acceptodds
Under review as a conference paper at ICLR 2027

A Unified Framework for Token-Level Watermarks

Abstract

LLM watermarks allow tracing AI-generated texts by inserting a detectable signal into their generated content. Recent works have proposed a wide range of token-level watermarking algorithms, each with distinct designs. Crucially, there is currently no general and principled formulation for token-level LLM watermarking that could enable deeper understanding of current schemes and ease the design of new variants. In this work, we propose the first unified framework for token-level watermarks based on a principled constrained optimization problem. Our formulation unifies most existing and popular token-level watermarking methods, and explicitly reveals the constraints that each method optimizes. In the process, it highlights an underexplored quality-diversity-power trade-off. Our experimental evaluation validates our framework: watermarking schemes derived from a given constraint consistently maximize detection power with respect to that constraint. At the same time, the framework provides a principled approach for designing novel watermarking schemes tailored to specific requirements, for instance minimizing perplexity of the watermarked text.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM, watermarks

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

This paper studies how to design token-level watermark sampling rules while limiting changes to the language model's output. With the hashing rule and detector fixed, it maximizes the expected token score under constraints on either each sampling distribution or its average over random scores. The framework recovers several existing methods and introduces new sampling rules based on chi-square and perplexity constraints.

Critical strengths
  1. The framework connects existing methods. Connecting KGW, AAR/KTH, and SynthID to different constraints helps explain their sampling rules and differences.

  2. The new methods show measurable gains. With matched binomial scores and the same detector, Table 3 reports 98.8% detection at a PPL budget of 4 for soft PPL, compared with 73.5% for KGW.

  3. The experiments cover several important settings. The paper evaluates detection, text quality, diversity, and text-editing attacks. Comparisons with matched scores and detectors help show the contribution of the sampling rule.

Critical weaknesses
  1. The claims about optimal detection need stronger support. Maximizing the expected token score does not generally maximize detection power at a fixed false-positive rate, even for a fixed sum-based detector. The connection to the error-based guarantees in [1, 2] needs to be established.

  2. The two optimization problems are not generally equivalent. Equation (2) imposes a separate hard constraint for every score vector gg, which generally requires a multiplier depending on gg, whereas Equation (3) uses one constant penalty parameter.

  3. The KL calculation does not match the theoretical constraint. In the supplementary metrics.py, the hard-KL calculation uses KL(p∥q)KL(p\|q) instead of KL(q∥p)KL(q\|p) and divides by the vocabulary size for a one-dimensional input. Whether this code produced the reported constraint comparisons needs clarification.

References

[1] Li et al. A Statistical Framework of Watermarks for Large Language Models: Pivot, Detection Efficiency and Optimal Rules. The Annals of Statistics, 2025.

[2] Li et al. Robust Detection of Watermarks for Large Language Models Under Human Edits. Journal of the Royal Statistical Society Series B, 2026.

Questions for authors
  1. Connection to detection power. What guarantee on detection power or Type II error follows from maximizing the expected token score? Do the gains over AAR/KTH remain when that baseline uses the detection rules from [1] or Tr-GoF from [2]?

  2. Equivalence of the optimization problems. Under what additional conditions are Equations (2) and (3) equivalent for hard constraints? How can a single penalty parameter enforce the same constraint level across different score vectors?

  3. Verification of the KL results. Was the supplied metrics.py used for Figures 2 and 5 and Table 6? Please verify the KL direction and normalization, and explain whether correcting the implementation changes the reported conclusions.

2 comments

We are happy to see that the reviewer has found our work clearly presented, informative, and sound. We address their remaining concerns below, and have updated our submission. All changes in the submission are colored in blue for clarity.

Q1: Does the framework provide guarantees on detection power? Why do the authors use a sum-based objective?

We agree with the reviewer that maximizing EG[G⋅q(G)]\mathbb E_G[G\cdot q(G)] does not generally maximize detection power at a fixed FPR. Our theorems establish optimality with respect to the constraint and objective only, not to the power. Hence, we don’t provide a guarantee on the detection power. Prior works that provide guarantees [1,2] need to make assumptions on the LLM next-token distribution which we argue are impractical (e.g. [1]’s Δ\Delta-regularity requires every next-token probability to be at most 1−Δ1-\Delta, which fails on low-entropy tokens).

Importantly, we opted for a sum-based objective (i) because it captures existing schemes (and even new ones such as OpenAI’s TextGrain [3], now in Section 4.1) and is therefore practically relevant, and (ii) because maximizing power directly, as in [1,2], requires making unrealistic assumptions on the LLM next-token distribution which, as we see in Q2, does not yield better detection than a well-chosen sum-based detector and limits practicality.

Q2: For AAR, how does the sum-based detector compare with the optimal detector?

Thanks to the reviewer's comment, we now compare in Appendix F.1 (and in the table below, for distortion-free AAR) the AAR detector with the detector from [1] (with fixed Δ=0.1\Delta=0.1), as well as the sum-based Gamma detector from [4, 5]. We find that the Gamma detector performs on par with [1]’s detector: it is slightly better at 1% and 0.1% FPR, and slightly worse at 10−610^{-6} FPR. This further supports that sum-based detectors are appropriate for building powerful schemes.

DetectorTPR at 1% FPRTPR at 0.1% FPRTPR at 10−610^{-6} FPR
Original KS82.3%61.4%15.4%
Gamma sum92.1%82.4%41.8%
Li h0.1⋆h^\star_{0.1}91.0%80.9%44.9%

Q3: Under which conditions are equations 2 and 3 equivalent?

We don’t consider Equations 2 and 3 to be equivalent. In Section 3, we highlight that the shape of the optimal solution is the same, e.g. hard-KL constraint in both cases leads to exponential tilting, though in Equation 2 the multiplier may depend on gg. Thanks to the reviewer, we have now updated our phrasing in Section 3 to make our statement clearer.

Q4: Which KL metric was used for plotting Figure 2 and 5?

We thank the reviewer for actively checking our code. We indeed found that Figures 2 and 5 used KL(p∣q)KL(p|q) instead of KL(q∣p)KL(q|p). We have now updated Figures 2 and 5 as well as Table 8 (previously Table 6) to use KL(q∣p)KL(q|p), summed over the vocabulary without normalization by its size. Importantly, our conclusion remains the same: Red-Green watermarks are optimal when it comes to the trade-off between detectability and Hard KL.

[1] Li, Xiang, et al. "A statistical framework of watermarks for large language models: Pivot, detection efficiency and optimal rules." Annals of statistics 53.1 (2025): 322
[2] Li et al. “Robust Detection of Watermarks for Large Language Models Under Human Edits”. Journal of the Royal Statistical Society Series B, 2026
[3] Li, Xiang, et al. “textGrain: Entropy-Calibrated Watermarking for Language Model Text.” OpenAI technical report (2026)
[4] Sander, Tom, et al. “TextSeal: A Localized LLM Watermark for Provenance & Distillation Protection.” Advances in Neural Information Processing Systems (2026)
[5] Fernandez, Pierre, et al. “Three Bricks to Consolidate Watermarks for Large Language Models.” IEEE International Workshop on Information Forensics and Security (2023)

Thanks for the detailed response. My concerns have been addressed. I have thus raised my rating .

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

This paper introduces a unified constrained optimization framework for token-level LLM watermarking. The authors frame the sampling mechanism as maximizing expected pseudorandom scores (detectability) subject to statistical distortion bounds (quality/diversity). Through this formulation, prominent existing methods—including KGW, Aaronson/KTH, and SynthID—are unified and shown to be optimal under specific constraints (e.g., hard KL, soft KL, χ2\chi^2). Leveraging this insight, the authors introduce novel schemes that directly constrain perplexity (Hard/Soft PPL) to achieve state-of-the-art detectability-quality trade-offs. Additionally, the paper theoretically and empirically demonstrates an inherent quality-diversity-power trade-off, showing that soft-constrained (distortion-free) watermarks inevitably degrade output diversity under repeated prompts.

Critical strengths

 The paper does a great job unifying seemingly unrelated heuristics (KGW, AAR/KTH, SynthID) under a clean constrained optimization framework. It gives a very neat theoretical explanation of what each method actually optimizes under the hood.  The framework isn't just theoretical; it leads to a practical new method, PPL (Soft), which empirically beats existing baselines on the quality vs. detectability frontier across several LLMs.  The formal distinction between hard and soft constraints offers a crisp explanation for why "distortion-free" watermarks end up collapsing generation diversity under prompt repetition.  The experimental evaluation is thorough, testing constraint optimality, downstream benchmarks (HumanEval, GSM8K), robustness attacks, and score distributions.

Critical weaknesses

 High Computational Overhead for PPL (Soft): PPL (Soft) relies on per-step root-finding via bisection and Monte Carlo score sampling. This adds non-trivial inference latency compared to closed-form baselines, yet the paper lacks concrete wall-clock time and throughput (tokens/sec) benchmarks.  Scope Limited to Fixed Detectors: The theoretical optimality holds with respect to fixed hashing and sum-based detectors, which does not address the joint optimization of sampling and detection statistics.  Vulnerability to Heavy Paraphrasing: Similar to prior token-level schemes, detectability degrades sharply under strong paraphrase attacks (e.g., GPT-5.4 Mini or high-diversity DIPPER), indicating that token-level quality constraints alone cannot prevent semantic signal loss.  Coarse Diversity Metric: Self-BLEU captures n-gram repetitions but does not distinguish semantic collapse from syntactic variation. Semantic entropy or embedding-based metrics would provide a clearer assessment of diversity loss.

Questions for authors

 What are the actual inference latency and computational overheads of Soft-PPL and Hard-PPL in practice compared to KGW, AAR/KTH, and SynthID? Providing a direct efficiency benchmark (e.g., tokens/second or generation wall-clock time) would help clarify the deployment viability of the proposed schemes.  How sensitive is the detectability-quality trade-off of Algorithm 7 to the number of Monte Carlo samples nn? Can nn be lowered (e.g., n≤16n \le 16) to speed up generation without violating the target distortion constraint ε\varepsilon?  How well do the proposed schemes hold up against more intentional, adaptive watermark-removal attacks? Also, to solidify the claims around diversity collapse, have the authors considered validating the hard vs. soft constraint trade-off using semantic diversity metrics across a wider variety of models and datasets?

2 comments

We thank the reviewer for their positive review. We address their remaining concerns below, and have updated our submission. All changes in the submission are colored in blue for clarity.

Q1: What are the computational overheads of the proposed schemes?

Thanks to the reviewer's comment, we now additionally measure the wall clock time and throughput (tok/s) for all the evaluated schemes in Appendix E.7 and in the table below. As we can see, our new schemes (PPL hard and soft) have similar overhead to prior works.

Schemewall clock time (s)tok/s
Unwatermarked8.2303614.908
AAR/KTH8.2433605.331
KGW8.3113556.433
Chi28.3693516.357
PPL (Hard)8.2973602.350
PPL (Soft)8.3483562.708
SynthID8.5343490.980

Q2: Can you ablate the number of samples in Monte-Carlo for Soft-PPL?

We now measure in Appendix E.8 the empirical constraint value, (p−EG[q(G)])⋅log⁡p(p - \mathbb{E}_G[q(G)])\cdot \log p, with respect to the number of samples used for Monte-Carlo. We find that with 512 samples, the constraint is already approximately satisfied (at most 0.017 nats per token above ε\varepsilon). We also find that from 88 to 20482048 samples the generation speed (wall clock time) is the same (i.e. latency variations due to uncontrollable external factors, such as cluster load, are larger than those due to the number of samples).

Q3: Can the authors elaborate on how restrictive the scope of sum-based detectors is?

Certainly! First, we agree with the reviewer that we explicitly focus on sum-based detectors. However, we find that this does not limit the applicability of our approach but rather provides a reasonable set of assumptions that allow a clear unified framework. Importantly, it is also practical as it allows us to capture all currently deployed watermark families (Red-Green, AAR, SynthID, and even OpenAI’s TextGrain released after the submission deadline (now in Section 4.1)), and easily design new practical schemes (i.e. with reasonable latency).

Q4: How do the proposed schemes hold up against more targeted attacks?

Thanks to the reviewer's comment, in Appendix E.5 and in the table below, we now also measure robustness against paraphrasing with specific watermark removal paraphrasers [2]. In line with all prior works, adaptive paraphrasing remains an active challenge for LLM watermarks. Our soft-PPL scheme still retains the highest TPR, although most of its signal is removed.

SchemeOriginal TPR (%)Anti-watermark paraphrasing TPR (%)
PPL (Soft)99.13.7
PPL (Hard)72.02.7
AAR/KTH91.91.7
KGW13.70.7
χ²19.02.1
SynthID93.52.7

Q5: Can the diversity experiment be extended to semantic diversity metrics?

In Appendix E.9, we now extend our diversity-quality-power trade-off measurement to semantic diversity metrics. More precisely, using the same experimental protocol as in Section 5.4, we replace the Self-BLEU score with the average pairwise cosine similarity between replies, using MiniLM-L6 [3] as an embedder. The observed results remain similar to those with the Self-BLEU score.

[1] Li, Xiang, et al. “textGrain: Entropy-Calibrated Watermarking for Language Model Text.” OpenAI technical report (2026)
[2] Diaa et al., “Optimizing Adaptive Attacks against Watermarks for Language Models.” Proceedings of the 42nd International Conference on Machine Learning 267 (2025)
[3] Wang, Wenhui, et al. “MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformers.” Advances in Neural Information Processing Systems 33 (2020)

The additional efficiency, semantic diversity, and attack evaluations have addressed my main concerns. I already supported acceptance in my original review and continue to do so.

Reviewer_We6eOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a categorization of inference-based watermarking for auto regressive LLM. The categorization is based on casting the watermarking problem as maximizing a (additive) score given a constraint on the watermarked token distribution. The authors distinguish between hard constraint -- constraining the perturbation directly on a given watermarked distribution -- and a soft constraint -- constraining the expected watermarked distribution over the secret keys -- leading to the standard distortion-free formulation. The authors demonstrate that this simple framework captures most of current schemes. Interestingly, they show that most prominent schemes all optimize the score under a f-divergence constraint (they don't formulate it this way, but this is what it amounts to!), which makes them all equivalent in a certain sense. Another important finding, which has been conjectured privately in the watermarking community for some time now but never proven to the best of my knowledge, is that an optimal distortion-free scheme can always be written deterministically. The authors then propose to apply their framework to produce a new "optimal" watermarking scheme under a perplexity constraint (if the authors allow me some playfulness for a second, a better title might have been a pastiche of the classic Valiant and Valiant statistics paper [1], "An Automatic Optimal Watermarking Generator") . All these results are validated experimentally.

[1] Valiant, Gregory, and Paul Valiant. 2014. “An Automatic Inequality Prover and Instance Optimal Identity Testing.” 2014 IEEE 55th Annual Symposium on Foundations of Computer Science, October, 51–60. https://doi.org/10.1109/FOCS.2014.14.

Critical strengths

The current situation of scientific practice combined with the rising interest in LLM watermarking due to recent regulations have led to a field which looks a bit like the wild west in terms of methodology and in the amount of new methods proposed each year.

Consequently, I wholeheartedly welcome papers that propose to give some order to the modern watermarking discipline. This paper does it in a simple manner, with a clear methodology and an application demonstrating the relevance of the approach: one cannot really ask for more.

Overall, the paper is well-written, with the formalism, terminology and notation introduced in a timely manner and comprehensively.

The results themselves, though I believe the authors could have gone a bit further in the abstraction, are relevant theoretically: 1) Aaronson, KGW, SynthID all maximize a score constrained on a f-divergence and 2) all optimal distortion-free schemes can be written as a deterministic algorithm.

Critical weaknesses

I will state here some weaknesses, though they are not critical in the sense that don't point to a rejection of the paper.

Bad choice for the experimental protocol: It is beyond me why the authors chose to report TPR@FPR=1% given the fact that a methodology to compute a statistically meaningful p-value for all the studied algorithm has been shown (except for KTH): see [2] for AAR and KGW, and the Synth-ID supplementary on the Frequentist detector. Using a p-value for the score would allow to work at extremely low FPR, even when computing only a few hundred samples, which would make Figure 2 infinitely more convincing. As is, the figure barely shows the difference between the schemes. Given the extremely high variance of Figure 4, I also highly doubt that the results are statistically significant.

Focus on general score: Maximizing a score does not necessarily lead to an optimal scheme in an hypothesis testing sense. The more important question is: what score is best to maximize in order to maximize detectability for a given token distribution? The simplicity of the current work relies on general scores, but I would appreciate it if the authors could relate their approach to recent theoretical works that demonstrate the shape of a truly optimal token-level watermarking scheme for auto-regressive LLM such as [3]. In particular I would be interested in the "detectability" gap between such an optimal scheme and a scheme built using the author's method and an arbitrary f-divergence constraint.

[2] Fernandez, Pierre, Antoine Chaffin, Karim Tit, Vivien Chappelier, and Teddy Furon. 2023. Three Bricks to Consolidate Watermarks for Large Language Models. Version 1. https://doi.org/10.48550/ARXIV.2308.00113.

[3]Li, Xiang, et al. "A statistical framework of watermarks for large language models: Pivot, detection efficiency and optimal rules." Annals of statistics 53.1 (2025): 322.

Questions for authors

The strengths of the paper far outweigh its weaknesses. Furthermore, I am concerned that such papers that favors theoretical studies and explanations over the creation of novel schemes will get less traction from other reviewers, whereas I believe this is the direction the discipline should strive towards.

Consequently, I believe this paper should be accepted given the strengths I highlighted. This judgement is conditional on thing: even though I read over the proof, I could not check them carefully in the allotted time for the reviews. I will decrease my score if another reviewer demonstrates a crucial flaw, notably for Theorem F.4 and F.5.

Some questions and demands that can still improve my score:

  • Please, try to perform the experiments for a relevant FPR, say 10−610^{-6} , instead of the useless 11%, possibly using p-values instead of scores.
  • Please, answer my "Focus on general score" question in the Weakness section.
  • One popular watermarking scheme, Watermax [4], demonstrates that there is not a definite detectability-quality trade-off as long as one does not work at the token-level but at the n-gram level (the real cost being computational complexity). The authors don't discuss this type of approach despite talking about SynthID and other rejection sampling schemes in the appendix. I believe one could nevertheless generalize the author's approach to the n-gram level quite simply in their framework, am I wrong? In this case, what is the system solved by Watermax (which is originally cast as the aggregation of many rank tests) in your framework, and how can it be that, for example, KGW is optimal if we admit Watermax can always do better given enough compute?

[4] Giboulot, Eva, and Teddy Furon. "WaterMax: breaking the LLM watermark detectability-robustness-quality trade-off." Advances in neural information processing systems 37 (2024): 18848-18881.

2 comments

We thank the reviewer for their positive review. We address their remaining concerns below, and have updated our submission. All changes in the submission are colored in blue for clarity.

Q1: Could you report results with TPR@$10^{-6}$?

We thank the reviewer for their suggestion and have updated Figure 2 with TPR@$10^{-6}$ (as well as Table 8, previously Table 6). Our conclusion remains the same: a scheme optimizing a given constraint is indeed optimal with respect to that constraint. We also provide full ROC curves in Appendix E.6 for all evaluated schemes.

Q2: Are the results in Figure 4 statistically significant?

Good point. Based on the reviewer's comment, we have updated Figure 4 with confidence intervals (instead of standard deviations) computed with bootstrapping ($n=20,000$ resamples of prompts with replacement), and we confirm that the main result is significant: the 95% CIs on Self-BLEU of the soft-constrained schemes (AAR/KTH and PPL (Soft)) do not overlap with those of any hard-constrained scheme.

Q3: Are the schemes optimal with respect to power? Which score is best to maximize detectability?

We agree with the reviewer that our optimality results are to be taken with respect to sum-based detectors. In general, for a given distribution, an optimal detector can’t be derived without assumptions on the LLM next-token distribution (e.g. in [1] they assume that the next-token distribution satisfies a $\Delta$-regularity assumption). This is why we opted for simplicity and restricted ourselves to sum-based detectors. At the same time, a sum-based detector formulation already recovers all existing watermarking schemes used in practice (e.g. KGW, SynthID, AAR) and even captures newer schemes designed after our work such as OpenAI’s TextGrain [2] (now in Section 4.1)!

Nonetheless, in Appendix F.2, we now apply [1]’s framework on all evaluated score distributions (approximating the watermarked score distribution with a histogram fitted on a separate training split of watermarked replies) and compare the TPR at 1% and $10^{-6}$ FPR between the sum-based and fitted likelihood-ratio (LLR) detectors. We find that, for all schemes but AAR, there is no significant gap between the two detectors. For AAR, replacing our KS detector by the sum-based Gamma detector (sum of exponential scores) from [3, 4] closes the gap between the LLR detector and a sum-based one.

Q4: Can the framework generalize to n-gram scores?

The reviewer intuition is correct and our framework generalizes to the n-gram scores setting, we now derive it in Appendix G.1. We replace the vocabulary $\Sigma$ by blocks $\mathcal{B}=\Sigma^n$, with block probability $p_b=\prod_{j=1}^n p(b_j\mid\text{prefix},b_{<j})$ and block score gb=∑j=1ngj(bj)g_b=\sum_{j=1}^n g_j(b_j) (or any fixed block-scoring rule). Equation 2 then applies unchanged, with q(g)∈Δ(B)q(g)\in\Delta(\mathcal B).

Q5: Can WaterMax be expressed in this generalized framework?

Yes, we now derive it in Appendix G.2. Greedy WaterMax samples KK i.i.d. blocks B1,…,BK∼pB_1,\ldots,B_K\sim p and outputs the highest-scoring one. A block in A⊆BA\subseteq\mathcal B can only be selected if at least one draft lands in $A$, giving the hard constraint
∀g,∀A⊆B,q(g)(A)≤1−(1−p(A))K.\forall g, \forall A\subseteq\mathcal B,\quad q(g)(A)\le 1-(1-p(A))^K. Selecting the highest-scoring draft attains this bound for every upper-score set, and hence maximizes the expected score. WaterMax is thus the solution of our framework under this constraint with $\varepsilon=0$.

Q6: Does this contradict KGW’s optimality?

No it does not. KGW remains optimal for token-level scores with a hard-KL penalty, whereas WaterMax is optimal for n-gram scores with its specific budget constraint.

Q7: Can the authors double-check the proofs, especially Theorems F.4 and F.5?

We thank the reviewer for highlighting the relevance of both Theorems. We now provide Lean proofs for all theorems (auto-formalized by coding agents) in our work, including Theorems H.4 and H.5 (previously F.4 and F.5, as the proofs moved to Appendix H). They are in the lean/ folder of the updated supplementary material.

We note that for F.5 we made a small technical assumption that held for all studied constraints (see new Corollary F.5.1) but that we did not explicitly state (that the problem admits an optimal solution). We have clarified this in the latest revision and thank the reviewer for pointing us in this direction.

[1] Li et al. “A statistical framework of watermarks for large language models: Pivot, detection efficiency and optimal rules.” Annals of Statistics (2025).
[2] Li et al. “textGrain: Entropy-Calibrated Watermarking for Language Model Text.” OpenAI technical report (2026).
[3] Fernandez et al. “Three Bricks to Consolidate Watermarks for Large Language Models.” WIFS (2023).
[4] Sander et al. “TextSeal: A Localized LLM Watermark for Provenance & Distillation Protection.” NeurIPS (2026).

I thank the authors for the additional study on n-gram scores and the better experimental visualization. I am somewhat disappointed that they changed the FPR for a single Figure (and the tables) given that it allows such a better visualization of the result. It won't play a role in my score, but I truly enjoin the authors to harmonize the figures to a FPR=1e-6.

Since I am not fluent in Lean (and I believe few people truly are), and that the formalization was performed by coding agent, I cannot trust a proof simply on the ground provided by the authors. I am not fluent, but I know enough Lean to understand that it is easy to fake a proof, or generate a incorrect formalization that still passes.

Given that the proofs have substantially changed (notably H.5, previously F.5 which is exactly the one I am interested in), and given the small timeframe alloted to reviewers, I will allow myself to take more time before increasing my score since I need to re-check them seriously.

Comments on the paper

1 comment

Dear Reviewers,

As the discussion deadline (Nov 18, AoE) approaches, please review the authors' rebuttal if you have not done so, and post a follow-up indicating whether your main concerns have been addressed or whether any issues remain.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.