-guided regularization mechanism for more flexible regularization at each training iteration. Our evaluation on MARL environments, i.e., SMAC, Pettingzoo MPE2, and MAMuJoCo shows that the proposed method significantly achieves better stability-plasticity than the existing state-of-the-art methods. Our extended analysis shows that the proposed method can maintain its performance under various numbers of experts and neurons.

Primary area:
reinforcement learning
Keywords:
Dual gating, Sub-networks, continual, multi-agent, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KU2gOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies rehearsal-free continual multi-agent reinforcement learning, focusing on catastrophic forgetting and negative policy transfer across sequential tasks. It proposes GASPOL, a dual-gating architecture that combines global experts for preserving task-shared knowledge with local experts for learning task-specific policies. The method further introduces learned task representations, a soft-routing mechanism for expert selection, and advantage-guided regularization to balance stability and plasticity.

Critical strengths
  1. The paper addresses catastrophic forgetting and negative transfer in continual multi-agent reinforcement learning under a rehearsal-free setting. This is practically relevant, especially when storage is limited or past trajectories cannot be retained.
  2. GASPOL uses global experts to preserve shared knowledge across tasks and local experts to learn task-specific policies, providing an intuitive approach to balancing stability and plasticity.
  3. The gating mechanism directly selects local experts at inference time without requiring environmental rewards, task IDs, or additional probing trajectories, making the method more practical than existing approaches such as OWL and MACPRO.
Critical weaknesses
  1. Figure 1(b) appears to contain a factual inconsistency: the randomly initialized policy reportedly starts with a success rate above 90%, while the transferred policy starts near zero. This contradicts the discussion of negative transfer in the Introduction and should be clarified.
  2. The claim of “highlighting catastrophic forgetting and negative transfer” is overstated, as both are well-known problems in continual learning rather than novel findings of this paper.
  3. The methodological novelty is limited. GASPOL appears largely to apply a standard MoE-style structure to the policy network and label its components as global and local experts. The paper does not clearly distinguish the proposed framework from existing MoE-based or expandable-network approaches.
  4. Figures 4 and 5 do not show consistent improvements over the baselines; GASPOL performs similarly or even worse in some stages. Therefore, the claimed advantages in stability and plasticity are not fully supported by the learning curves.
Questions for authors

Please refer to weakness.

4 comments
AuthorsOpenReview ↗

Authors' Response to Reviewer KU2g

W1. Figure 1(b) appears to contain a factual inconsistency: the randomly initialized policy reportedly starts with a success rate above 90%, while the transferred policy starts near zero. This contradicts the discussion of negative transfer in the Introduction and should be clarified.

Our response: Figure 1(b) shows the comparison of the transferred policy versus the trained-from-scratch policy for task-2 in MAMuJoCo games. The policy negative transfer from task-1 to task-2 causes the task-2 performance to start from a lower magnitude than the performance of randomly initialized (trained from scratch) policy. Rather than boosting the policy learning on task-2, the transferred model resists achieving plasticity. This is aligned with our narration in the paragraph.

W2. The claim of “highlighting catastrophic forgetting and negative transfer” is overstated, as both are well-known problems in continual learning rather than novel findings of this paper.

Our response: We believe that showing catastrophic forgetting and negative transfer for continual multi-agent reinforcement learning (MARL), with empirical data and visualization is something new to existing studies. To our knowledge, we have not found the detailed analysis on negative transfer impact to currently learned policy, in the previous studies.

W3. The methodological novelty is limited. GASPOL appears largely to apply a standard MoE-style structure to the policy network and label its components as global and local experts. The paper does not clearly distinguish the proposed framework from existing MoE-based or expandable-network approaches.

Our response: As we mentioned in introduction section, there are at least 2 components that distinguish the proposed framework from existing MoE-based or expandable-network approaches. First, we build a customized dual-expert structure tailoring MoE for global experts, a switch mechanism for local experts, which is connected with customized task representation design (K^o, K^f, K^p). Second, our proposed architecture is designed to learn a sequence of tasks in a rehearsal-free manner. While all global experts are trained on all tasks, a local expert is trained on the current task and has a low probability of being trained on upcoming tasks. To our knowledge, these 2 aspects are distinguished from the existing MoE/evolving methods, including the existing continual MARL methods, i.e., OWL and MACPRO.

W4. Figures 4 and 5 do not show consistent improvements over the baselines; GASPOL performs similarly or even worse in some stages. Therefore, the claimed advantages in stability and plasticity are not fully supported by the learning curves.

Our response: As we mentioned in our narration, Figures 4 and 5 show the adaptability to the current task; thus, the Figures show the performance for the currently learned task only, not the average of all learned tasks. The performance of all learned tasks is shown in Table 2. Indeed, in some stages, our proposed method achieves slightly lower performance than some methods for the currently learned task. However, looking at the performance of all learned tasks (on average), our methods achieve the highest performance among the consolidated methods. This means that our proposed method achieves the highest stability and plasticity.

I thank the authors for their response and revisions. However, the rebuttal has not changed my core assessment. First, Figure 1(b) appears to have been redrawn without explaining the previously reported results or the reason for the change. Second, I agree that a transferred policy can initially perform worse than a randomly initialized policy on a new task. However, related observations have been discussed in Efficient Online Reinforcement Learning Fine-Tuning Need Not Retain Offline Data and work on neural-network plasticity, offline-to-online RL, and continual learning. Demonstrating catastrophic forgetting and negative transfer in continual MARL therefore appears to extend established phenomena rather than introduce a fundamentally new problem setting. Finally, while the architectural clarification is helpful, the global/local expert design still appears to be an incremental adaptation of existing MoE and expandable-network approaches.

AuthorsOpenReview ↗

Authors’ 2nd Response to Reviewer KU2g

Dear Reviewer KU2g, Thank you for your timely response and feedback.

Figure 1(b): We revised Figures 1(a) and 1(b) because an error in the normalization values in the previous plot caused the normalized success rate to exceed 1.0. We have corrected the figures accordingly. The raw numerical results and normalization values are presented in Appendix E. We added the reasons for the changes to our summary of changes.

Transferred vs. randomly initialized policy, and related to Efficient Online Reinforcement Learning Fine-Tuning Need Not Retain Offline Data (WSRL): Our paper differs from the referenced work in several fundamental ways.

1. Setting: While WSRL focuses on a single online reinforcement learning task, our study focuses on a sequence of multi-agent reinforcement learning tasks.

2. Highlight: While WSRL highlights the impact of offline data unavailability (no-retention) during online fine-tuning, our study highlights the impact of policy negative transfer from the previous task to the currently learned task. While WSRL highlights distribution shift between offline and online data, our work highlights policy misalignments between two similar tasks (games).

3. Objective and Evaluation: While WSRL attempts to achieve effective fine-tuning on a single task without offline data, our work focuses on handling catastrophic forgetting and achieving stability-plasticity in a sequence of tasks. As an implication, rather than being evaluated in a single task as in WSRL, our proposed method is evaluated on all learned tasks, from the first task to the current task. In addition, we explicitly quantify forgetting and measure it, while WSRL does not.

4. Neural Plasticity and Continual Learning: While WSRL focuses on plasticity, our work focuses on both stability and plasticity. To our understanding, WSRL discusses “continually training on offline data during fine-tuning”, not “continual reinforcement learning”, let alone “continual multi-agent reinforcement learning”. Please kindly refer to the previous 3 points.

Demonstrating catastrophic forgetting and negative transfer: Indeed, both terms have been mentioned in numerous continual learning studies. However, to our knowledge, in the area of “continual multi-agent reinforcement learning”, our work is the first study to demonstrate it with empirical data and visualization (plot), highlighting how a policy model from a previous task holds back its learning process on the current task. Particularly for continual MARL with continuous actions, our work is the first study.

Architecture: We respect your view on our architectural design. We believe that our bi-level (dual) gating system, which tailors the mixture and switch of experts, customized task representation, our original routing mechanism, and learning objectives, is distinguished from the existing MoE and expandable-network approaches. We believe that how the architecture works and how it is trained are counted as important elements. For instance, MoE and sparse MoE have the same architecture, but each has unique mechanisms, which distinguish one from the other.

Finally: We kindly request that you take the update and our responses into account and reconsider your score.

Kind Regards,

Authors of Submission 42421

AuthorsOpenReview ↗

Follow up on Reviewer KU2g Feedback

Dear Reviewer KU2g,

We would like to kindly follow up on our latest responses to your comments. We would appreciate it if you could let us know whether you have any further questions, concerns, or clarifications regarding our responses and the revisions made to the manuscript.

We would be happy to address any remaining concerns. Thank you very much for your time and consideration.

Kind regards,

Authors of Submission 42421

Reviewer_WSimOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GASPOL addresses rehearsal-free continual multi-agent reinforcement learning using shared global experts and growing task-specific local heads. A similarity-based router selects a head from each agent’s local observation and features, without test-time rewards or probing episodes. The method also introduces an adaptive regularizer. Experiments cover four task sequences across SMAC, MPE2, and MAMuJoCo, with component ablations and architecture-sensitivity analyses.

Critical strengths
  1. The architecture addresses a concrete deployment constraint: selecting task-specific heads from local observations without reward feedback or probing at test time. Combining shared global experts with local heads offers a useful approach to balancing shared knowledge and task specialization.
  2. The empirical study spans discrete and continuous action spaces and includes component ablations, expert-count and hidden-width sensitivity, and raw taskwise results. This provides useful evidence about the proposed components beyond an aggregate final score.
Critical weaknesses
  1. The regularizer in Eq. (7) is unclear: parameter importance is not clearly distinguished from stored parameter values, and the scaling factor can be negative. Please provide the full implemented formula and explain its connection to advantage-guided regularization and forgetting prevention. The proofs of Theorems 1 and 2 do not fully correspond to the implemented algorithm: the former uses an optimization objective different from Eq. (8), while the latter does not cover the full routing score in Eq. (2). Please provide the missing justification or restrict the scope of the guarantees accordingly.
  2. The experimental results require checking. Table 2 and Eq. (34) imply an OWL CSR of approximately 80.76 on SMAC SZ, rather than 49.17 in Table 1. Tables 1 and 2 repeat EWC and MAS results on MPE2, while Table 6 reports identical mean/standard-deviation pairs for EWC and OWL across all ten entries. Please verify these entries against the source runs, reconcile the main and ablation results, and specify the reward-normalization rule.
Questions for authors
  1. What is the exact implemented regularizer, and how is its potentially negative scaling factor handled?
  2. Under what conditions do the guarantees cover the full training objective and routing rule?
  3. Can you explain the duplicated entries and CSR discrepancy, reconcile the main and ablation tables, and identify the source runs and normalization rule?
3 comments
AuthorsOpenReview ↗

Authors' Response to Reviewer WSim

W1. The regularizer in Eq. (7) is unclear: parameter importance is not clearly distinguished from stored parameter values, and the scaling factor can be negative. Please provide the full implemented formula and explain its connection to advantage-guided regularization and forgetting prevention. The proofs of Theorems 1 and 2 do not fully correspond to the implemented algorithm: the former uses an optimization objective different from Eq. (8), while the latter does not cover the full routing score in Eq. (2). Please provide the missing justification or restrict the scope of the guarantees accordingly.

Our response:

Regularizer: Thank you for pointing this out. We have improved the explanation of our proposed regularization, including parameter importance and how the advantage function participates. We calculated the norm of current-old logits’ difference, ensuring the non-negativity of its magnitude. Please see our revised manuscript.

Theorem 1: Theorem 1 covers policy improvement in a model-free perspective, which is applicable to any deep neural network architecture, including our gated sub-network structure. As shown in the latter part of Theorem 1, we have analyzed how policy improvement works for commonly used architectures such as MLP and our gated sub-network policy.

Theorem 2: Theorem 2 shows the proven leverage of using dual key references (K^o and K^f) instead of using only 1 key (K^f), assuming both keys have been trained well enough with comparable accuracy to the task representation. Optimization objective different from Eq. (8): Thank you for correcting our writing error. Eq. 8 tunes K^p as well as the other mentioned parameters, aligning with routing in Eq. 2. We have revised the learning-objective formula in our revised manuscript.

W2. The experimental results require checking. Table 2 and Eq. (34) imply an OWL CSR of approximately 80.76 on SMAC SZ, rather than 49.17 in Table 1. Tables 1 and 2 repeat EWC and MAS results on MPE2, while Table 6 reports identical mean/standard-deviation pairs for EWC and OWL across all ten entries. Please verify these entries against the source runs, reconcile the main and ablation results, and specify the reward-normalization rule.

Our Response:

Tabulation errors: Thank you for the correction. We have fixed the tabulation errors, including OWL CSR in Table 2, duplicate entries (MAS and EWC for MPPE) in Tables 1 and 2, and OWL-EWC duplication in Table 6. Please kindly check our revised manuscript.

Q1. What is the exact implemented regularizer, and how is its potentially negative scaling factor handled?

Our response: As in W1, we have detailed the implementation of the proposed regularization. We calculated the norm of current-old logits’ difference, ensuring the non-negativity of its magnitude. Please see our revised manuscript.

Q2. Under what conditions do the guarantees cover the full training objective and routing rule?

Our response: As our response to W1, we have fixed our writing errors; thus, the learning objective in Eq. 8 covers full routing that is presented in Eq. 2.

Q3. Can you explain the duplicated entries and CSR discrepancy, reconcile the main and ablation tables, and identify the source runs and normalization rule?

Our response: The duplication and CSR discrepancy are tabulation errors that we have fixed in our revised manuscript.

AuthorsOpenReview ↗

Gentle Follow Up on Reviewer WSim Feedback

Dear Reviewer WSim,

We would like to kindly follow up on our previous response to your comments and questions. We would appreciate it if you could let us know whether you have any further questions, concerns, or clarifications regarding our responses and the revisions made to the manuscript.

We would be happy to address any remaining concerns. Thank you very much for your time and consideration.

Kind regards,

Authors

Thank you for your response and revisions. I appreciate your efforts to address my concerns. However, several issues remain insufficiently clarified. 1.Regularization (Eq. 7): While the norm of the logits difference is non-negative, this does not necessarily resolve the issue of a potentially negative advantage-dependent scaling coefficient. Please clarify the exact implemented regularizer and how this issue is handled. 2.Theoretical guarantees: Correcting Eq. (8) improves consistency with the routing mechanism, but does not establish that the theoretical guarantees apply to the complete training objective. The assumptions and applicability of Theorems 1 and 2, particularly regarding the full routing mechanism, should be clarified. 3.Experimental results: I appreciate the corrections to the reported tables. However, given the multiple inconsistencies, the corrected values, reward-normalization procedure, and whether these changes affect the main conclusions should be explicitly documented. Overall, while the revisions address some concerns, the theoretical justification and reliability of the experimental results still require further clarification. I therefore maintain my original rating.

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

The paper proposes GASPOL for replay-free continual multi-agent reinforcement learning. The method builds on HAPPO and combines a shared global expert, task-specific action heads, task routing based on observation and feature similarity, and a policy-change-guided regularization mechanism. The paper validates the method on four task sequences: SMAC Marine, SMAC SZ, MPE2, and MAMuJoCo. It reports better final average performance than most compared methods and analyzes the role of each component through ablation experiments.

Critical strengths
  1. The problem is practically meaningful. The method aims to simultaneously avoid replaying historical task data, reward feedback at inference time, and additional task probing, selecting action heads solely based on local observations. These constraints bring the research closer to application scenarios with limited storage and decentralized execution.
  2. The reported performance gains are attractive. The experiments cover both discrete and continuous action spaces, and on SMAC Marine the method achieves a final average win rate of 95.18%, substantially higher than most compared methods. The ablation results also provide preliminary support for the expert structure and regularization design.
Critical weaknesses
  1. The core regularization term and its theoretical basis remain insufficient. Equation (7) defines θF\theta^F as Fisher parameter importance but uses it as a reference for parameter distance; the scaling factor log⁡π−log⁡πold\log\pi-\log\pi^{old} is not guaranteed to be non-negative, and it is not clearly specified how the advantage function participates. In addition, Theorem 1 follows the proof for idealized updates and does not prove that training still satisfies the relevant conditions after incorporating routing and auxiliary losses.
  2. The experimental reporting contains internal inconsistencies that affect credibility. In Table 6, the per-task raw means and standard deviations of EWC and OWL are exactly identical, yet their aggregate metrics differ in Table 1. The FT and full GASPOL results in Table 3 are also inconsistent with Table 1, and the difference in settings is not explained. The number of random seeds and the normalization formula for continuous-task rewards are likewise not clearly specified, making the performance gains and ablation conclusions difficult to fully verify.
Questions for authors

The following are clarification requests that could change the evaluation:

  1. Please provide the exact implementation of Equation (7): how do Fisher importance, old parameters, and the advantage function each participate? How is the scaling coefficient guaranteed to keep the regularization acting as a constraint? Why does Equation (8) not include the projection key loss LKp\mathcal L^{K^p}?
  2. How does Theorem 1 cover the auxiliary losses and routing updates actually used? Can Theorem 2 hold for the complete routing that includes the Softmax and projection terms?
  3. Please explain the duplication of raw results for EWC and OWL in Table 6, and the reason for the different results between Table 1 and Table 3, and provide the normalization formula, the number of independent random seeds, and the corresponding statistical method. After correction, do the main comparisons and ablation conclusions still hold?
3 comments
AuthorsOpenReview ↗

Authors' Response to Reviewer yBNY

W1. The core regularization term and its theoretical basis remain insufficient. Equation (7) defines (\theta^F) as Fisher parameter importance but uses it as a reference for parameter distance; the scaling factor (\log\pi-\log\pi^{old}) is not guaranteed to be non-negative, and it is not clearly specified how the advantage function participates. In addition, Theorem 1 follows the proof for idealized updates and does not prove that training still satisfies the relevant conditions after incorporating routing and auxiliary losses.

Our Response:

Fisher parameter importance: Thank you for pointing out our writing error. We have fixed our writing error regarding the Fisher-based regularization loss. Please kindly check it in our revised manuscript. Scaling factor: We have improved the explanation of our proposed regularization, including how the advantage function participates. We calculated the norm of current-old logits’ difference, ensuring the non-negativity of its magnitude. Please see our revised manuscript.

Theorem 1: In the latest part of Theorem1, we have theoretically analyzed that Theorem1 is applicable to our proposed gated network architecture. In addition, Figures 4 and 5 show the policy improvement and convergence with our proposed gating network.

W2.The experimental reporting contains internal inconsistencies that affect credibility. In Table 6, the per-task raw means and standard deviations of EWC and OWL are exactly identical, yet their aggregate metrics differ in Table 1. The FT and full GASPOL results in Table 3 are also inconsistent with Table 1, and the difference in settings is not explained. The number of random seeds and the normalization formula for continuous-task rewards are likewise not clearly specified, making the performance gains and ablation conclusions difficult to fully verify.

Our Response:

Tabulation: Thank you for your correction. Indeed, we have tabulation errors in our manuscript. The tabulation errors have been fixed in our revised manuscript.

Different settings: Tables 1 and 3 have different settings and configurations. Table 1 shows results from 3 different random seeds, while Table 3 shows results from 1 seeded run. FT in Table 1 utilizes an MLP with 160-sized hidden layers, while FT in Table 3 utilizes an MLP with 128-sized hidden layers. In Tables 1 and 2, all the consolidated methods are run with a comparable backbone size. Since our proposed method grows its sub-networks, we intentionally set a smaller size for its hidden layers. Thus, they show different results for both FT and GASPOL. We have elaborated the captions in both tables to improve clarity. We also elaborated the detailed experimental settings in Section 6.1.c.

Q1.Please provide the exact implementation of Equation (7): how do Fisher importance, old parameters, and the advantage function each participate? How is the scaling coefficient guaranteed to keep the regularization acting as a constraint? Why does Equation (8) not include the projection key loss (\mathcal L^{K^p})?

Our response:

Equation 7: Similar to W1, we have improved our writing on the proposed regularization mechanism, including how Fisher importance, old parameters, and the advantage function participate. Projection key loss: Thank you for pointing out our writing error; the total loss includes the projection key loss, as shown in our ablation study. We have fixed the writing of Equation 8 and Section 4.3.e.

Q2. How does Theorem 1 cover the auxiliary losses and routing updates actually used? Can Theorem 2 hold for the complete routing that includes the Softmax and projection terms?

Our response:

Theorem 1: Similar to our response on W1, Theorem 1 covers policy improvement from a model-free perspective, which is applicable to any deep neural network architecture, including our gated sub-network structure. As shown in the latter part of Theorem 1, we have analyzed how policy improvement works for commonly used architectures such as MLP and our gated sub-network policy. In addition, Figures 4 and 5 show the policy improvement and convergence with our proposed gating network.

Theorem 2: Theorem 2 shows the proven leverage of using dual key references (K^o and K^f) instead of using only 1 key (K^f), assuming both keys have been trained well enough with comparable accuracy for selecting the task representation. Bringing Theorem 2 from cosine value ranges [-1, 1] to Softmax ranges [0,1] is directly applicable. In addition, Theorem 2 assumes a positive cosine function for both keys.
We include K^p as an auxiliary reference in case K^o and K^f are not discriminative enough in the same cases. Our ablation study shows that K^p contributes to improving its routing that leads to performance increasing.

AuthorsOpenReview ↗

Authors' Response to Reviewer yBNY COnt

Q3. Please explain the duplication of raw results for EWC and OWL in Table 6, the reason for the different results between Table 1 and Table 3, and provide the normalization formula, the number of independent random seeds, and the corresponding statistical method. After correction, do the main comparisons and ablation conclusions still hold?

Our Response:

Table 6: Thank you for the correction. We have fixed the tabulation errors in Table 6. We have provided the normalization formula for the raw numerical results in Tables 6 and 7. Please see Appendix E.

Tables 1 and 3: As mentioned in W2, Tables 1 and 3 have different settings and configurations. Table 1 shows results from 3 different random seeds, while Table 3 shows results from 1 seeded experiment. FT in Table 1 has 3 MLP layers, while FT in Table 3 has 2 MLP layers, as its third layer is structured as MoE (global experts). Thus, they show different results for both FT and GASPOL. We have elaborated the captions on both tables to improve clarity. We also elaborated the detailed experimental setup. After the correction, the main conclusion still holds.

AuthorsOpenReview ↗

Gentle Follow Up on Reviewer yBNY Feedback

Dear Reviewer yBNY,

We would like to kindly follow up on our previous response to your comments and questions. We would appreciate it if you could let us know whether you have any further questions, concerns, or clarifications regarding our responses and the revisions made to the manuscript.

We would be happy to address any remaining concerns. Thank you very much for your time and consideration.

Kind regards,

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers and Summary of Updates

Dear Program Chairs (PC), Senior Area Chairs (SAC), Area Chairs (AC), and Reviewers of ICLR Submission 42421,

We would like to express our gratitude to all the reviewers for their constructive and insightful comments on our paper. We are honored that all the reviewers highlighted the strengths of our paper; i.e., reviewer yBNY pointed out the practical, meaningful, and attractive performance gain of our work, reviewer WSim pointed out the usefulness of our work from the approach and empirical study points of view, and reviewer KU2g pointed out that our work is relevant, intuitive, and practical. The reviewers also acknowledged the presentation, informativeness, and soundness of our paper.

Second, we have revised our paper, following the reviewers' feedback. We have improved the quality and presentation of our paper with the summary of changes as follows:

  1. Revising the formulation and narration of Advantage-Guided Regularization

  2. Fixing the writing error in the Final Objective and Loss function

  3. Fixing tabulation errors and inconsistencies for Tables 1, 2, 3, and 6.

  4. Fixing Figures 1, 4, and 5 based on the latest results/updates or error findings. Figures 1(a) and 1(b) are updated due to an error in the normalization values in the previous plot caused the normalized success rate to exceed 1.0. We have corrected the figures accordingly. Figures 4 and 5 are updated due to tabulation errors for consolidated methods.

  5. Adding normalization values for Tables 6 and 7.

The changes w.r.t. feedback from reviewers yBNY and WSim are colored in red, while the changes w.r.t. feedback from reviewer KU2g are colored in blue.

We have addressed the reviewers' concerns and questions by replying to their respective comments.

Third, we look forward to receiving the reviewers' feedback and discussion on our responses.

Kind Regards,

Submission 42421

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

Related papers

Discussion (0)

Sign in to comment.