acceptodds
Under review as a conference paper at ICLR 2027

MDSeq: Sequence-Level Verification for Multi-Draft Speculative Decoding

Abstract

Speculative decoding has emerged as a pivotal paradigm for accelerating large language model inference through a draft-then-verify mechanism. To further enhance efficiency, modern frameworks have transitioned from single-draft to multi-draft to increase the expected acceptance length. However, existing methods primarily rely on token-level verification, which leads to suboptimal performance. In this work, we propose MDSeq, a novel multi-draft sequence-level verification framework. By carefully designing the acceptance probabilities and residual distributions, MDSeq combines the improved candidate coverage of multi-draft speculation with the benefits of jointly verifying entire sequences. We provide theoretical guarantees that MDSeq is valid and attains a near-optimal expected acceptance length. Extensive experiments across various models, scales and tasks demonstrate that MDSeq consistently outperforms current state-of-the-art methods, providing substantial acceleration gains. Ablation studies examine the impact of various hyperparameters on performance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM, inference, speculative decoding

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_MFxmOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies verification of multiple independently sampled draft sequences. It derives an upper bound on expected accepted length and proposes MDSeq, which can accept a longer prefix even when a shorter prefix failed its individual acceptance test. A scaling factor, residual sampling, and subsequent distribution modification compensate for the accepted probability mass. The paper proves target-distribution preservation and, under a relative depth-level acceptance assumption, a constant-factor lower bound relative to the derived upper bound. Experiments on Spec-Bench report higher average accepted length and modest wall-clock improvements over the strongest evaluated baselines.

Critical strengths
  1. The upper bound accounts for both the chance that a prefix appears among the sampled drafts and the probability mass the target assigns to it. The proof of distribution preservation includes residual sampling and the modification of future distributions. The worked example in Appendix A helps explain why that correction is necessary.

  2. The comparison with SpecTr and recursive rejection sampling uses the same five independent length-five drafts, making this part of the experimental comparison informative about the verifier itself. MDSeq achieves the best reported average wall-clock speedup on both main model pairs, with additional results for a separate Tiny-Vicuna/Vicuna pair.

Critical weaknesses
  1. The near-optimality result depends on Assumption 1, but the paper does not establish when this assumption holds in practice. Uniformly relaxing the prefix threshold does not imply the required inequality between normalized acceptance rates. For example, take K=2K=2, L=2L=2, vocabulary {a,b}\lbrace a,b\rbrace, and p(aa)=p(ab)=p(ba)=p(bb)=0.25p(aa)=p(ab)=p(ba)=p(bb)=0.25. Set (q(aa),q(ab),q(ba),q(bb))=(0.89,0.01,0.05,0.05)(q(aa),q(ab),q(ba),q(bb))=(0.89,0.01,0.05,0.05). The root is ρ⋆≈1.430074\rho^\star\approx1.430074, but

    β2(ρ⋆)β1(ρ⋆)≈0.573616<0.610000=β2(2)β1(2).\frac{\beta_2(\rho^\star)}{\beta_1(\rho^\star)}\approx0.573616<0.610000=\frac{\beta_2(2)}{\beta_1(2)}.

    Thus, Assumption 1 can fail even in a small example. This does not contradict the conditional theorem or the distribution-preservation result. The description of the condition as mild needs justification, and the abstract should state that near-optimality is conditional.

  2. The wall-clock gains over the best baselines are small, and their variability is not reported. Average speedup improves from 1.87 to 1.94 on Vicuna and from 1.82 to 1.86 on Llama2, but only means over three runs are given. A timing breakdown is also needed for solving for ρ⋆\rho^\star, evaluating residuals, and maintaining the modified distributions after correction. These operations are necessary for exact verification, and their cost matters when the relative speed advantage is only 2.2–3.7%.

Questions for authors
  1. Can you provide sufficient conditions for Assumption 1, or measurements of how often it holds and how much it is violated in the evaluated draft–target pairs? What can be said about near-optimality when Assumption 1 fails, as in the example above?

  2. How is Algorithm 2 represented and evaluated after a correction introduces a prefix outside the original draft paths? What happens if a new speculative round starts while some positions are still subject to a previous distribution modification? Please explain which model evaluations are needed and whether all these costs are included in WS.

  3. What is the run-to-run variability of the wall-clock comparison with RRSw and SpecTr? How much time is spent on drafting, target scoring, and verifier calculations?

2 comments

Response to W1 and Q1:

Can you provide sufficient conditions for Assumption 1, or measurements of how often it holds and how much it is violated in the evaluated draft–target pairs? What can be said about near-optimality when Assumption 1 fails, as in the example above?

Thanks for the comment. We have added Appendix G, Discussion of Assumption 1, to further investigate the necessity of this assumption and provide both theoretical and empirical evidence supporting our method.

  1. Theoretical guarantee without Assumption 1. We prove that MDSeq achieves an expected acceptance length at least as large as that of Greedy Block Verification [1], without requiring Assumption 1. Specifically, EMDSeq[τ]≥EBV[τ]\mathbb E_{\mathrm{MDSeq}}[\tau]\geq \mathbb E_{\mathrm{BV}}[\tau].

  2. Assumption-free lower bound for ρ=K\rho=K. We further prove that, when the division factor is set to ρ=K\rho=K, the lower bound on the expected acceptance length in Theorem 3 holds without Assumption 1:

Eρ=K[τ]≥[1−(1−1K)K]⋅∑i=1L∑xi∈Vimin⁡(1−(1−p(xi))K,q(xi)).\mathbb E_{\rho=K}[\tau] \geq \left[1-\left(1-\frac1K\right)^K\right]\cdot\sum_{i=1}^{L}\sum_{x^i\in\mathcal V^i}\min(1-(1-p(x^i))^K,q(x^i)).

  1. Empirical evidence for the practical choice of ρ\rho. Recall that βi(ρ)=∑xi∈Vimin⁡(q(xi)ρ,p(xi))\beta_i(\rho)=\sum_{x^i\in\mathcal{V}^i}\min\left(\frac{q(x^i)}{\rho},p(x^i)\right). Computing βi(ρ)\beta_i(\rho) exactly requires summing over all possible prefixes in Vi\mathcal V^i, making direct verification of Assumption 1 computationally intractable in practice.

To further investigate its practical implications, we compare the BE of MDSeq using ρ=ρ∗\rho=\rho^* and ρ=K\rho=K. Our experiments show that ρ=ρ∗\rho=\rho^* consistently achieves higher BE than ρ=K\rho=K across all evaluated datasets. Since the lower bound is guaranteed to hold for ρ=K\rho=K without Assumption 1, the consistently superior empirical performance of ρ=ρ∗\rho=\rho^* suggests that this lower-bound guarantee also holds for ρ=ρ∗\rho=\rho^* across the evaluated workloads.

Dataρ=ρ∗\rho=\rho^* WSρ=ρ∗\rho=\rho^* BEρ=K\rho=K WSρ=K\rho=K BE
Math2.023.321.822.89
QA1.712.711.602.49
Trans.1.592.691.402.31
Sum.1.722.881.582.61
Con.2.113.421.882.98
RAG1.702.931.542.63
AVG1.812.991.642.65

Response to W2:

Only means over three runs are given.

We respectfully note that we have reported the standard deviations in Table 1 of the main experiments.

A timing breakdown is also needed for solving for ρ\rho, evaluating residuals, and maintaining the modified distributions after correction.

Thanks for the comment. We have added Appendix I, Computational Overhead Breakdown, to provide a detailed analysis of the computational costs.

Distribution modification introduces only a small additional cost of 0.035 seconds per question, which is outweighed by the savings in drafting and target scoring.

MDSeq’s verifier cost is higher than RRSw’s primarily because MDSeq additionally solves for ρ\rho. However, MDSeq solves for ρ\rho only once per decoding iteration, whereas SpecTr requires multiple solves. This reduces MDSeq’s cumulative ρ\rho-solving cost to 0.022 seconds per question, compared with 0.061 seconds for SpecTr, and contributes to its lower overall verifier cost.

Methodρ\rho solvingAccTestResidual evaluationVerifier total
SpecTr0.061±0.0130.047±0.0110.004±0.0010.112±0.025
RRSw–0.036±0.0110.007±0.0020.043±0.012
MDSeq0.022±0.0060.038±0.0110.005±0.0010.066±0.019

Response to Q2:

How is Algorithm 2 represented and evaluated after a correction introduces a prefix outside the original draft paths?

Following greedy block verification [1], we evaluate Algorithm 2 lazily: we store only the parameters, prefix probability masses, and remaining horizon of each distribution modification, without explicitly constructing all possible continuations. When a correction token leaves the original draft paths, the next round generates new drafts from the actual emitted prefix and scores them with the target model, obtaining the conditional distributions required along the new paths.

Unfinished modifications persist across rounds and are applied in chronological order. Each new modification is constructed from the effective target distribution after applying the existing modifications. The corresponding states are updated whenever a token is emitted, until their horizons expire. Distribution modification involves only probability-vector operations and does not require an additional model forward pass for each modification. All costs, including processing corrected prefixes, subsequent drafting and target scoring, distribution modification, and KV-cache updates, are included in the end-to-end timing used to compute WS.

[1] Sun et al. Block Verification Accelerates Speculative Decoding. ICLR 2025.

Response to Q3:

What is the run-to-run variability of the wall-clock comparison with RRSw and SpecTr? How much time is spent on drafting, target scoring, and verifier calculations?

We have added Appendix I, "Computational Overhead Breakdown".We repeated each experiment three times and reported the run-to-run variability in the wall-clock time breakdown comparison with RRSw and SpecTr. The results show that MDSeq achieves the lowest total runtime and highest throughput among baseline methods. Specifically, MDSeq incurs lower cumulative drafting and target-scoring costs because its higher block efficiency (BE) allows more tokens to be generated per decoding iteration, thereby reducing the number of iterations required to produce an output of a given length.

MethodDraftingTarget scoringVerifierDistModPrefillOtherTotalTokens/s ↑
SpecTr0.49±0.121.44±0.330.112±0.025–0.04±0.000.02±0.002.11±0.4877.65±2.41
RRSw0.45±0.131.42±0.420.043±0.012–0.04±0.000.02±0.011.98±0.5781.01±3.86
MDSeq0.39±0.111.13±0.320.066±0.0190.035±0.0100.04±0.000.02±0.001.68±0.4783.97±2.31
Reviewer_jt9HOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes MDSeq, a sequence-level verification algorithm for multi-draft speculative decoding. Theoretically, the authors provides an upper bound on the expected acceptance length for multi-draft verification and prove that, under an additional assumption, MDSeq achieves at least a 1−1/e1-1/e fraction of this bound. Empirically, MDSeq improves average speed by 2%–4% over the strongest baselines.

Critical strengths
  1. Despite combining multi-draft speculative decoding and block verification is not new, the authors provide principled approach which offers theoretic guarantee. The proposed approach also can achieve better acceptance rate compared to prior strong baselines.
  2. The paper is clear and easy to follow.
  3. It makes clear theoretical contribution to multi-draft speculative decoding.
Critical weaknesses

My major concern is about the experiments. Currently, the authors only conduct experiments on two model pairs, which is very limited. Moreover, the authors do not test model with larger capacity, which is more important in real-world application.

Questions for authors

See above.

1 comment

Response to W1:

My major concern is about the experiments. Currently, the authors only conduct experiments on two model pairs, which is very limited. Moreover, the authors do not test model with larger capacity, which is more important in real-world application.

Thank you for your comment. In Appendix B, "Experiments on Vicuna-13B," we evaluate MDSeq using Tiny-Vicuna-1B as the draft model and Vicuna-13B-v1.3 as the target model, demonstrating its effectiveness with a different draft model architecture from EAGLE.

Furthermore, we have added Appendix K, "Experiments on Vicuna-33B," to evaluate MDSeq on a larger-scale model, using EAGLE-Vicuna-33B-v1.3 (0.56B) as the draft model and Vicuna-33B-v1.3 as the target model. The results show that MDSeq also outperforms the baselines in both block efficiency (BE) and wall-clock speedup (WS), further demonstrating its effectiveness across different draft models and target model scales.

DataVanilla WSVanilla BEBV WSBV BERRSw WSRRSw BESpecTr WSSpecTr BEMDSeq WSMDSeq BE
Math1.713.031.723.061.833.221.843.321.883.40
Con.1.762.691.782.831.692.941.673.001.763.12
QA1.392.211.402.281.402.521.352.491.442.64
RAG1.402.371.412.421.422.641.432.661.452.76
Sum.1.482.451.512.501.482.761.482.801.522.85
Trans.1.452.401.452.401.442.681.482.811.462.78
AVG1.532.531.552.581.542.791.542.851.592.93
Reviewer_eBLKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MDSeq Multi draft speculative decoding framework, that performs sequence level verifications to improve the expected acceptance length. Existing SD verification methods perform token level verification across multiple drafts or sequence verification of single draft. MDSeq, combines these both by considering sub-sequence probabilities across multiple drafts.

MDSeq introduces hierarchical acceptance probabilities to verify sub-sequences of different lengths and select the longest accepted prefix, even when shorter-prefix verification fails. This is combined with a global scaling factor to control the acceptance probability across multiple drafts and a correction step to ensure the output follows the target model's distribution. The paper derives an upper bound on the expected acceptance length for any valid multi-sequence verification algorithm. It also establishes a validity guarantee and derives a lower bound on its expected acceptance length under an additional assumption.

Experiments on Spec-Bench dataset show improvements in block efficiency and wall clock speedup over existing verification algorithms.

Critical strengths
  • The paper presents a novel approach that combines multi-draft SD with multi-sequence verification. This allows longer prefixes to be accepted even when shorter-prefixes or individual token level verifications fail. The theoretical formulation is rigorous and well motivated.

  • The results show that MDSeq acheives improvements to block efficiency and wallclock speed over several baselines and ablation study with the effect of hyper-parameters.

Critical weaknesses
  • The lowerbound near-optimality guarantee in Theorem 3 relies on Assumption 1, which the authors describe as mild. However, this assumption does not always hold, even for simple binary distributions. For example, let K=L=2K=L=2, with p(aa)=p(ba)=0.4p(aa)=p(ba)=0.4, p(ab)=p(bb)=0.1p(ab)=p(bb)=0.1, and qq uniform over all four sequences. Here, ρ∗=1\rho^*=1, but β2(1)/β1(1)=0.7\beta_2(1)/\beta_1(1)=0.7 is less than β2(2)/β1(2)=0.9\beta_2(2)/\beta_1(2)=0.9, violating Assumption 1. A more permissive scaling factor can increase first-token acceptance without proportionally improving longer-prefix acceptance.

  • The paper does not study closely related GBV (Thomas & Pal, 2026) and SpecTr-GBV (Lin et al., 2026) in its experiments. Those both papers also cover multi-draft sequence-level verifications and would serve as important baselines to better establish the empirical advantages of MDSeq.

Questions for authors
  • Would it be possible for authors to establish a weaker approximation guarantee without this Assumption 1? or under what conditions of draft and target distributions is Assumption 1 true?

  • Are the upper and lower bounds the tighest bounds for multi-sequence verification? did you also verify them emperically? It would be great to see some plots for the same on real datasets, if any.

  • Can the authors include GBV and SpecTr-GBV in their experiment suite and highlight how this work stands out?

  • Can the authors quantify the computational overhead of MDSeq's verification and distribution modification steps, and how it scales with the number and length of drafts?

  • Can you add some reasoning/justification for various steps in the Theorem 3 proof in Appendix?

6 comments

Response to W1 and Q1:

Would it be possible for authors to establish a weaker approximation guarantee without this Assumption 1? or under what conditions of draft and target distributions is Assumption 1 true?

Thank you for your comment. We have added Appendix G, "Discussion of Assumption 1," to further investigate the necessity of this assumption and provide both theoretical and empirical evidence supporting our method.

  1. Theoretical guarantee without Assumption 1. We prove that MDSeq achieves an expected acceptance length at least as large as that of Greedy Block Verification [1], without requiring Assumption 1. Specifically, EMDSeq[τ]≥EBV[τ]\mathbb E_{\mathrm{MDSeq}}[\tau]\geq \mathbb E_{\mathrm{BV}}[\tau].

  2. Assumption-free lower bound for ρ=K\rho=K. We further prove that, when the division factor is set to ρ=K\rho=K, the lower bound on the expected acceptance length in Theorem 3 holds without Assumption 1:

    Eρ=K[τ]≥[1−(1−1K)K]⋅∑i=1L∑xi∈Vimin⁡(1−(1−p(xi))K,q(xi)).\mathbb E_{\rho=K}[\tau] \geq \left[1-\left(1-\frac1K\right)^K\right]\cdot\sum_{i=1}^{L}\sum_{x^i\in\mathcal V^i}\min(1-(1-p(x^i))^K,q(x^i)).

  3. Empirical evidence for the practical choice of ρ\rho. Recall that βi(ρ)=∑xi∈Vimin⁡(q(xi)ρ,p(xi))\beta_i(\rho)=\sum_{x^i\in\mathcal{V}^i}\min\left(\frac{q(x^i)}{\rho},p(x^i)\right) . Computing βi(ρ)\beta_i(\rho) exactly requires summing over all possible prefixes in Vi\mathcal V^i, making direct verification of Assumption 1 computationally intractable in practice.

    To further investigate its practical implications, we compare the block efficiency (BE) of MDSeq using ρ=ρ∗\rho=\rho^* and ρ=K\rho=K. Our experiments show that ρ=ρ∗\rho=\rho^* consistently achieves higher BE than ρ=K\rho=K across all evaluated datasets. Since the lower bound is guaranteed to hold for ρ=K\rho=K without Assumption 1, the consistently superior empirical performance of ρ=ρ∗\rho=\rho^* suggests that this lower-bound guarantee also holds for ρ=ρ∗\rho=\rho^* across the evaluated workloads.

Dataρ=ρ∗\rho=\rho^* WSρ=ρ∗\rho=\rho^* BEρ=K\rho=K WSρ=K\rho=K BE
Math2.023.321.822.89
QA1.712.711.602.49
Trans.1.592.691.402.31
Sum.1.722.881.582.61
Con.2.113.421.882.98
RAG1.702.931.542.63
AVG1.812.991.642.65

[1] Sun et al. (2025). Block Verification Accelerates Speculative Decoding. In International Conference on Learning Representations (ICLR).


Response to W2 and Q3:

Can the authors include GBV and SpecTr-GBV in their experiment suite and highlight how this work stands out?

Thank you for your comment. We have added Appendix J, "Experiments on GBV," to provide a direct comparison with GBV. We adopt the same experimental setting as in our main experiments, using EAGLE-Vicuna-7B-v1.3 (0.24B) as the draft model and Vicuna-7B-v1.3 as the target model. The experimental results demonstrate that MDSeq consistently outperforms GBV in both BE and WS.

Regarding SpecTr-GBV, we would like to clarify that it does not provide an unconditional lossless guarantee. Its proof of distribution preservation relies on the assumption of conditionally independent couplings, which is not generally satisfied in practice. Consequently, SpecTr-GBV does not guarantee exact preservation of the target distribution without this assumption. In this sense, it can be viewed as a form of relaxed speculative decoding that potentially improves acceptance rates at the expense of distributional fidelity. Therefore, we believe that a direct performance comparison with MDSeq, which guarantees lossless decoding, would not be an apples-to-apples comparison.

DataGBV WSGBV BEMDSeq WSMDSeq BE
Math1.99±0.033.26±0.032.21±0.043.40±0.04
QA1.60±0.022.63±0.011.80±0.042.72±0.04
Trans.1.53±0.032.56±0.061.74±0.022.69±0.01
Sum.1.56±0.012.81±0.031.87±0.032.93±0.01
Con.1.85±0.073.14±0.092.22±0.033.36±0.03
RAG1.70±0.062.91±0.081.80±0.043.02±0.03
AVG1.702.891.943.02

Response to Q2:

Are the upper and lower bounds the tighest bounds for multi-sequence verification? did you also verify them emperically? It would be great to see some plots for the same on real datasets, if any.

Thank you for your comment. We have added Appendix H, "Tightness of the Acceptance-Length Bounds," to provide a more detailed theoretical and empirical analysis.

  1. Complexity of the optimal multi-sequence verification problem. We first clarify that optimal multi-sequence verification can be formulated as an optimal transport (OT) problem involving ∣V∣L(K+1)|\mathcal{V}|^{L(K+1)} transport variables and ∣V∣LK+∣V∣L|\mathcal{V}|^{LK}+|\mathcal{V}|^L marginal constraints. Due to the complexity of this formulation, the exact optimum generally does not admit a closed-form expression, making it challenging to derive the tightest possible upper bound analytically.

  2. Approximation guarantee of our lower bound. We clarify that there is no unique notion of the "tightest lower bound" without specifying a particular class of bounds. More importantly, Theorem 3 establishes that our lower bound approximates the optimal expected acceptance length within a multiplicative factor of at least 1−1/e1-1/e, and MDSeq inherits the same approximation guarantee.

  3. Empirical evaluation of bound tightness. Since both the upper and lower bounds involve summations over all possible prefixes in Vi\mathcal{V}^i, evaluating them exactly on real-world language modeling distributions is computationally intractable. We therefore conduct controlled experiments using two synthetic distributions: bernoulli and uniform distributions. The experimental results verify that

    LowerBound≤EMDSeq[τ]≤OPT≤UpperBound,\mathrm{LowerBound}\leq\mathbb{E}_{\mathrm{MDSeq}}[\tau]\leq\mathrm{OPT}\leq\mathrm{UpperBound},

    where OPT\mathrm{OPT} denotes the optimal expected acceptance length obtained by solving the multi-sequence OT problem. Furthermore, we observe that our upper bound becomes increasingly tight as the number of draft sequences KK increases, progressively approaching the multi-sequence OT optimum.


Response to Q4:

Can the authors quantify the computational overhead of MDSeq's verification and distribution modification steps, and how it scales with the number and length of drafts?

Thank you for your comment. We have added Appendix I, Computational Overhead Breakdown, to provide a detailed analysis of the computational costs.

The results show that MDSeq achieves the lowest total runtime and highest throughput among baseline methods. Specifically, MDSeq incurs lower cumulative drafting and target-scoring costs because its higher BE allows more tokens to be generated per decoding iteration, thereby reducing the number of iterations required to produce an output of a given length. Furthermore, distribution modification introduces only a small additional overhead of 0.035 seconds per question, which is more than offset by the savings in drafting and target scoring. These results demonstrate that MDSeq's efficiency gains outweigh its additional computational overhead.

Effect of the number of drafts KK. Table 9 illustrates the trade-off between higher BE and greater computational cost per iteration. Increasing KK provides more candidate drafts, which can improve BE and reduce the number of decoding iterations needed to generate an output of a given length. However, constructing and scoring more candidates increases the cost of each iteration.

Effect of draft length LL. Table 10 shows a similar trade-off for draft length. Longer drafts allow more tokens to be accepted in each iteration, potentially increasing BE and reducing the number of decoding iterations. At the same time, they require more work per iteration for draft generation, target scoring, and distribution modification.

KKDrafting(s)Target scoring(s)Verifier(s)DistMod(s)Prefill(s)Other(s)Total(s)Tokens/s
10.31±0.091.54±0.460.092±0.0280.062±0.0170.04±0.000.02±0.012.07±0.6078.91±3.79
30.48±0.131.48±0.410.089±0.0240.051±0.0110.04±0.000.02±0.012.17±0.5877.23±3.61
50.39±0.111.13±0.320.066±0.0190.035±0.0100.04±0.000.02±0.001.68±0.4783.97±2.31
70.41±0.101.23±0.310.073±0.0190.038±0.0100.04±0.000.02±0.001.81±0.4581.06±0.29
LLDrafting(s)Target scoring(s)Verifier(s)DistMod(s)Prefill(s)Other(s)Total(s)Tokens/s
30.29±0.081.43±0.350.078±0.0170.014±0.0030.04±0.000.02±0.001.87±0.4582.83±4.78
50.39±0.111.13±0.320.066±0.0190.035±0.0100.04±0.000.02±0.001.68±0.4783.97±2.31
80.68±0.161.23±0.310.078±0.0190.091±0.0210.05±0.000.02±0.012.15±0.5174.80±0.65

Response to Q5:

Thank you for your helpful suggestion. We have added further explanations and justifications to the proof of Theorem 3 in the Appendix to improve its clarity.

Thank you again for your valuable feedback! As the discussion deadline is approaching, we would appreciate your feedback. We hope that we have fully addressed your concerns. If so, we would appreciate it if you could consider raising your score accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.