acceptodds
Under review as a conference paper at ICLR 2027

Fine-Tuning on Self-Generated and Reward-Weighted Data: Learning Dynamcis, Convergence Rates, and Benefits of Off-Policyness

Abstract

We study the learning dynamics of fine-tuning a policy model on self-generated and reward-weighted rollout data, with particular focus on a generalized version of classic REINFORCE — referred to as RE(S) — that updates the rollout distribution once every gradient steps. Prior work in bandits and reinforcement learning has developed rich theory for REINFORCE and policy gradient methods, and on-policy sampling (i.e., a small value of , ideally ) is often viewed as crucial to their success; yet in prominent application like post-training large language models, reward-guided self-training has proved to be effective even when the rollout distribution is updated infrequently, but theoretical understanding remains limited for the learning dynamics and convergence properties of these off-policy methods. To bridge these gaps, we develop a unified theory for RE(S) that covers the full spectrum of : it can be interpreted as a stage-wise iterative optimization process, where each stage takes gradient steps for minimizing the Kullback–Leibler distance to a fixed reward-weighted rollout distribution. For multi-arm bandits with softmax policies, our in-depth analysis and numerical experiments reveal three key findings: (1) for any fixed , RE(S) enjoys global convergence to the optimal policy as the number of rollout distribution updates , where denotes the total number of gradient steps; (2) we prove tight two-sided bounds for the convergence rate of RE(S), and show that its suboptimality gap achieves an asymptotic convergence rate, while only affects the length of a burn-in phase; (3) when initialized at a weak policy with a small optimal-action probability, RE(1) gets trapped around suboptimal policies for a long period, whereas RE(S) with a suitable avoids the detour and achieves significantly faster convergence to the global optimum, highlighting the benefits of off-policyness in this case.

Primary area:
learning theory
Keywords:
multi-arm bandit, reinforcement learning, off-policy RL, policy gradient, learning dynamics, convergence analysis

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_q3YqOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

Overall, I find the paper very didactic, with a theoretical motivation guided by empirical practice. The main research question is: "How do delayed updates to the data-generating policy affect learning in the REINFORCE algorithm?" This connects well with on-policy and off-policy learning. Due to my lack of familiarity with the topic and the constrained time to provide a review, I'm not able to fully judge, for instance, the contribution of the provided theoretical results compared to previous work in the literature. My current impression is that the contribution may be somewhat incremental, with limited empirical analyses. I'll remain open to raising my initial score.

UPDATE: I have raised my initial score.

Critical strengths

Clearly written and easy to follow.

Critical weaknesses

While the use of algorithms such as REINFORCE and its applications to LLM fine-tuning serves as motivation for the work, the authors acknowledge in the limitations section that further empirical analyses would be an extension of the work to bridge the theoretical analyses provided with their application. Therefore, I don't think the lack of practical experiments is a weakness due to the theoretical nature of the work. However, I believe the authors could expand the given empirical analyses to highlight the results; for instance, how the initial probability of the best action impacts the results when keeping K and S fixed. This analysis seems to be particularly important due to assumptions made in Thm 3.6.

Questions for authors

I don't have any questions.

3 comments
AuthorsOpenReview ↗

Response to Reviewer q3Yq

Thank you for your thoughtful and constructive review! We address your questions and concerns in the following.


Question: Contribution of the provided theoretical results compared to previous work in the literature.

We thank the reviewer for this question and appreciate the opportunity to clarify the contribution of our work. To compare our theoretical results with previous work:

  • Regarding problem formulation: Several works [1, 2] study reward-guided self-training, but they assume S=∞S=\infty. Another line of work [3, 4] studies softmax policy gradient algorithms, but they focus on the on-policy setting. In contrast, we study the more challenging and practical setting with a finite S≥1S \geq 1.
  • Regarding theoretical techniques: As discussed in Remark 3.5, existing theoretical techniques cannot be transferred to our setting due to (i) incomplete fitting within each stage, (ii) vanishing inner-loop gradients at suboptimal stage targets, and (iii) sign mismatch between off-policy updates and true policy gradients. We overcome these difficulties through new analytical techniques.
  • Furthermore, our work has revealed counterintuitive and important findings — e.g., the provable benefits of off-policyness as described in Section 3.3 — which are novel results to the best of our knowledge.

[1]: Štrupl, Miroslav, et al. "Reward-weighted regression converges to a global optimum." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 36. No. 8. 2022.

[2]: Arnal, Charles, et al. "Asymmetric reinforce for off-policy reinforcement learning: Balancing positive and negative rewards, 2025." arXiv preprint arXiv:2506.20520.

[3]: Mei, Jincheng, et al. "On the global convergence rates of softmax policy gradient methods." International conference on machine learning. PMLR, 2020.

[4]: Liu, Jiacai, Wenye Li, and Ke Wei. "Elementary analysis of policy gradient methods." arXiv preprint arXiv:2404.03372.


Suggestions for additional experiments.

We appreciate the reviewer for your valuable suggestion! We agree that empirically examining the role of initialization can make our work stronger. Following your suggestion, we have conducted additional experiments that investigate the impacts of the initial best-action probability xx. The empirical results can be found in Figures 3 and 4 in Appendix D.1 of our revised manuscript. It is shown that as xx decreases, on-policy RE(1)\text{RE}(1) and RE(S)\text{RE}(S) with small SS plateau at the suboptimal gap for a long period, whereas increasing SS progressively lowers the threshold of xx needed to avoid the trap and enables steady convergence to the global optimum, confirming Theorem 3.6.

Besides these, we have added other experiments to the revised manuscript, including:

  • Figure 2 (left): we have added lines for more values of SS.
  • Figures 5 and 6 in Appendix D.2: experiments for finite-sample settings, where we investigate Algorithm 1 with batch sizes N∈32,256N \in \\{32, 256\\} over 2020 random seeds. These results — as finite-sample counterparts of Figures 1 and 2 respectively — suggest that our main findings continue to hold in the finite-sample setting: Figure 5 shows that RE(SS) maintains the Θ(1/t)\Theta(1/t) asymptotic convergence rate, and Figure 6 shows that RE(SS) with a suitable SS can avoid the suboptimal trap that delays on-policy RE(1) and converges much faster to the global optimum.

Thank you again for your review, which helps us a lot in improving the current work! We hope that our clarifications on comparisons with prior literature and the newly added empirical experiments have provided satisfactory answers to your questions. In light of these revisions, we would greatly appreciate it if you would consider updating your rating of our work accordingly :)

Thank you. I have updated my initial score to reflect the new additions to the paper.

Thank you for taking the time to read our response and update your score! We are glad to know that you find the new additions useful, and grateful for your constructive suggestions that have helped us a lot in improving the quality of this work.

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

The paper studies RE(SS), a reward-weighted fine-tuning algorithm that refreshes its rollout distribution every SS gradient steps. It interprets each stage as fitting a reward-weighted target through forward-KL minimization. For softmax bandits with exact gradients, it proves global convergence for any fixed SS, two-sided asymptotic Θ(1/T)\Theta(1/T) rates, and a weak-initialization regime where suitable off-policy staging converges much faster than on-policy updates.

Critical strengths
  1. The analysis covers finite inner-loop optimization for every fixed SS, rather than requiring each reward-weighted target to be fitted to convergence. Global convergence and matching Θ(1/T)\Theta(1/T) rates in total gradient steps provide a substantive extension of the stage-wise interpretation.
  2. The authors carefully examine difficult initializations in which the optimal action has very low probability, exposing behavior that the shared asymptotic convergence rate alone does not capture. Theorem 3.6 shows that on-policy updates can drive the policy toward a suboptimal corner and delay recovery, while suitably chosen off-policy stages avoid this detour. I find this analysis particularly valuable because it identifies a concrete benefit of off-policy learning beyond tolerating stale rollouts, and makes the initialization conditions supporting that benefit explicit.
Critical weaknesses

The rate constants depend on the trajectory-dependent minimum optimal-action probability, which may be very small and may itself depend on SS. The explicit lower bound on that probability requires a favorable initialization. Consequently, the shared Θ(1/T)\Theta(1/T) order does not by itself establish that SS affects only burn-in under arbitrary initialization: SS could also enter the asymptotic constants through the optimization trajectory. Clarifying this dependence is important to assessing one of the paper's central conclusions about rollout staleness.

Questions for authors

Clarification: Does the claim that SS affects only burn-in concern the asymptotic order alone, or also the rate constants? The stated upper bound involves the trajectory-dependent minimum optimal-action probability, which may depend on SS. Can the analysis establish SS-independent asymptotic constants after an SS-dependent transient?

2 comments
AuthorsOpenReview ↗

Response to Reviewer XSNg

Thank you for your thoughtful and constructive review! We address your questions and concerns in the following.


Question: “Does the claim that SS affects only burn-in concern the asymptotic order alone, or also the rate constants?” “Can the analysis establish SS-independent asymptotic constants after an SS-dependent transient?”

Thank you for the insightful and important questions! The short answer is yes: our claim that SS affects only the burn-in phase holds not only for the asymptotic Θ(1/T)\Theta(1/T) order, but also for the rate constants.

In our original statement of Theorem 3.3, to present a single unified bound across all stages, we uniformly lower-bounded πb,0(1)\pi_{b,0}(1) by its whole-trajectory minimum pminp_{\text{min}}, which was slightly conservative. With a straightforward additional derivation based on existing analysis in our manuscript, we can strengthen Theorem 3.3 and confirm our original claim: after an SS-dependent transient phase, the rate constant in the upper bound is completely independent of SS and the optimization trajectory.

The core proof flow is as follows:

  1. First, applying Lemma 3.2 at stage bb, as in the proof of Eq.27, gives the single-stage recurrence inequality that depends only on the optimal-action probability πb,0(1)\pi_{b,0}(1) at the start of the current stage: 1D(πb,s)−1db≥λ(πb,0(1))min⁡ηs,1db\frac{1}{D(\pi_{b,s})} - \frac{1}{d_b} \ge \lambda(\pi_{b,0}(1)) \min\\{\eta s, \frac{1}{d_b}\\} where the constants in function λ(⋅)\lambda(\cdot) (defined in Appendix B.1) depends only on μ\mu and η\eta.

  2. By Theorem 3.1, for any finite initialization, db=D(πb,0)d_b = D(\pi_{b,0}) is monotonically non-increasing and converges to 00, which together with db≥Δ(1−πb,0(1))d_b \ge \Delta(1 - \pi_{b,0}(1)) implies πb,0(1)→1\pi_{b,0}(1) \to 1. Therefore, there exists a finite stage b~\tilde{b} (corresponding to an SS-dependent transient step t~=b~S\tilde{t} = \tilde{b} S) such that 1/db≥ηS1/d_b \ge \eta S and πb,0(1)≥1/2\pi_{b,0}(1) \ge 1/2 hold for all b≥b~b \ge \tilde{b}. Substituting λ(πb,0(1))≥λ(1/2)\lambda(\pi_{b,0}(1)) \ge \lambda(1/2) into the recurrence inequality above and summing from stage b~\tilde{b} to bb gives, for all b≥b~b \ge \tilde{b}: D(πb,s)≤1db~−1+λ(1/2)η(t−t~)D(\pi_{b,s}) \le \frac{1}{d_{\tilde{b}}^{-1} + \lambda(1/2)\eta(t - \tilde{t})} Here the asymptotic constant λ(1/2)\lambda(1/2) depends only on μ\mu and η\eta, and is independent of SS and pmin⁡p_{\min}.

Consequently, the potential dependence of early-stage pmin⁡p_{\min} on SS affects only the transient time t~\tilde{t} required to reach this local regime, not the asymptotic rate constants.

We have incorporated this refined statement into Theorem 3.3 (and the corresponding proofs into Appendix B.2 in page 20) in our revised manuscript.


Thank you again for your review, which helps us a lot in improving the current work! We hope that our responses and revisions have provided satisfactory answers to your questions, and we would greatly appreciate it if you would consider updating your rating of our work accordingly :)

To further supplement our response above, we can give an explicit upper bound on the length of the burn-in phase b~\tilde{b}. As we stated above, the b~\tilde{b} only need to satisfy 1/db~≥ηS1/d_{\tilde{b}} \ge \eta S and πb~,0(1)≥1/2\pi_{\tilde{b},0}(1) \ge 1/2. These two conditions hold when we define b~=min⁡b≥0:db≤min⁡Δ/2,1/(ηS)\tilde{b} = \min\\{b \ge 0 : d_b \le \min\\{\Delta/2, 1/(\eta S)\\}\\}, by noticing that db≥Δ(1−πb,0(1))d_b \ge \Delta(1 - \pi_{b,0}(1)).

At stage bˉ\bar{b}, the proof in Theorem 3.3 gives $ d_{\bar{b}}^{-1} \geq \eta S $. After another nn stages, we have dbˉ+n−1≥dbˉ−1+cηSn≥ηS(1+cn),c=λ(pmin)d_{\bar{b}+n}^{-1} \geq d_{\bar{b}}^{-1} + c\eta Sn \geq \eta S(1+cn), c = \lambda(p_{\text{min}})

The bound already meets the ηS\eta S threshold. So we only need to have ηS(1+cn)≥2/Δ\eta S(1+cn) \geq 2/\Delta. Therefore, for any S≥1S \geq 1, we have b~≤bˉ+⌈max⁡(2/(ηSΔ)−1,0)c⌉=⌈max⁡(2/(ηSΔ)−1,0)c⌉(if ηSd0≤1)or⌈log⁡(ηSd0)log⁡(1+c)⌉+⌈max⁡(2/(ηSΔ)−1,0)c⌉(if ηSd0>1).\tilde{b} \le \bar{b} + \left\lceil \frac{\max(2/(\eta S\Delta) - 1,0)}{c} \right\rceil = \left\lceil \frac{\max(2/(\eta S\Delta) - 1,0)}{c} \right\rceil (\text{if } \eta S d_0\leq 1) \text{or} \left\lceil\frac{\log(\eta S d_0)}{\log(1+c)}\right\rceil + \left\lceil \frac{\max(2/(\eta S\Delta) - 1,0)}{c} \right\rceil (\text{if } \eta S d_0 > 1).

We hope this additional response helps address your question regarding the transient phase. Please feel free to let us know if you have any further questions or comments, and we would be more than happy to discuss!

Reviewer_DCWFOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies theoretical properties of a version of the off-policy REINFORCE algorithm. In particular, the authors consider the bandit setting with simple softmax policies and show that the algorithm (operating on the population objective) converges with an Θ(t−1)\Theta(t^{-1}) rate. Furthermore, the authors consider a case where the policy is heavily mis-initialized and provide there a separation between the on-policy and an off-policy algorithm.

Critical strengths

The paper is very well written, with clear motivation and description of the contributions. The paper provides a theoretical analysis of an off-policy version of the REINFORCE algorithm, which is a nice and timely contribution to the field of theoretical RL. The separation between on-policy and off-policy in the case of poorly initialized policies is a very interesting finding.

Critical weaknesses

I did not identify any critical weaknesses in the manuscript. Arguably, the contributions of the paper are a bit thin, as it is currently restricted to an idealized theoretical study (my main concern is the treatment of population objectives). I would have appreciated a further development of Section 3.3 and perhaps some experiments in more complicated settings.

Questions for authors
  1. How general do you think the separation of Section 3.3 is? I understand that the choice of SxS_x is not very specific, but can we say more general for more values of SS? Also, do you think it is possible to identify conditions of the initialization that predict whether there is a separation between on-policy and off-policy methods?
  2. Can you add lines for more values of SS in Figure 2 (left)? You could also perhaps sweep the original probability xx. This would be a nice addition to the paper.
  3. Please define BB in Algorithm 1. (I know it is defined in the introduction as T/ST/S).
  4. I noticed a typo in Line 697: you forgot to close the parenthesis in the argument of the final gradient of that line.
4 comments
AuthorsOpenReview ↗

Response to Reviewer DCWF (1)

Thank you very much for your support, positive evaluation, and insightful feedback on our work! We have carefully addressed your questions and incorporated your suggestions into the revised paper, as detailed below.


Question about Section 3.3: “How general do you think the separation of Section 3.3 is? I understand that the choice of SxS_x is not very specific, but can we say more general for more values of SS?”

Thank you for this insightful question. Indeed, our analysis in Theorem 3.6 naturally extends to a broader range of staleness SS beyond the stated choice SxS_x, which is shown below.

Recall that in Appendix C.2, the KL fitting error within each stage has an O(1/S)O(1/S) upper bound (Lemma C.2 in our paper). Under the assumptions of Section 3.3, and for sufficiently small xx, once SS reaches the threshold Sx=⌈cSx−1⌉S_x = \lceil c_S x^{-1} \rceil (where the explicit constant cSc_S is defined in Eq. 49), each stage before the target accuracy is reached is guaranteed to increase the optimal action's log-odds by at least Δ2μmax⁡\frac{\Delta}{2\mu_{\max}} (Lemma C.3 and Eq. 50). Reaching accuracy ϵ\epsilon therefore requires at most Bx=⌈(4μmax⁡/Δ)log⁡(1/x)⌉B_x = \lceil (4\mu_{\max}/\Delta)\log(1/x) \rceil stages (Eq. 51). Since the fitting error bound decreases with SS, the argument leading to Eqs. (50) and (51) remains valid for any S≥⌈cSx−1⌉S \ge \lceil c_S x^{-1} \rceil, and the total number of gradient steps satisfies the explicit upper bound Tϵ(S)≤BxST_\epsilon(S) \le B_x S.

Comparing this with the lower bound Tϵ(1)≥cx−(K−1)T_\epsilon(1) \ge c x^{-(K-1)} for RE(1) in Appendix C.1, where c>0c>0 is independent of xx, we see that for any integer SS in the following interval: ⌈cSx−1⌉≤S<cx−(K−1)Bx\left\lceil c_S x^{-1} \right\rceil \le S < \frac{c x^{-(K-1)}}{B_x}, we strictly have Tϵ(S)<Tϵ(1)T_\epsilon(S) < T_\epsilon(1). If, in addition, S=S(x)S=S(x) satisfies BxS(x)xK−1→0B_xS(x)x^{K-1} \to 0 as $ x \to 0 $, then $ T_\epsilon(S(x))/T_\epsilon(1) \le c^{-1}B_xS(x)x^{K-1} \longrightarrow 0$.


Question about Section 3.3: “Do you think it is possible to identify conditions of the initialization that predict whether there is a separation between on-policy and off-policy methods?”

We appreciate the reviewer for raising this question. Building upon the asymptotic condition in Theorem 3.6 (where Tϵ(S)/Tϵ(1)→0T_\epsilon(S)/T_\epsilon(1) \to 0 as x→0x \to 0 for suitable choices of SS), we can indeed go a step further and establish explicit non-asymptotic conditions. For a non-asymptotic separation, we can provide — based on the analyses from our paper — explicit (though conservative) sufficient conditions on the initialization and SS under which we have $T_\epsilon(S)/T_\epsilon(1) < \gamma $, where γ\gamma is an arbitrary real number in (0,1)(0,1):

  1. κ:=J(π0,0)−max⁡3≤a≤Kμ(a)>0\kappa := J(\pi_{0,0}) - \max_{3 \le a \le K} \mu(a) > 0,
  2. 0<x≤μmin⁡ϵμmax⁡(μmax⁡−μmin⁡)0 < x \le \frac{\mu_{\min}\epsilon}{\mu_{\max}(\mu_{\max} - \mu_{\min})},
  3. The positive integer SS satisfies ⌈cSx−1⌉≤S<γ⋅C0−1Φ(π0,0)−22ημmax⁡Bx\left\lceil c_S x^{-1} \right\rceil \le S < \gamma \cdot \frac{C_0^{-1}\Phi(\pi_{0,0}) - 2}{2\eta\mu_{\max} B_x}, where Φ(π0,0)=∏a=2Kπ0,0(a)π0,0(1)K−1\Phi(\pi_{0,0}) = \frac{\prod_{a=2}^K \pi_{0,0}(a)}{\pi_{0,0}(1)^{K-1}}, C0=eK(1+ημmax⁡)(1+(K−2)μmax⁡κ)(μmax⁡κ)K−2C_0 = e^{K(1+\eta\mu_{\max})}\big(1 + \frac{(K-2)\mu_{\max}}{\kappa}\big)\big(\frac{\mu_{\max}}{\kappa}\big)^{K-2}, and cSc_S and BxB_x are defined in Eqs. (49) and (51), respectively.

Proof: Under Condition 1 and the monotonicity of J(πb,0)J(\pi_{b,0}) (Eqs. 16), Eq. (37) in Appendix C.1 holds directly from b0=0b_0=0 without taking x→0x\to0. Using bϵb_\epsilon and bcb_c as defined in Appendix C.1, we have $ \ell_{b_\epsilon}> \frac1K\log\Phi(\pi_{0,0})>1$. The first inequality follows from Eq. (39) and πbϵ,0(1)>1/2\pi_{b_\epsilon,0}(1)>1/2 (since ϵ<Δ/2\epsilon<\Delta/2, with Δ\Delta as defined in Section 3.3), and the second from Φ(π0,0)>2C0>eK\Phi(\pi_{0,0})>2C_0>e^K, by Condition 3 and the definition of C0C_0. Thus bcb_c exists with bc≤bϵb_c\le b_\epsilon. Substituting b0=0b_0=0 into Eqs. (39) to (43) gives 1−πbc,0(2)≤C0Φ(π0,0)<121-\pi_{b_c,0}(2) \le \frac{C_0}{\Phi(\pi_{0,0})}<\frac12 at stage bcb_c. Since πbc,0(1)≤1−πbc,0(2)<1/2<πbϵ,0(1)\pi_{b_c,0}(1)\le1-\pi_{b_c,0}(2)<1/2<\pi_{b_\epsilon,0}(1), we have bc<bϵb_c<b_\epsilon. Lemma C.1 directly yields the non-asymptotic lower bound

\frac{C_0^{-1}\Phi(\pi_{0,0})-2}{2\eta\mu_{\max}}.$$ Meanwhile, under Condition 2 (Eq. 47), Eqs. (49) to (51) in Appendix C.2 directly guarantee the non-asymptotic upper bound $T_\epsilon(S)\le B_xS$ for any $S\ge\lceil c_Sx^{-1}\rceil$. Combining these two bounds with Condition 3 immediately gives $T_\epsilon(S)/T_\epsilon(1)<\gamma$, which concludes our proof.
AuthorsOpenReview ↗

Response to Reviewer DCWF (2)

Supplementary experiments on your comment regarding the finite-sample setting.

Thank you for your constructive comment. In our paper, we focused on the exact-gradient dynamics to cleanly isolate the core algorithmic mechanisms of off-policy staleness SS. Inspired by your comment, we have further conducted experiments where the gradient is estimated from finite samples, to verify how our theoretical insights carry over to the finite-sample settings.

The empirical results can be found in Figures 5 and 6 in Appendix D.2 in the revised manuscript, where we investigate Algorithm 1 with batch sizes N∈32,256N \in \\{32, 256\\} over 2020 random seeds. These results — as finite-sample counterparts of Figures 1 and 2 respectively — suggest that our main findings continue to hold in the finite-sample setting: Figure 5 shows that RE(SS) still follows Θ(1/t)\Theta(1/t) asymptotic convergence rate, and Figure 6 shows that RE(SS) with a suitable SS can avoid the suboptimal trap that delays on-policy RE(1) and converges much faster to the global optimum.


Suggestions for additional experiments.

Thank you for these constructive suggestions! In our revised manuscript, we have expanded our empirical experiments as follows:

  • Figure 2 (left): we have added lines for more values of SS.
  • Figures 3 and 4 in Appendix D.1: experiments that sweep the value of the initial optimal-action probability xx. The results confirm that as xx decreases, on-policy RE(1)\text{RE}(1) and small-SS RE(S)\text{RE}(S) get trapped at the suboptimality plateau, whereas RE(S)\text{RE}(S) with larger SS still converges steadily. An appropriately larger SS lowers the required threshold of xx for avoiding this suboptimality trap.

Suggestions: fix minor issues.

We appreciate the reviewer's careful reading. In our revised manuscript, we have clearly defined BB as ⌊T/S⌋\lfloor T / S \rfloor in Section 2 before Algorithm 1, and fixed the typo in Line 697.


Thank you again for your support and valuable feedback, which helps us a lot in improving the current work! We hope that our responses and revisions have provided satisfactory answers to your questions.

Reviewer_DCWFOpenReview ↗

Thanks

Thank you for your comments and clarifications. I maintain my already high score.

Thank you for your reply! We sincerely appreciate your encouraging feedback and constructive suggestions, which have helped us a lot in improving this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.