acceptodds
Under review as a conference paper at ICLR 2027

WOSpec: Watermarked Online Speculative Decoding

Abstract

Large language models (LLMs) increasingly rely on speculative decoding for efficient inference and on watermarking for output provenance. Online speculative decoding improves efficiency by training the draft model on the queries it serves, but when the watermark is embedded through the draft, every draft update can change the watermark carried by the generated text. Therefore, we show that to avoid this problem, drafter-invariant speculative decoding is needed, and in such an online setting the generated text is the watermarked text for every draft, including drafts trained on earlier watermarked outputs. In contrast, the watermark strength of watermarked speculative sampling with the standard acceptance rate depends on the draft, which was unstable. We propose WOSpec, a watermarked online speculative decoding framework built on watermarked Gumbel-max list sampling, which naturally has a drafter-invariance property, achieves desriable sampling efficiency, and maintains stable watermark detectability during online training. On Llama-2-7B-chat and Vicuna-13B, the watermark strength of our WOSpec remains unchanged under online training and matches that of a recently proposed watermark-maintaining sampling scheme, while WOSpec accepts 25–28% more tokens per verification step in terms of sampling efficiency.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Speculative decoding, online learning, LLM watermarking, channel simulation, relative entropy coding, information theory.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_C7soOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes WOSpec, a method that integrates Gumbel-max watermarking with online speculative decoding. The key contribution is Watermarked Gumbel-max List Sampling (WMGLS), which uses exponential order statistics to construct a multi-draft coupling that preserves the exact watermarked output regardless of draft model updates. By combining this coupling with existing online draft training, the method improves decoding efficiency while maintaining stable watermark strength. Experiments on Llama-2 and Vicuna models demonstrate higehr sampling efficiency and throughput.

Critical strengths

The reverse construction of draft sampling randomness is a neat idea. By leveraging the properties of exponential order statistics, the authors design a coupling that preserves the exact Gumbel-max watermarked output while retaining the sampling efficiency of multi-draft speculative decoding.

Critical weaknesses
  1. Unclear novelty and contribution. The main technical contribution is not entirely clear to me. While the reverse construction is a neat idea, the overall approach seems quite similar to existing work, especially Watermarkable Multi-Draft Speculative Sampling via Poisson Processes (arXiv:2609.21858), which already studies multi-draft coupling with watermark preservation. Also, the online draft setting feels more like a natural extension rather than a core technical contribution, since the proposed coupling already works for arbitrary draft models. Without a direct comparison to MPFR, it is hard to see how much additional value this method brings.

  2. Limited generality and practical applicability. Although the authors present WOSpec as a framework, the proposed method is specifically designed around Gumbel-max watermarking. There are many other watermarking methods, such as Google's SynthID-Text, and the proposed approach is hard (or impossible) to extend to them. This makes the method more like a specific solution for Gumbel-max watermarking rather than a general framework, which limits its broader practical applicability.

Questions for authors

My biggest concern is the novelty and contribution of this work. As I understand it, the core idea seems quite similar to existing work such as MPFR, while the online draft setting feels more like a natural extension rather than a new technical challenge. Am I missing any key technical differences or contributions? I would appreciate it if the authors could clarify what makes this work fundamentally different from prior methods.

3 comments
AuthorsOpenReview ↗

Novelty and Comparison to MPFR

We thank the reviewer for the review and for recognizing the reverse construction as a neat idea.

Before addressing the questions, we would like to respectfully clarify that MPFR (arXiv:2609.21858) is contemporaneous work, first made available on arXiv on September 18, 2026, around the submission of WOSpec. According to the ICLR 2027 Reviewer Guidelines (FAQ on contemporaneous work), authors are not required to compare against contemporaneous or arXiv-only papers, and the absence of such comparisons should not itself be grounds for rejection.

Nevertheless, we appreciate the reviewer's suggestion and have now implemented MPFR in our online framework and conducted a direct comparison, showing comparable sampling efficiency and watermark preservation.

More importantly, we would like to emphasize that WOSpec and MPFR make distinct contributions. While MPFR develops a multi-draft coupling based on Poisson processes, WOSpec studies watermark preservation under online draft updates and introduces WMGLS through a reverse exponential order-statistics construction. Furthermore, our online serving-and-learning framework is compatible with different watermark-preserving coupling schemes, including both WMGLS and MPFR, as demonstrated by our new experiments. Thus, the online framework and its analysis provide contributions beyond any particular list-coupling algorithm, and compared to MPFR, our contribution is orthogonal.

We believe that these theoretical and empirical contributions provide independent value beyond the contemporaneous MPFR work. We will further clarify their relationship and the specific novelty of WOSpec in the revised manuscript.

We now respond point by point as follows.

Weakness 1: Novelty and Contribution

As we stated above, although MPFR is contemporaneous work and orthogonal to our contribution, we nevertheless incorporated it into our online learning framework and compared its performance with our original scheme.

We integrated MPFR into our online learning framework of WOSpec by replacing only the multi-draft coupling module. The implementation follows Algorithm 4 of the MPFR paper and matches it on 3000 random cases. A ll other settings follow the WMGLS experiment (same initial draft, training recipe, prompts and seeds):

Llama-2-7B-chat, GSM8K, K=6K=6, L=5L=5, 3 seeds, 100 chunksAATPSAATPS per seedANLPPTAATPS, chunks 1–20 →\rightarrow 81–100
MPFR1.5981.597 / 1.595 / 1.6030.00821.475 →\rightarrow 1.645
WMGLS1.5951.593 / 1.593 / 1.5970.00821.471 →\rightarrow 1.641
GLS (unwatermarked)1.5941.594 / 1.595 / 1.594--1.472 →\rightarrow 1.642
PFR1.2651.264 / 1.264 / 1.2680.00821.190 →\rightarrow 1.294

Under matched conventions, MPFR commits the same keyed token as WMGLS, and its ANLPPT equals that of WMGLS in our runs (0.0082). Its AATPS is higher by 0.002-0.006 in each seed. We therefore do not claim that WMGLS is a better coupling than MPFR. Instead, our experiments demonstrate that the online framework of WOSpec is compatible with both coupling schemes, without changing the online training procedure or watermark detector. This makes our contribution (a unified online, watermarked speculative decoding pipeline) even stronger.

Are key technical differences missing?

The committed token is the same, so the technical difference lies in how the drafts are generated. MPFR takes the first arrivals of the same keyed clocks under qq. WMGLS splits the watermark minimum among the active drafts, so that each draft keeps the proposal law of conditional GLS (Proposition 2(i)). The acceptance of WMGLS can therefore be compared directly with unwatermarked GLS on the same draft, whereas MPFR has its own proposal law, which has no unwatermarked counterpart.

What is fundamentally different from prior methods, and is the online setting only a natural extension?

The reviewer is right that a drafter-invariant coupling works for any draft, and this is the property the online setting needs. What the online setting adds is the question of which methods have it, and what the others lose. Hu and Huang (2024) show that, for a fixed single draft, a watermarked speculative method keeps either the watermark strength (PFR) or the acceptance rate (MSE), not both. Under online training the draft changes after every chunk, so the trade-off gains a dimension: how the strength and the efficiency of a method change as its draft is updated. Our contributions follow this dimension.

AuthorsOpenReview ↗

Novelty and Comparison to MPFR (continued)

  1. The online dimension and its criterion. We formalize the requirement (Definition 1: the committed text does not depend on the draft) and show that MSE fails it. For a draft independent of the decoder randomness, its watermark strength depends on the draft: the loss in expected token score relative to PFR is at most 34DTV(p,q)\frac{3}{4}D_{\mathrm{TV}}(p,q), and the constant is tight (Proposition 1). Under online training, this independence assumption need not hold, and MSE's watermark strength can change as the draft is updated, as observed in Figure 3(c), while the committed watermark of PFR and WMGLS remains unchanged under matched target, key, and fallback conventions. Our SynthID-Text experiments show a similar empirical distinction (W2).

  2. Strength and efficiency together, online. A single draft at the strength end (PFR) keeps the watermark but has a lower acceptance rate. WMGLS stays at the strength end and recovers efficiency with several drafts and online training: its AATPS is 0.26–0.38 higher than that of MSE, with the ANLPPT of PFR (Table 1), and within 0.01 of unwatermarked GLS on Llama-2-7B-chat and Vicuna-13B. Under draft–randomness independence, WMGLS also reproduces the conditional GLS proposal law (Proposition 2(i)); its comparable efficiency under online training is supported empirically.

  3. What carries over to a draft trained on its own watermarked outputs. The pathwise identity (Proposition 2(ii)) holds for any update rule under matched target, key, and fallback conventions. The acceptance analysis of conditional GLS (Proposition 2(i)) assumes a draft independent of the decoder randomness, which training on same-key responses can violate, so we measure online acceptance directly. On 100 held-out prompts (3 seeds, six target–workload settings), the final online draft decoded by WMGLS with fresh keys is within 0.032 AATPS of the same draft decoded by GLS. With the deployment key, WMGLS is higher by 0.105 and 0.140 on Vicuna-7B GSM8K and Code, respectively, and differs by at most 0.030 in the other settings; with the template contexts masked, the difference on Vicuna-7B GSM8K and Code is within 0.003.

  4. A system into which keyed couplings plug. The WOSpec online serving-and-learning framework can accommodate different compatible watermark-preserving multi-draft couplings. In particular, MPFR can be integrated by replacing a single coupling function and, under matched conventions, preserves the same pathwise watermarked output while using the same online training loop and detector.

AuthorsOpenReview ↗

Generality beyond Gumbel-max, and Summary

Weakness 2: Generality and Practical Applicability

We agree that WMGLS is specific to watermarks built on an exponential (Gumbel-max) race, which is the watermark of Aaronson and Kirchner (2023) used by the work on watermarked speculative sampling that we build on (Hu and Huang, 2024; He et al., 2026). Two parts of the work are not tied to it: the online loop and system, and the online dimension of the strength--efficiency trade-off. To check both, we integrated the two speculative constructions for SynthID-Text (Dathathri et al., 2024), a high-detectability one and a fast one, into our system; only the draft rule and the detector change. Both use one draft, our online loop, the same EAGLE-3 head and training as the paper, and the opening masking of Appendix D. The detection rate uses a threshold calibrated to a 1% false-positive rate on unwatermarked GLS responses:

Vicuna-7B, GSM8K, over 2 seeds, 100 chunksAATPS, chunks 1–20 → 81–100Detection at 128 tokensMean gg-value (change in standard errors)
SynthID-High1.162 → 1.24160.2% → 58.6%0.5252 → 0.5247 (−0.8-0.8)
SynthID-Fast1.172 → 1.25537.1% → 31.8%0.5186 → 0.5166 (−3.5-3.5)

Both constructions run in our system and become faster with online training. SynthID-High, whose committed token has a law that does not depend on the draft, keeps its mean gg-value along the stream within noise, as PFR does for the Gumbel-max watermark. SynthID-Fast commits draft-dependent tokens, and its target-key mean gg-value and detection rate change as the draft is trained, as MSE does. The online dimension of the trade-off is therefore not specific to Gumbel-max.

What is specific is how WMGLS resolves it. SynthID-High, like PFR, keeps the watermark with a single draft, and its AATPS over the stream (1.215) is close to that of PFR (1.220) and well below that of WMGLS at K=6K=6 (1.546; same streams, seeds and keys, opening-masked). WMGLS recovers the efficiency with several drafts because the exponential race allows a multi-draft coupling whose committed token is the target's keyed race. A multi-draft coupling for the SynthID tournament is an interesting future direction.

Summary

We sincerely thank the reviewer again for the constructive comments and valuable suggestions, which have helped us strengthen both the contributions and empirical evidence of our work.

In response, we have addressed the two main concerns:

  • Novelty, contributions, and MPFR: We have clarified the distinct contributions of WOSpec relative to the contemporaneous MPFR work. Furthermore, following the reviewer's suggestion, we have implemented MPFR within our online serving-and-learning framework and conducted direct comparisons, demonstrating that our framework is more general and can accommodate different watermark-preserving coupling schemes.

  • Generality, practical applicability, and SynthID: We have extended our investigation to SynthID-Text, demonstrating the applicability of our online framework beyond Gumbel-max watermarking and further examining the trade-off between watermark strength and speculative efficiency.

We hope that our clarifications, additional analyses, and new experiments adequately address the reviewer's concerns. We would be very grateful if the reviewer could kindly reconsider their evaluation in light of these new results.

Thank you again for your time and consideration!

Reviewer_rH3COpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

WOSpec studies watermarking when speculative drafts are updated online. Its WMGLS sampler couples multiple draft sequences to a fixed Gumbel-max watermark and uses OnlineSPEC's training procedure to update the draft. The paper also derives MSE's expected token score and a tight bound on its loss relative to the target watermark. Experiments on three workloads and 7B/13B targets compare watermark strength and inference efficiency, with draft-count and frozen-draft ablations.

Critical strengths
  1. The reverse splitting construction makes the watermark compatible with conditional GLS. It starts with the target's watermark random numbers and constructs draft random numbers whose minimum selects the same target token, regardless of the active drafts. Under ideal randomness and draft independence, it also preserves conditional GLS's distribution of tokens committed per step. With the target probabilities, key, and fallback randomness fixed, the output identity does not require draft independence. It therefore continues to hold when the draft learns from earlier watermarked responses. This gives a concrete way to adapt the draft without changing the ideal watermarked text. The paper correctly distinguishes this guarantee from the GLS acceptance result, which does not automatically carry over to online training.
  2. The experiments support the intended combination of watermark preservation and efficient verification. In Table 1, WMGLS matches PFR's reported ANLPPT while committing 2525–28%28\% more tokens per verification step with six drafts rather than one. More directly, with the same six-draft budget, WMGLS closely matches unwatermarked GLS in tokens per step and is within 4%4\% in TPS. The frozen-draft ablation shows that online updates improve inference efficiency without a measured loss in watermark strength. These comparisons demonstrate the value of the construction beyond the formal identity.
Critical weaknesses
  1. Total online savings remain unmeasured. TPS explicitly excludes training, although each 40-prompt chunk is followed by target evaluation and two epochs for each of three learners. The frozen-draft comparison therefore establishes faster inference, but not whether the saved time repays these updates. Reporting cumulative serving and update times would clarify the practical benefit. Relative efficiency against the concurrent watermarked multi-draft method MPFR is also untested. These omissions limit the assessment of the full system, but do not contradict the reported inference results or the comparison with GLS at the same draft count.

  2. The optional masking experiment has a generation/detection mismatch. With mask_opening enabled, state.py remembers the contexts of the first three output positions, while detect.py omits them from its seen set. A later recurrence can therefore use fallback randomness during generation but receive a watermark score during detection. For prompt tokens (1,2,3)(1,2,3) and output (1,2,3,4,5,6)(1,2,3,4,5,6), this occurs at output position 4. Table 6 should be checked with consistent rules. The effect on its metrics is unknown, and this configuration is not used in the main experiments.

  3. Table 3 incorrectly applies Definition 1 to GLS that takes the minimum over all KK drafts. For p=(1/2,1/2)p=(1/2,1/2), S(1)=(1,2)S^{(1)}=(1,2) and S(2)=(3,0.1)S^{(2)}=(3,0.1), the rule selects token 1 with K=1K=1 and token 2 with K=2K=2. Definition 1 requires invariance even when KK changes. Qualifying this entry as holding for fixed KK resolves the comparison error without affecting WMGLS's own result.

Questions for authors
  1. Including target evaluation and draft updates, how does total time compare with the frozen-draft stream? Do cumulative inference savings exceed the update cost within the evaluated stream?
  2. How often does the masking mismatch affect Table 6's evaluated prefixes, and do its conclusions hold after generation and detection use the same rule?
4 comments
AuthorsOpenReview ↗

Total time including target evaluation and updates, and MPFR

Thank you very much for the positive assessment and for the precise checks of the code and of Table 3. We are glad that the reviewer finds that the output identity holds without draft independence, and that online updates improve inference efficiency without a measured loss in watermark strength.

Before addressing weaknesses, we would like to respectfully clarify that MPFR (arXiv:2609.21858) is contemporaneous work (available on arXiv on Sep 18, 2026, around ICLR submission) and according to the ICLR 2027 Reviewer Guidelines, authors are not required to compare against it. Importantly, we would like to note that WOSpec and MPFR make distinct contributions. While MPFR develops a multi-draft coupling, WOSpec studies watermark preservation under online draft updates and introduces WMGLS through a reverse exponential order-statistics construction. Our online framework is compatible with different watermark-preserving coupling schemes, includingboth WMGLS and MPFR. Thus, our contribution is orthogonal to MPFR. Nevertheless, as asked by you and also other two reviewers, we now have implemented MPFR within our WOSpec online framework and rigorously conducted a comparison.

We now respond point by point as follows.

Weakness 1 and Question 1: Total time including target evaluation and updates, and comparison to MPFR

We agree that the cost of the updates is needed to assess the full system, and we now report it. We follow the online setting of OnlineSPEC (Qian et al., 2026), in which the update runs asynchronously on separate training GPUs and the updated draft is synchronized to the inference server, and we measured the end-to-end wall-clock time of the stream, which includes the target evaluation and the updates of the three learners.

Q1(a): total time against the frozen-draft stream.

The setting is Vicuna-7B/GSM8K with WMGLS, over 2 seeds, with chunks of 40 prompts (88--96 of the 100 chunks per seed). The frozen draft is served in the same way (one persistent server):

Vicuna-7B, GSM8K, WMGLS (over 2 seeds)Wall-clock time per chunkvs frozenAATPSANLPPT
Frozen draft176.3 s--1.270.039
Online, asynchronous update150.9 s-14%1.530.039

With online updates, the wall-clock time per chunk is 14% shorter than with the frozen draft. AATPS rises from 1.27 to 1.53, and ANLPPT is the same (0.039), consistent with Proposition 2(ii).

Q1(b): do the cumulative inference savings exceed the update cost?

Yes. The updates run on the training GPU while the server keeps serving, so the time saved by the better draft is not offset by the updates: the stream is 14% shorter than the frozen-draft stream. We will add these times to Section 5.

MPFR

We now have integrated MPFR into WOSpec and ran it with the same initial draft, training recipe, prompts and seeds as WMGLS:

Llama-2-7B-chat, GSM8K, K=6K=6, L=5L=5, 3 seeds, 100 chunksAATPSAATPS per seedANLPPTAATPS, chunks 1–20 → 81–100
MPFR1.5981.597 / 1.595 / 1.6030.00821.475 → 1.645
WMGLS1.5951.593 / 1.593 / 1.5970.00821.471 → 1.641
GLS (unwatermarked)1.5941.594 / 1.595 / 1.594--1.472 → 1.642
PFR1.2651.264 / 1.264 / 1.2680.00821.190 → 1.294

Under matched conventions, MPFR commits the same keyed token as WMGLS, and its ANLPPT is equal to that of WMGLS in our runs (0.0082). Its AATPS is higher by 0.002--0.006 in each seed (0.2%), and it rises over the stream like that of WMGLS. We regard the two as comparable in efficiency. We report AATPS because our MPFR port is not kernel-optimized. We will add this comparison to Section 5 and discuss MPFR as concurrent work in Section 2.

AuthorsOpenReview ↗

Opening-mask generation/detection mismatch, Fixed-K qualification in Table 3, and Summary

Weakness 2 and Question 2: Opening-mask generation/detection mismatch

The reviewer is right, and we thank them for the precise example. We reproduced it: with prompt (1,2,3)(1,2,3) and output (1,2,3,4,5,6)(1,2,3,4,5,6), output position 4 is generated with fallback randomness but scored by the detector, because the detector did not add the masked opening contexts to its seen set. We fixed the detector so that every visited output context, including the masked opening contexts, enters the seen set, and masked and repeated contexts are not scored, as in generation. We also added a regression test, with the reviewer's example and random texts, that checks that the detector skips exactly the positions that generation treats as repeated.

Question 2: Table 6 with the same rule in generation and detection.

We recomputed Table 6 with the corrected detector (Vicuna-7B, key of seed 0, GSM8K and Code, chunks 1--20; 800 masked responses per workload). Detection depends only on the text and the key, and the fix does not change generation, so we rescored the saved masked responses. The table shows the masked rows as in Table 6 and with the corrected detector; Length, AATPS and TPS do not depend on the detector, and the rows without masking are unaffected:

WorkloadMethodLengthAATPSTPSANLPPTDet. (%)
GSM8KGLS168.81.45044.18----
WMGLS111.21.61646.770.03730.8
WMGLS, masked (Table 6)173.91.44542.660.03731.8
WMGLS, masked (corrected detector)173.91.44542.660.03731.8
CodeGLS396.91.35441.46----
WMGLS285.91.55245.200.04645.2
WMGLS, masked (Table 6)407.51.36340.670.05548.5
WMGLS, masked (corrected detector)407.51.36340.670.05548.5

The corrected values equal the stored values of these runs (ANLPPT 0.0366 and 0.0547 at four decimals, detection 31.8% and 48.5%), and the score of no response changes. The two rules agree at all 463562 output positions of the masked responses, so the mismatch does not affect any evaluated prefix. The conclusions of Table 6 hold: with masking, the response length and the AATPS of WMGLS are close to those of GLS, and ANLPPT and the detection rate do not decrease.

We also checked the 36 new 100-chunk masked online runs of this rebuttal (PFR, MSE and WMGLS on Vicuna-7B GSM8K and Code, Llama-2-7B-chat GSM8K and TinyLlama-1.1B GSM8K; 3 seeds each; 144000 responses). There, the rule of the paper disagrees with generation at 629 of 33.1M output positions (0.002%), which can only happen when a masked opening context recurs later in the response. With the consistent rule there is no disagreement, no run's detection rate changes, and the ANLPPT@128 of a run changes by at most about 0.0001. The main experiments do not use opening masking and are not affected.

Weakness 3: Table 3: GLS with the minimum over all K drafts.

The reviewer is right, and we thank them for the counterexample. We verified it: for p=(1/2,1/2)p=(1/2,1/2), S(1)=(1,2)S^{(1)}=(1,2) and S(2)=(3,0.1)S^{(2)}=(3,0.1), the rule selects token 1 with K=1K=1 (scores 2 and 4) and token 2 with K=2K=2 (scores 2 and 0.2). Definition 1 requires the committed token not to change with KK, so this rule satisfies it only for a fixed KK. We will correct the GLS entry of Table 3 to "for fixed KK". WMGLS is not affected: by Proposition 2(ii), its committed token at a new context is arg⁡min⁡vSc,vκ/p(v)\arg\min_v S^\kappa_{c,v}/p(v), which does not depend on KK, LL or the draft.

Summary and Appreciation

We sincerely thank the reviewer again for the positive assessment of our work, as well as the careful technical checks and constructive suggestions, which have helped us further strengthen the paper.

In response, we have addressed all three concerns:

  1. Online efficiency and MPFR comparison: We evaluated end-to-end wall-clock performance, demonstrating 14% lower elapsed time with asynchronous online training compared with a frozen draft, and conducted a direct comparison with the contemporaneous MPFR method.
  2. Implementation correctness: We reproduced and corrected the generation–detection masking inconsistency identified by the reviewer, added regression tests, and verified that the original Table 6 results remain unchanged.
  3. Theoretical clarification: We acknowledged the reviewer's counterexample and clarified the fixed-KK qualification for GLS in Table 3.

We sincerely hope that these additional experiments, corrections, and clarifications satisfactorily address the remaining concerns and further strengthen the reviewer's confidence in our work. We would be very grateful if the reviewer could kindly consider increasing their rating in light of the new evidence, which would greatly strengthen our paper's prospects for acceptance.

Thank you again for your encouraging and positive evaluation, valuable feedback, and consideration!

Thank you for the point-by-point response and revisions covering timing and MPFR (Sec. 5.4), masking (App. E), and the fixed-K qualification (Table 4). These clarifications and revisions have addressed my concerns, and I maintain my original score.

Thank you again for your constructive feedback and continued support! Your valuable suggestions have helped us substantially improve the paper.

Reviewer_93TtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies speculative decoding with a draft model that is continually updated from its own watermarked outputs. It argues that conventional watermark-preserving or efficiency-preserving speculative decoders can behave differently as the draft changes. The proposed WMGLS construction starts from the keyed Gumbel-max random variables that select the target-side watermarked token, then uses an inverse exponential order-statistics construction to generate draft-side random variables for multiple draft sequences. This makes the committed output equal to standard Gumbel-max watermarked generation Wκ(p)W^\kappa(p) for any draft, while recovering the conditional GLS proposal law when the draft is independent of the decoder randomness.

Critical strengths
  • The paper identifies a meaningful issue that is not captured by fixed-draft watermark/speculative-decoding analyses, where a draft trained on previous watermarked outputs can become dependent on the watermark key and its generated trajectories. The distinction between maintaining an acceptance rate and maintaining the actual committed watermarked text is insightful. Proposition 2(ii) provides a particularly strong statement: the output is pathwise identical to Wκ(p)W^\kappa(p), rather than merely matching its marginal distribution.

  • The inverse order-statistics construction in Eq. 4 is technically elegant. It fixes the minimum exponential race to the watermark randomness while preserving the required draft-side exponential variables under the stated independence assumptions.

Critical weaknesses
  • Proposition 2(i) requires the draft to be independent of the decoder randomness, but Section 4.3 explicitly acknowledges that this condition fails in the paper’s intended online loop because the draft is trained on earlier responses generated with the same key. The paper consequently proves that WMGLS preserves the committed text Wκ(p)W^\kappa(p) online, but does not prove that its proposal law, acceptance behavior, or AATPS remains that of conditional GLS online. The statement that “Part (i) does not carry over” is important, yet the abstract, contribution list, and experimental framing still closely associate GLS-level efficiency with online watermark-preserving decoding. Could the authors separate more clearly which claims are pathwise guarantees in the closed loop and which are empirical observations? In particular, Figure 3 and Table 1 show that WMGLS is close to GLS in AATPS, but do not establish that this remains true under arbitrary key-dependent online updates.

  • The closest concurrent baseline, MPFR, is not evaluated or formally compared to WMGLS. Table 3 states that MPFR already couples several drafts through keyed Poisson processes and has a stopping-time form of drafter invariance, while Section 5.1 omits it on the grounds that its AATPS is “close to that of GLS.” This does not establish whether WMGLS provides a stronger invariance property, better online behavior, higher acceptance, lower cost, or merely an alternative construction.

  • The main efficiency comparison in Table 1 uses K=6K=6 draft sequences for WMGLS and GLS but only one draft sequence for PFR and MSE. Thus, the reported 2525--28%28\% AATPS gain over PFR is partly the expected result of using several draft proposals rather than one. Table 11 confirms that AATPS rises substantially as KK increases, but it reports only one seed and one workload. Could the paper compare methods under a matched draft-compute or tree-size budget, report a broader KK sweep, and include MPFR under the same budget?

  • TPS excludes online training time by design (Section 5.1 and Table 5), despite the system updating three learners for two epochs after every chunk of 40 prompts. AATPS is a useful decoder-level metric, but it does not by itself establish an end-to-end advantage for online serving. What are the total throughput, update cost, and latency impact when training is included? This is particularly relevant because the proposed motivation is an online serving loop rather than a frozen-draft decoding benchmark.

  • Appendix D identifies a substantial key-reuse artifact: because the initial contexts after a common chat template are shared across requests, a fixed key can systematically alter response openings and lengths. For one Vicuna-7B key, watermarked responses on GSM8K and CodeSearchNet-Python are about one-third shorter than GLS responses, and Table 6 shows that masking the first three template-overlapping contexts materially changes both length and AATPS. However, this masking is not used in the principal experiments. Since Table 1 averages AATPS and TPS per response, can some apparent efficiency differences be driven by key-induced changes in response length rather than only by speculative matching? The main results should either adopt this control consistently or report a systematic key-sensitivity analysis.

  • The watermark evaluation is conditional on responses having at least 128128 tokens (Appendix E), but the paper does not report the retained fraction for each method, workload, and seed. Appendix D already shows that the key can affect response length. The low detection rates in Table 7 also limit the practical interpretation: on Llama-2-7B-chat, WMGLS detects only 2.1%2.1\%, 11.8%11.8\%, and 15.8%15.8\% of eligible responses across the three workloads. Matching PFR’s ANLPPT establishes preservation of the chosen Gumbel-max watermark, but not necessarily strong practical provenance detection.

  • The experiments use Llama-2-7B-chat, Vicuna-7B, and Vicuna-13B, one A100 GPU, and one request at a time. Will it be effective with smaller LLM like million scale?

Questions for authors

ee questions in weaknesses.

4 comments
AuthorsOpenReview ↗

Responses to pathwise guarantees and comparison with MPFR

We thank the reviewer for the detailed review and for recognizing the importance of studying watermark preservation under online draft updates. We appreciate the reviewer's positive assessment of our pathwise invariance guarantee and the inverse order-statistics construction.

Before addressing concerns, we would like to respectfully clarify that MPFR (arXiv:2609.21858) is contemporaneous work, first made available on arXiv on Sep 18, 2026, around ICLR submission. According to the ICLR 2027 Reviewer Guidelines, authors are not required to compare against it.

Nevertheless, following the reviewer's suggestion, we have implemented MPFR within our WOSpec online framework and conducted a comparison. Detailed are in our response to Weakness 2.

Weakness 1: Online guarantees versus measured acceptance

We agree, and we will separate the two in the paper. The guarantee is Proposition 2(ii): at every new context, the committed token is arg⁡min⁡vSc,vκ/p(v)\arg\min_v S^\kappa_{c,v}/p(v) whatever the drafts propose. Under any update rule, including training on earlier same-key responses, the committed text therefore equals Wκ(p)W_\kappa(p). That WMGLS keeps the AATPS of GLS online is an empirical finding, not a guarantee for arbitrary key-dependent updates. Proposition 2(i) assumes a draft independent of the decoder's random numbers, which training on same-key responses can violate, so we measure the online acceptance directly.

Measurement. We decode the same final online draft on 100 held-out prompts with the deployment key, with three fresh keys, and with unwatermarked GLS (K=6K=6, 3 seeds, AATPS as in Table 1):

Same draft, decoded byGLSWMGLS, fresh keysWMGLS, deployment keyWMGLS, deployment key, masked
Vicuna-7B, GSM8K1.5681.5641.6731.570
Vicuna-7B, Code1.4851.4531.6251.482
Vicuna-7B, Finance1.4861.4861.4961.482
Llama-2-7B-chat, GSM8K1.6331.6441.6391.635
Llama-2-7B-chat, Code1.5701.5851.6001.598
Llama-2-7B-chat, Finance1.5711.5751.5761.579

With fresh keys, which the draft has never seen, WMGLS is within 0.032 of GLS on the same draft. With the deployment key, it is within 0.030 in four settings and higher by 0.105 and 0.140 on Vicuna-7B GSM8K and Code. These are the two settings in which the key shortens responses through the shared openings after the chat template (W5); with the three template contexts masked, the deployment key is within 0.003 of GLS there as well. In no setting is the AATPS under the deployment key lower than that of GLS by more than 0.004.

Revisions. In the abstract, the contribution list, Section 4.3 and the captions of Figure 3 and Table 1, we will state the identity with Wκ(p)W_\kappa(p) as the guarantee and GLS-level online efficiency as an empirical finding supported by this table.

Weakness 2: Comparison to MPFR

As we stated above, although MPFR is contemporaneous work and orthogonal to our contribution, we nevertheless incorporated it into our online learning framework and compared its performance with our original scheme.

Specifically, we integrated MPFR by replacing only the multi-draft coupling module. The implementation follows Algorithm 4 of the MPFR paper and matches its proposal rule on 3,000 random cases. All other settings follow the WMGLS experiments (Llama-2-7B-chat/GSM8K, K=6K=6, L=5L=5, 100 chunks, 3 seeds, with the same initial draft, training recipe, prompts, and seeds).

MethodAATPSAATPS per seedANLPPTAATPS, chunks 1–20 → 81–100
MPFR1.5981.597 / 1.595 / 1.6030.00821.475 → 1.645
WMGLS (ours)1.5951.593 / 1.593 / 1.5970.00821.471 → 1.641
GLS (unwatermarked)1.5941.594 / 1.595 / 1.594--1.472 → 1.642
PFR1.2651.264 / 1.264 / 1.2680.00821.190 → 1.294

The results show that MPFR and WMGLS achieve comparable sampling efficiency and identical watermark strength. Under matched conventions, both commit the same target-keyed token, arg⁡min⁡vSc,v/p(v)\arg\min_v S_{c,v}/p(v), preserving the committed watermarked text under online draft updates. Their ANLPPT is 0.0082, equal to PFR. MPFR achieves 1.598 AATPS versus 1.595 for WMGLS, with an advantage of only 0.002–0.006 across individual seeds. Both improve similarly during online training.

Importantly, WOSpec and MPFR make distinct, orthogonal and complementary contributions. MPFR develops a multi-draft coupling, whereas WMGLS uses a reverse exponential order-statistics construction that reproduces the conditional GLS proposal law under draft–randomness independence (Proposition 2(i)). Moreover, our WOSpec online serving-and-learning framework is not restricted to a particular watermark-preserving coupling: both WMGLS and MPFR can be incorporated while retaining the same online training procedure and watermark detector. This demonstrates the modularity and broader applicability of our framework.

AuthorsOpenReview ↗

Matched budget, draft-count sensitivity, total online cost and service latency

Weakness 3: Matched budget, draft-count sensitivity

We agree that part of the gain over PFR and MSE comes from using several drafts. This is our purpose: single-draft PFR keeps the watermark unaffected by online updates but has a low acceptance, and WMGLS recovers efficiency with several drafts while keeping watermark. The fair test is against multi-draft methods with the same K,LK,L. Against GLS at the same K,LK,L, WMGLS is within 0.01 AATPS on Llama-2-7B-chat (Table 1; the larger gaps on Vicuna-7B GSM8K and Code are the key effect of W5, which disappears with masking). Against MPFR (W2), the AATPS is 1.595 versus 1.598.

Matched budget.

K⋅LK\cdot LMethodGSM8K AATPSGSM8K tok/sCode AATPSCode tok/sFinance AATPSFinance tok/s
5PFR, K=1K=1, L=5L=51.29142.01.25240.71.23641.2
10PFR, K=1K=1, L=10L=101.29133.01.25232.01.23632.3
WMGLS, K=2K=2, L=5L=51.41845.71.37845.11.35144.1
20PFR, K=1K=1, L=20L=201.29123.21.25222.41.23622.7
WMGLS, K=2K=2, L=10L=101.41934.71.37434.51.35434.6
WMGLS, K=4K=4, L=5L=51.55648.61.51748.01.49248.0
30PFR, K=1K=1, L=30L=301.29117.71.25117.21.23617.3
WMGLS, K=3K=3, L=10L=101.49736.61.45635.81.43135.7
WMGLS, K=6K=6, L=5L=51.63950.51.60050.11.57649.3

Extending the single draft from L=5L=5 to 3030 leaves AATPS unchanged, while the token rate falls to 42%. This is because verification stops at the first mismatch, and a single draft rarely survives beyond 5 tokens. At the same 30 tokens, 6 drafts of length 5 reach 27% higher AATPS at 2.9 times the token rate. Allocating the budget to more drafts, rather than to a longer chain, is what helps.

K sweep. We swept K=1,2,4,6,8K=1,2,4,6,8 with frozen final drafts on held-out prompts in 3 settings, each with 3 seeds (Vicuna-7B GSM8K and Code, Llama-2-7B-chat GSM8K). AATPS increases with K in each setting. E.g., on Llama-2-7B-chat/GSM8K it is 1.29, 1.42, 1.56, 1.64 and 1.70. On Llama-2-7B-chat Code and Finance, AATPS also increases from K=2K=2 to 4 to 6 (1.38, 1.52, 1.60 and 1.35, 1.49, 1.58; 3 seeds). The token rate saturates at K=6K=6 to 8. In addition to Table 11, we ran a 100-chunk online sweep on Llama-2-7B-chat/GSM8K:

KK1 (PFR)2468
AATPS, full stream1.2641.3841.5131.5921.650
AATPS, chunks 1–20 → 81–1001.19 → 1.291.29 → 1.421.40 → 1.561.47 → 1.641.52 → 1.70
ANLPPT0.00850.00850.00860.00850.0086

AATPS increases over the stream for every KK, and ANLPPT does not change with KK: additional drafts raise efficiency without affecting the watermark, as Proposition 2(ii) implies.

MPFR under the same budget. MPFR with the same K=6K=6 and L=5L=5 (W2) reaches online AATPS 1.598, against 1.595 for WMGLS and 1.265 for PFR, with equal ANLPPT.

Weakness 4: Total online cost, service latency

We agree that AATPS alone does not establish an end-to-end advantage, and we now report the cost including updates.

End-to-end time, compared across methods. The table gives the end-to-end time of each 100-chunk online stream, averaged over seeds; in these runs the updates run on the serving GPU between chunks. It runs from the first server start to the end of the last update and includes serving, target evaluation, the three learners, merging and server restarts.

End-to-end hours, 100 chunksPFRMSEGLSWMGLSWMGLS vs PFR / MSE
Llama-2-7B-chat, GSM8K9.009.308.038.07-10.3% / -13.2%
Llama-2-7B-chat, Code19.3719.2316.5317.00-12.3% / -11.6%
Llama-2-7B-chat, Finance21.6121.6818.4318.99-12.1% / -12.4%
Vicuna-13B, GSM8K11.6811.6311.0710.80-7.5% / -7.1%
Vicuna-7B, GSM8K / Code / Finance8.41 / 15.55 / 13.769.32 / 16.40 / 13.628.19 / 15.03 / 12.017.61 / 13.98 / 12.34-9.5% to -10.3% / -9.4% to -18.3%

Including updates, WMGLS takes 7-12% less end-to-end time than PFR and 7-1% less than MSE, and is close to unwatermarked GLS.

Online updates versus a frozen draft. We follow the online setting of OnlineSPEC (Qian et al., 2026), in which the update runs asynchronously on separate training GPUs and the updated draft is synchronized to the inference server, and we measure the end-to-end wall-clock time. On Vicuna-7B/GSM8K (over seeds), the time per 40-prompt chunk is 150.9 s, against 176.3 s with a frozen draft served the same way, 14% shorter. AATPS rises from 1.27 to 1.53, and ANLPPT is the same (0.039), consistent with Proposition 2(ii).

Latency. Training runs concurrently with serving, so updates introduce no training-related pauses in these runs. Online training raises the serving token rate from 39 to 47 token/s, which reduces the time per token by about 16%.

AuthorsOpenReview ↗

Key-dependent openings, lengths, efficiency, coverage and detectability

Weakness 5: Key-induced response length

On Vicuna-7B, a fixed key can shorten responses through the openings after the chat template (Appendix D), and the per-response average then gives the shorter responses more weight. This accounts for the higher AATPS of WMGLS than of GLS on Vicuna-7B GSM8K and Code in Table 1 (1.619 and 1.557 against 1.544 and 1.455), and we withdraw the statement that WMGLS is more efficient than GLS. The comparison with PFR is not affected: WMGLS and PFR use the same key and commit the same text (Proposition 2(ii)), and their mean response lengths differ by at most 2.5 tokens in all 21 runs of Table 1.

Following the suggestion, we first ran a key-sensitivity analysis and then reran the sensitive settings with the masking control.

Key-sensitivity analysis. We decode the final draft of each run with its own key and 3 fresh keys (3 seeds, 100 prompts). The mean response length varies strongly with the key on Vicuna-7B GSM8K (108-251 tokens; GLS 174) and Code (164-465; GLS 373), but little on Llama-2-7B-chat GSM8K (180-195; GLS 190). Masking the three template contexts reduces the Vicuna-7B ranges to 169-196 and 303-425.

Masking in the sensitive settings. On Vicuna-7B GSM8K and Code, we reran the full 100-chunk online streams of all keyed methods with masking, 3 seeds each, keeping the rest of our configuration:

MethodLengthRefusal openingAATPSANLPPT
Vicuna-7B, GSM8K: GLS178.21.2%1.544--
PFR, masked179.61.1%1.2200.0363
MSE, masked180.11.0%1.2310.0067
WMGLS, masked178.11.1%1.5470.0366
Vicuna-7B, Code: GLS415.43.7%1.455--
PFR, masked422.24.2%1.1770.0566
MSE, masked413.63.9%1.1860.0060
WMGLS, masked422.14.2%1.4670.0558

With masking, the response lengths and refusal openings of all keyed methods are close to those of GLS, and WMGLS is within 0.012 of GLS in AATPS (1.547 against 1.544 on GSM8K, 1.467 against 1.455 on Code). Its gain over PFR is 26.8% and 24.6% (28.5% and 27.5% in Table 1), and its ANLPPT and detection rate are close to those of PFR.

Weakness 6: Retained fraction and detection rates

We agree that matching PFR shows that WMGLS preserves the Gumbel-max watermark. Our purpose is to show that the watermark strength is stable under online updates and equal to that of PFR; how strong the watermark is depends on the watermark and target. We report the retained fractions, the detection rate at a calibrated 1% false-positive rate, and the reason for the low rates on Llama-2-7B-chat.

Retained fraction. Share of responses with at least 128 tokens (over 3 seeds):

Model / DatasetPFRMSEGLSWMGLS
Llama-2-7B-chat, GSM8K / Code / Finance77.9 / 98.4 / 99.8%78.9 / 98.3 / 99.8%78.2 / 98.4 / 99.8%77.9 / 98.3 / 99.8%
Vicuna-13B, GSM8K64.0%63.1%67.1%63.6%
Vicuna-7B, GSM8K / Code / Finance48.8 / 80.0 / 89.5%56.1 / 78.6 / 89.6%58.4 / 88.6 / 89.6%48.7 / 80.1 / 89.4%

WMGLS and PFR use the same key and commit the same text (Prop 2(ii)); their retained fractions agree: within 0.5 points in every setting mean and within 0.85 points in every of the 21 seeds. The detection comparison is thus made on the same cohort. The lower coverage on Vicuna-7B GSM8K and Code comes from the key-dependent short openings discussed in W5.

Detection at 1% false-positive rate.

The Chernoff threshold of the paper is conservative: on unwatermarked GLS responses scored with the same keys, its false-positive rate at p≤0.01p \le 0.01 is 0.1-0.2%. We therefore also set the threshold to a 1% false-positive rate, calibrated on the unwatermarked responses of the other two seeds (the measured rate on the held-out seed is 1.0--1.3%). For WMGLS on Llama-2-7B-chat, the detection rate rises from 2.1%, 11.8% and 15.8% to 9.9%, 29.7% and 36.7% on GSM8K, Code and Finance, and WMGLS and PFR are within 0.6 points in every setting. Other tests of the Aaronson watermark, such as the Gamma-tail statistic of −log⁡(1−U)-\log(1-U), are also applicable; we keep the statistic of the paper so that all methods are compared with the same test.

Low rates on Llama-2-7B-chat.

The score of a committed token exceeds 1/2 only to the extent that the target is uncertain: at a token with probability pv→1p_v \to 1 the expected excess vanishes (Prop 4). Low-entropy responses provide less information for detecting this watermark. Llama-2-7B-chat has the lowest LPPL of our targets (0.124, 0.215 and 0.276 on GSM8K, Code and Finance, against 0.379, 0.521 and 0.586 for Vicuna-7B) and short GSM8K responses. This limits every Gumbel-max watermarked method equally, including Wκ(p)W_\kappa(p) itself, independently of speculation. Giboulot and Furon (2024) also find very low detectability for the Aaronson watermark on Llama-2-7b-chat, the model with the lowest entropy in their study.

AuthorsOpenReview ↗

Smaller targets, deployment scope and summary

Weakness 7: Smaller models.

The construction and its guarantees do not depend on the size of the target. To test a smaller target empirically, we ran the full online protocol on TinyLlama-1.1B-Chat. We trained a one-layer EAGLE-3 draft from scratch on the same 800 GSM8K warm-up examples, then ran the GSM8K stream (100 chunks of 40) with the stream, sampling seeds, training seeds and keys of the Llama-2-7B-chat runs (K=6K=6, L=5L=5; 3 seeds; AATPS as in Table 1). As discussed in W5, a fixed key can make certain template-adjacent openings dominate, so the keyed methods use the masking control of Appendix D:

TinyLlama-1.1B-Chat, GSM8KAATPSAATPS, chunks 1–20 → 81–100Tokens/sANLPPTDetection @128
PFR, masked1.1741.128 → 1.19845.20.18895.1%
MSE, masked1.1971.146 → 1.22246.00.0179.4%
GLS1.4161.330 → 1.45654.3----
WMGLS, masked1.4181.332 → 1.45951.60.18995.1%

The method is effective at this scale. WMGLS is within 0.002 of the AATPS of GLS (1.418 against 1.416) and exceeds PFR by 20.8% (1.174), compared with 23--29% at 7B. Its token rate is 14% above that of PFR and 5% below that of GLS. Its watermark strength is that of PFR (ANLPPT 0.189 against 0.188, detection 95.1% for both), whereas MSE retains little watermark strength (ANLPPT 0.017, detection 9.4%). All methods speed up over the stream as the draft is trained.

We did not test targets far below 1B. There, the target forward pass is cheap relative to drafting, which limits speculative decoding itself, independently of the watermark.

Summary

We sincerely thank the reviewer again for the careful evaluation, constructive comments, and valuable suggestions. In response to the seven concerns raised, we have conducted extensive additional experiments and analyses, which have substantially strengthened our work.

Our main additions include:

  1. Theoretical clarifications (W1): Clearly distinguished the pathwise guarantee of watermark preservation from the empirically observed online sampling efficiency, supported by new experiments using deployment and fresh watermark keys.
  2. Additional baselines and efficiency studies (W2–W3): Integrated the contemporaneous MPFR method into WOSpec, demonstrating comparable watermark preservation and sampling efficiency. We also conducted matched nominal-budget comparisons and extensive draft-count sweeps across multiple workloads and seeds.
  3. End-to-end efficiency (W4): Measured wall-clock performance including online updates, demonstrating 7–12% lower total time than PFR across the evaluated settings and 14% lower elapsed time than frozen-draft serving in the asynchronous training experiment.
  4. Watermark robustness and detection (W5–W6): Conducted systematic key-sensitivity analyses, additional masked online experiments, and detection evaluations at a calibrated 1% false-positive rate, clarifying the effects of key reuse, response length, and detection coverage.
  5. Smaller-model applicability (W7): Extended our experiments to TinyLlama-1.1B-Chat, confirming comparable watermark strength and 20.8% higher AATPS than PFR.

We are grateful to the reviewer for motivating these substantial additional experiments and clarifications. We believe that the new results provide stronger theoretical and empirical support for WOSpec and address the reviewer's main concerns.

We sincerely hope that these additions satisfactorily address the concerns raised, and we would be very grateful if the reviewer could kindly reconsider their evaluation in light of the new evidence.

Thank you again for your time, consideration, and valuable feedback!

Comments on the paper

5 comments
AuthorsOpenReview ↗

Author Response Summary

Author Response and Summary of Rebuttal

We sincerely thank the Area Chair and all reviewers for their careful evaluations, constructive comments, and valuable suggestions.

In response, we have provided detailed point-by-point replies and conducted extensive additional experiments and analyses. Our main updates include:

  1. Theoretical clarifications: Distinguished the pathwise watermark-preservation guarantees from empirical online efficiency, clarified the effects of key-dependent draft updates, and corrected the fixed-KK qualification for GLS.

  2. Comparison with contemporaneous MPFR: Although MPFR is contemporaneous work and it is not required to compare to it, we nevertheless implemented it within our WOSpec framework and conducted direct comparisons, demonstrating comparable watermark preservation and sampling efficiency. Importantly, these experiments further demonstrate that our online serving-and-learning framework is orthogonal to the choice of compatible multi-draft coupling, strengthening its modularity and generality.

  3. Additional efficiency and robustness experiments: Conducted matched nominal draft-budget comparisons, expanded draft-count ablations, and measured end-to-end performance including online updates. We further investigated key sensitivity, response-length effects, and calibrated watermark detection, including additional controlled experiments.

  4. Implementation corrections and broader applicability: Corrected and validated the generation–detection masking inconsistency, confirming that the original reported results remain unchanged. We also extended our evaluation to TinyLlama-1.1B and SynthID-Text, providing further evidence of practical applicability beyond the original experimental settings.

We believe these additional results, clarifications, and corrections substantially strengthen the theoretical and empirical contributions of WOSpec and address the reviewers' main concerns.

We warmly welcome any further questions, comments, or discussion from the reviewers and the Area Chair. If the reviewers find our responses and revisions satisfactory, we would be very grateful if they could kindly reconsider their assessments and ratings in light of the clarifications, corrections, and new results.

Thank you again for your time, consideration, and constructive feedback!

AuthorsOpenReview ↗

Follow-up on Author Responses and Manuscript Revisions

Dear Reviewers and Area Chairs,

As the discussion period approaches its end in less than 35 hours, we would like to kindly follow up on our responses.

We have provided detailed responses to the reviewers' concerns and substantially revised the manuscript based on your constructive feedback, with all changes highlighted in blue. Our main additions include:

  1. Theoretical clarifications: Distinguished the pathwise watermark-preservation guarantee from empirical online sampling efficiency and clarified key-dependent draft updates.

  2. Comparison with contemporaneous MPFR: Although MPFR is a contemporaneous work, we implemented it within WOSpec and conducted experimental comparisons, even strengthening our contribution by showing our online framework supports different watermark-preserving samplings.

  3. Efficiency and robustness experiments: Conducted matched-budget comparisons, expanded draft-count sweeps, and evaluated key sensitivity and watermark detection. Importantly, online training reduces wall-clock time by 14% compared with frozen-draft serving, while WMGLS achieves approximately 7–12% lower end-to-end time than PFR across the evaluated settings.

  4. Broader applicability: Extended our evaluation to TinyLlama-1.1B and SynthID-Text. On TinyLlama, WMGLS achieves 20.8% higher AATPS than PFR while maintaining the same 95.1% detection rate, further supporting the broad applicability of our framework!

We sincerely hope these revisions have adequately addressed your concerns. We would greatly appreciate any further feedback from the reviewers and Area Chairs, particularly if any concerns remain or additional clarification would be helpful.

We would also be very grateful if you could kindly reconsider your assessments in light of these clarifications, corrections, and new experimental results.

Thank you again for your time, thoughtful feedback, and consideration!

The Authors

AuthorsOpenReview ↗

A Gentle Reminder as the Discussion Period Draws to a Close

Dear Reviewers and Area Chairs,

As the discussion period approaches its end (in 9 hours), we would like to kindly follow up on our responses.

We have provided detailed responses to all reviewers' concerns and substantially revised the manuscript based on your constructive feedback, with all changes highlighted in blue.

We particularly thank Reviewer rH3C for the constructive engagement and continued positive support. We would also greatly appreciate any further feedback from the other reviewers.

We sincerely hope these revisions have adequately addressed your concerns. We would be grateful if you could kindly reconsider your assessments in light of the strengthened contributions, extensive additional experiments, and revised presentation.

Thank you again for your time, thoughtful feedback, and consideration!

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.