acceptodds
Under review as a conference paper at ICLR 2027

Watermarkable Multi-Draft Speculative Sampling via Poisson Processes

Abstract

Large language models (LLMs) have achieved state-of-the-art performance across a wide range of tasks, motivating two important aspects of deployment: inference efficiency and output provenance, which can be tackled by speculative sampling and watermarking, respectively. However, recent works have shown that combining these two goals is highly nontrivial and can be potentially impossible. In this work, we develop a novel multi-draft speculative sampling algorithm based on Poisson processes that improves the frontier of this fundamental trade-off. The proposed algorithm has strong sampling efficiency on its own and, more interestingly, is naturally watermarkable: we can embed an unbiased watermark without degrading speculative acceptance. Moreover, our algorithm is based on an exact list-coupling-without-communication scheme and satisfies certain level of drafter invariance that provides robustness to drafter substitution. Our multi-draft construction jointly provides exact target sampling, stable target-keyed watermark strength as the number of drafts increases, and strong empirical sampling efficiency.

Primary area:
generative models
Keywords:
speculative decoding, LLM watermarking, channel simulation, relative entropy coding, information theory

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_v89dOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MPFR, a speculative sampling algorithm based on a Poisson process that uses multiple drafts to improve sampling efficiency while preserving watermark strength. Its core contribution is to extend sampling to multiple drafts using the mapping theorem for Poisson processes and to ensure drafter invariance through a keyed Poisson source indexed by context.

Critical strengths
  1. The proposed method has theoretical value and may support future work combining watermarking with speculative sampling.
  2. The stopping-time drafter invariance has practical value and offers a reasonable compromise between strong invariance and conditional invariance.
Critical weaknesses
  1. The writing is difficult to follow. More intuitive explanations and illustrations would improve readability.
  2. The core contribution is incremental. For example, Sections 4.3 and 4.4 largely restate existing work, with the main new contribution limited to the extension to multiple drafts.
Questions for authors
  1. Section 4.3 explicitly acknowledges that PFR with a single draft is equivalent to MWS. Is the paper’s core contribution therefore entirely the extension to multiple drafts?
  2. The efficiency improvement of MPFR over GLS in Table 1 appears small. Does this suggest that increasing BB offers only limited gains for MPFR? Could the authors report the mean and standard deviation across three different seeds to present the results in more detail?
  3. The method achieves drafter invariance at a stopping time, as defined in Definition 4.3, rather than strong drafter invariance. Is this because the stronger property cannot be achieved, or because it is unnecessary?
  4. The paper reports that ANLPPT remains stable as BB changes Δ≤0.001\Delta \leq 0.001. Is this because each token is still coupled as in the case of a single draft, making the watermark signal independent of BB by construction? Could the authors compare against an alternative in which the watermark signal naturally weakens when multiple drafts are used?
4 comments
AuthorsOpenReview ↗

Responses to presentation, contribution, and empirical gains

We have provided detailed responses to your points and substantially revised the manuscript using the permitted additional page, with all changes highlighted in blue.
Thank you for recognizing the theoretical value of our Poisson-process construction and the practical importance of stopping-time drafter invariance. We also appreciate the thoughtful questions about the novelty of the multi-draft extension, experimental variability, and watermark stability. We address weaknesses and four questions below.

W1: presentation

The revision adds a concise step-by-step PFR/MPFR workflow explanation, introduces the roles of lookahead LL and draft multiplicity BB with a simple two-draft example, adds a numerical comparison table against MWS/MSE and the multi-draft GLS baseline in Appendix D, and expands the captions of Figures 1 and 2 to make the intended trends explicit.

W2 / Q1 — "is the core contribution only the multi-draft extension?"

We respectfully argue that our contribution is more than a naive multi-draft extension. First, compared with GLS, our scheme relies heavily on the properties of Poisson processes, and the construction itself is nontrivial; moreover, it achieves better performance. Second, the key observation is that, although Hu and Huang (2024) identified an inevitable tradeoff between speculative sampling efficiency and watermark strength, we can use the multi-draft extension as a third dimension to address this fundamental tradeoff. Of course, achieving higher AATPS is straightforward; the key question is whether we can indeed achieve better wall-clock time without hurting watermark strength, and we answer this affirmatively. Moreover, technically, we also provide:

  • A context-indexed rather than draft-indexed keyed source. This is the design decision that makes the rest work; see Q4, where we show that the draft-indexed alternative loses detector evidence as BB grows.

  • A proof that the construction is drafter-invariant at B>1B>1. The submission defined multi-draft invariance (Appendix C.2) but proved it only for the single-draft Algorithm 3. We have closed that gap; the new result is stated in our response to Reviewer vB4D and holds for heterogeneous drafter sets, different temperatures and different BB.

  • An acceptance lower bound strictly tighter than the corresponding bound for the existing drafter-invariant multi-draft scheme (Theorem 5.1, B.1–B.4). Following Reviewer vB4D we will state precisely that this orders the bounds, not the true acceptance probabilities.

Such constructions, theoretic guarantees, and practical performance on both sampling and watermarking, are nontrivial contributions.

Q2 — "the gains in Table 1 look small; can you report mean ± std over seeds?"

Table 1 averages 3 seeds × 1000 prompts per cell; the per-seed standard deviations are in Appendix Tables 6-9 but were dropped from the main table, which makes the result look like a single run. We will restore them.

The underlying evidence is stronger than the cell means suggest. Across all 48 paired seed-by-configuration comparisons (identical prompts and seed protocol, both decoders run sequentially in one process on matched hardware), MPFR has the higher token rate in 48/48, with per-cell gains up to +5.4%. Per-cell seed variability is small relative to the gap: the standard deviation of the MPFR − GLS token-rate difference across seeds has median 0.21 tok/s, while the difference itself ranges 0.34–2.17 tok/s, and its sign never flips.

On whether increasing BB offers only limited gains — this is the more important half of the question, and the answer is the opposite: the gain grows with BB, and Table 1 understates it because it fixes L=4L=4, which is where the BB-dependence is weakest. On a unified-hardware (L,B)(L,B) sweep (see response to Reviewer ZXzX), the gain over the multi-draft baseline on the Llama pair shows a particularly pronounced gain at larger BB, especially at small LL (the intermediate values are not always monotone):

B=2B=2B=4B=4B=6B=6B=8B=8
L=1L=1+2.60%+4.63%+6.29%+9.25%
L=4L=4 (main table)+1.06%+1.49%+1.45%+4.62%

The same trend holds at scale and gets stronger, not weaker: on Qwen2.5-72B the paired tt-statistic for token rate goes from −0.48-0.48 at B=2B=2 to +9.94+9.94 at B=8B=8 (and +10.85+10.85 with a 1.5B drafter), i.e. +11.62% and +10.47% at B=8B=8. The mechanism is in the ZXzX reply: wider draft lists produce more duplicate prefixes, context merging collapses them (66% fewer expanded tree nodes at B=8B=8 versus 52% at B=4B=4), and the target's verification batch shrinks accordingly.

Increasing BB is where the method pays off; L=4L=4 with small BB is the least favourable corner of the grid. We keep L=4L=4 as the main configuration, since it matches the protocols of the prior work, and add the full grid alongside it so the dependence is more visible.

AuthorsOpenReview ↗

Responses to drafter invariance and ANLPPT's stability, and Summary

Q3 — "invariance holds at a stopping time, not strongly. Is strong invariance unachievable, or unnecessary?"

In short, we consider it unnecessarily strong for us. However, we clarify that the right distinction is between a variable-length speculative block and the complete generated response, rather than a general impossibility of strong invariance. The stopping length τ\tau depends on the drafter, so comparing the laws of Y1:jY_{1:j} without specifying whether position jj was emitted can be ambiguous. We do not claim that every coupling with a shortened block must fail a strong-invariance definition.

What we establish is a particularly operational property: for a fixed target, key, prompt, and shared context-indexed randomness, every emitted block is a prefix of the same target-keyed sequence ZZ, i.e. Y1:τ=Z1:τY_{1:\tau}=Z_{1:\tau} almost surely, regardless of the permitted draft choice. Conditional statements at a given block length are understood only on block-length events with positive conditional probability. When MPFR is iterated to a fixed total output length, the entire resulting output is pathwise identical across compatible drafters. Thus the drafter governs how many verification passes are needed, but not the target-keyed text that is produced. We now give the multi-draft proof in Appendix C (Proposition 5.2).

Empirically, extending the drafter-substitution ablation from B=1B=1 to B∈{4,8}B \in \{4,8\} and varying the drafter in both scale (1B → 3B) and temperature (0.5 / 1.0 / 1.5): AATPS spans 3.47–4.39 and token rate 36.4–45.1 across the four conditions and both draft-count settings, while ANLPPT-U varies by ≤0.0012\le 0.0012 and TPR@1%FPR by ≤1.7\le 1.7 points. Efficiency responds strongly to the drafter; the detector-visible signal does not.

Q4 — "is ANLPPT's stability in BB true by construction, and hence not a result? Can you compare against a scheme whose signal weakens with more drafts?"

It is true by construction — and we now prove why: each emitted token's keyed pivot is a function of (key, context, target distribution) only, never of BB. So the reviewer's reading is correct, and we will present the stability as a consequence of a design decision rather than as a separate empirical finding.

But the construction is not free, which is what the requested comparison shows. Stability requires indexing the Poisson source by context. Index it by draft slot instead — the natural way to watermark an active-set multi-draft scheme, which is the structure the active-set baseline uses — and the emitted token is determined by a maximum over the realised active draft subset. The detector cannot recover that subset from the output text, since it depends on the accept/reject trajectory; it can only regenerate the maximum over all BB rows, which is not the pivot that produced the token.

In a synthetic setting that isolates the keying index (∣V∣=64|\mathcal V| = 64, Gaussian logits, 50% active fraction, 2000 tokens), with each scheme calibrated against its own correct null so that the comparison does not rest on mis-calibration, the per-token evidence (E[A∣H1]−E[A∣H0])/sd[A∣H0](\mathbb E[A \mid H_1] - \mathbb E[A \mid H_0]) / \mathrm{sd}[A \mid H_0] is:

BBcontext-indexed (ours)draft-indexed
13.3433.222
23.2522.483
43.2792.432
63.2682.467
83.2532.427

Ours is flat in BB; draft-indexed keying loses about 25% as soon as B>1B>1 and does not recover. At B=1B=1 the two coincide, since the active subset is then necessarily the full set — a consistency check on the simulation. We will report this as a mechanism check and label it as such, not as a model-level result.

Separately, we have verified at B>1B>1 that keying costs nothing in efficiency. On the Llama-3.1-8B/3.2-1B pair at L=4L=4, comparing MPFR against an otherwise identical decoder whose keyed Poisson source is replaced by unkeyed fresh noise (five cells over two datasets and B∈{2,4,8}B \in \{2,4,8\}): Δ\DeltaAATPS lies in −0.01%…+0.57%-0.01\%\ldots+0.57\% and Δ\DeltaTR in −0.02%…+0.71%-0.02\%\ldots+0.71\%, while ANLPPT-U rises by 23–90×. The same holds in the single-draft regime (PFR 41.46 vs PFR-NoWM 41.72 tok/s at L=1L=1, and equally at every other LL).

Summary

We sincerely hope that these explanations, corrections, and new results address the reviewer's concerns. We have revised the manuscript accordingly, with changes highlighted in blue. We appreciate the reviewer's time and constructive feedback, and would be grateful if the reviewer could kindly reconsider their evaluation in light of the strengthened theoretical and empirical evidence.

I thank the authors for their response. However, the extensive references to prior work in Sections 4 and 5 still make the paper’s actual contribution difficult to assess. Given its focus on watermark sampling efficiency, the experiments should offer more direct insight than the theory, yet those in the main text appear limited. I therefore remain inclined to keep my original score.

AuthorsOpenReview ↗

One More Clarification of Our Contributions in the Revised Manuscript

We sincerely thank the reviewer for the follow-up and for taking the time to consider our paper. We understand your concern and agree that our contributions could be presented more clearly.

We have therefore further revised the manuscript, with all new changes highlighted in red\color{red}{\text{red}}, focusing on:

  1. Section 4 (Background): We clarify that (1) this section is mainly for background, and hence it contains components that resemble existing work; and (2) the single-draft version can be viewed as a key representation that coincides with MWS, but it is precisely this representation that allows us to exploit the properties of Poisson processes to construct our multi-draft algorithm. A direct extension of MWS is unclear, and this representation serves as a key translation.

  2. Section 5 (Contribution): In short, we design a novel multi-draft algorithm (Algorithm 4) with new theoretical analysis that achieves desirable sampling efficiency while preserving watermark strength. In contrast, the baseline GLS does not incorporate watermarking. More importantly, our key contribution is to introduce draft multiplicity as an additional dimension of the watermark–efficiency trade-off, improving efficiency while preserving watermark strength.

Moreover, we have also added the experimental results from our previous rebuttal to the revised paper, highlighting that MPFR achieves 48.30 tokens/s versus 33.62 for MWS, with comparable watermark strength.

We would be very grateful if you could take one more look at the red\color{red}{\text{red}}-highlighted revisions, particularly at the end of Sec 1, the beginning of Sec 4, and Sec 5. Our primary intention is to make the paper's contributions easier to identify and assess, rather than to persuade you to change your judgment (of course, we would be extremely happy if you agree that our revisions based on your suggestions make the paper clearer to you).

We fully respect your assessment, whatever your final decision may be. Thank you again for your time, thoughtful feedback, and help in improving our paper!

Reviewer_vB4DOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to combine exact speculative decoding with unbiased LLM watermarking. It proposes a Poisson Functional Representation (PFR)-based coupling in which the target model and draft model share key-dependent Poisson randomness. The single-draft construction is equivalent to a shared Gumbel/DeltaGumbel coupling, while the main proposed extension (MPFR) uses mapped Poisson-process order statistics to generate multiple draft samples from the same source.

Critical strengths

The paper addresses a relevant problem at the intersection of speculative decoding and unbiased watermarking. Its multi-draft construction is technically coherent, which uses context-indexed shared Poisson sources and mapped order statistics to couple one target-side sample with multiple draft proposals. Under the paper's stated idealized assumptions, the accompanying analysis establishes exact sampling from the processed target distribution, while the experiments consistently show that the measured per-token watermark signal remains stable as the number of drafts increases. The authors are also appropriately transparent that the single-draft construction is equivalent to the existing shared Gumbel/DeltaGumbel coupling and locate the potential novelty in the multi-draft extension.

Critical weaknesses
  1. The paper seems to overstate what the multi-draft construction establishes about the watermarking--efficiency trade-off. Hu and Huang's no-go result concerns a single-draft/information-structure setting. MPFR introduces an additional resource dimension---multiple draft proposals and the corresponding tree/branching computation---which can naturally increase the probability that a target-keyed winner appears in the proposal list. This is a reasonable way to improve efficiency while retaining a target-keyed watermark, but it does not by itself resolve or overturn the original trade-off. The paper should frame the contribution more precisely as constructing a multi-draft coupling that preserves per-token watermark evidence under additional proposal and branching budget, rather than broadly claiming a fundamental improvement of the trade-off frontier.

    Relatedly, Theorem 5.1 only lower-bounds the probability that at least one draft matches the target token at the current step. It is not a characterization of block-level accepted length, AATPS, or end-to-end throughput. In a multi-step speculative tree, later contexts, multiplicities, and merging behavior determine the effective number of drafts available at subsequent positions. Thus, a stronger first-step list-hit lower bound does not automatically imply better AATPS or wall-clock speed. Moreover, the comparison in Appendix B appears to show that the MPFR lower bound is stronger than a prior lower bound, rather than proving that MPFR's true acceptance probability is always higher than that of the prior method. Could the authors clarify precisely which quantity is ordered theoretically, and avoid interpreting a lower-bound comparison as a general acceptance or frontier dominance result?

  2. The drafter-invariance/robustness claim is central to the motivation but is not fully established for MPFR. Section 5 presents context-indexed Poisson sources and stopping-time drafter invariance as an important distinction from draft-indexed active-set schemes. However, Proposition C.3 proves the property only for the single-draft Algorithm 3. Appendix C.2 introduces multi-draft invariance definitions, but does not appear to provide a corresponding theorem and proof for Algorithms 4/10, i.e., the main MPFR method. The empirical drafter-substitution study likewise evaluates only single-draft PFR, while the main contribution is the B>1B>1 setting.

Does MPFR retain the claimed invariance when different draft models are used within the same list? Does it remain stable under heterogeneous draft models, different temperatures, or different draft counts? More generally, what operational guarantee does the proposed stopping-time notion provide after conditioning on both the shared randomness and the realized block length?

  1. The empirical evidence does not establish a practically meaningful efficiency improvement. MPFR has higher AATPS in 15 of 16 matched settings in Table 1, but the differences are generally small, and GLS is slightly better in the remaining setting. Token-rate improvements are also modest. Since MPFR requires keyed Poisson-clock generation, top-BB selection, and construction of a branching context tree, the paper should report the associated costs: number of visited contexts/tree nodes, merge rate, target batch size, memory usage, and the time spent on clock generation and selection. Without this breakdown, it is difficult to assess whether the observed wall-clock gains would persist in larger or production-like settings.

    The evaluation is also limited to 7--8B target models, L=4L=4, B≤8B\leq 8, two datasets, and a fixed top-kk decoding regime. Could the authors provide evidence that the claimed advantage persists as the target--draft gap, lookahead, draft count, token entropy, or model scale changes? In particular, a multi-draft unkeyed-MPFR/MPFR-no-watermark ablation is needed to directly verify that the watermark keying itself does not introduce an efficiency penalty in the setting where the paper claims its main improvement.

Questions for authors

See questions stated in weaknesses.

5 comments
AuthorsOpenReview ↗

Responses to trade-off frontier and Theorem 5.1

We have provided detailed responses to the reviewer's points and substantially revised the manuscript using the permitted additional page, with all changes highlighted in blue. The new theoretical result is stated in Section 5 and proved in Appendix C; the additional efficiency and ablation results are summarized in Section 6 and detailed in Appendix D.

We thank the reviewer for recognizing that the Poisson-process construction is technically coherent and that context-indexed coupling preserves a stable watermark signal as the draft count grows. We particularly appreciate the careful mathematical assessment, which identified two places needing sharper statements and one missing multi-draft proof. We address all three points below: (a) state precisely what our acceptance theorem establishes, (b) provide a multi-draft theorem and proof, and (c) report the requested cost accounting and additional ablations.

W1 — the trade-off frontier claim, and what Theorem 5.1 actually orders

We agree, and we will make the claims precise.

Which quantity is ordered theoretically. Precisely one: the first-step list-hit probability, and only as a lower bound. Theorem 5.1 bounds below the probability that at least one of the B proposals at the current context equals the target token. Appendix B.2 then compares our lower bound against the corresponding lower bound of the prior drafter-invariant scheme and shows ours is the tighter of the two. Neither statement orders the realized acceptance probabilities, and neither extends to block-level accepted length, AATPS or wall-clock throughput. The reviewer's multi-step reading is right: at subsequent positions the effective draft count is determined by context multiplicities and merging, so a tighter first-step bound does not propagate to a higher AATPS by itself.

We add one sentence immediately after Theorem 5.1 naming the ordered quantity, and restate Appendix B.2 as a bound-vs-bound comparison, so the lower-bound result cannot be read as acceptance or frontier dominance.

This scoping is borne out by our own measurements, which is why we are comfortable stating it: at a 48:1–144:1 target–draft ratio the first-step edge no longer shows up in accepted length, exactly as a first-step-only result predicts.

Concretely, we replace "improves the frontier of this fundamental trade-off" in the abstract, introduction and conclusion with the precise claim: a multi-draft coupling that preserves per-token target-keyed watermark evidence under an additional proposal/branching budget; add, after Theorem 5.1, a statement naming which quantity is ordered (the bounds) and which is not (the realized acceptance probabilities); revise Figure 1's caption and surrounding text accordingly; and keep the empirical acceptance comparison as the evidence for the practical claim, stated as such.

On the empirical side, the ordering we can support is a paired wall-clock one. Across 48 paired seed-by-configuration comparisons at L=4 (3 seeds × 2 model pairs × 2 datasets × B ∈ {2,4,6,8}, n=1000 prompts per seed and configuration; identical prompts and seed protocol, both decoders run sequentially in one process on one GPU), MPFR has the higher measured token rate in 48/48, with per-cell gains of +1.0% to +5.4%.

AuthorsOpenReview ↗

Responses to multi-draft drafter invariance

This was a real gap, and we thank the reviewer for locating it exactly: Appendix C.2 defines strong / conditional / stopping-time invariance for the multi-draft setting, but Proposition C.3 establishes the property only for the single-draft Algorithm 3, while the contribution is the B > 1 construction.

The missing result follows from a structural feature of Algorithm 10 that we had not drawn out. The target-side winner at a context c is y(c)=MPFR(P(⋅∣c),Π(c),1),Π(c)=G(k,c)y(c) = MPFR(P(·|c), Π(c), 1), Π(c) = G(k, c), whose arguments are the context, the key and the target distribution only — it does not depend on ν(c)\nu (c) (how many drafts occupy c), on D(c) (which tokens they proposed), on the draft model(s) Q, or on B. The drafts enter solely through the acceptance test y(c⋆) ∈ D(c⋆), i.e. they determine where the block stops, not what the emitted path is.

Proposition 5.2 (multi-draft pathwise / stopping-time drafter invariance; proof in Appendix C). Fix the target model, prompt hh, key kk, and common context-indexed randomness R={G(k,c):c a context}\mathcal R=\{G(k,c):c\text{ a context}\}. Define the target-keyed chain by Z1=PFR(P(⋅∣h),G(k,h))Z_1=\mathrm{PFR}(P(\cdot\mid h),G(k,h)) and Zs=PFR(P(⋅∣h∥Z1:s−1),G(k,h∥Z1:s−1))Z_s=\mathrm{PFR}(P(\cdot\mid h\|Z_{1:s-1}),G(k,h\|Z_{1:s-1})) for s≥2s\ge2. For any draft count BB, lookahead LL, and draft model configuration permitted by Algorithm 10, with emitted block length τ\tau, the emitted block satisfiesY1:τ=Z1:τalmost surely.Y_{1:\tau}=Z_{1:\tau}\quad\text{almost surely}.

Please refer to the revised paper for more details on theoretic statement and proof.

This answers the reviewer's three follow-up questions directly:

  • Does invariance hold when different draft models are used within the same list? Yes for the target-path property under the compatible heterogeneous extension just specified; the theorem for the submitted homogeneous Algorithm 10 requires no such extension.
  • Under different temperatures or draft counts? The target-path identity persists under any compatible variation because neither BB nor the draft distributions enter y(c)y(c); the acceptance bound for homogeneous drafts is a separate result.
  • What operational guarantee does the stopping-time notion provide after conditioning on both the shared randomness and the realized block length? Exactly this: the emitted prefix equals the target-side keyed chain. The drafter governs how many target forward passes are needed to walk that chain — efficiency — and not the conditional law of its contents — provenance.

The stopping-time formulation is needed because τ\tau is drafter-dependent: different drafters may emit blocks of different lengths. We do not claim a general impossibility theorem for strong drafter invariance. In fact, the pathwise result gives an especially useful stronger statement at the level of a complete response: when decoding is continued to the same prescribed total output length, the target-keyed output is exactly the same for every compatible drafter. Stopping-time invariance characterizes the corresponding guarantee for the individual variable-length blocks.

Empirically, we extended the drafter-substitution ablation from B=1 (Figure 2(a)) to B ∈ {4,8}, varying the drafter in scale (Llama-3.2-1B → 3B) and temperature (0.5 / 1.0 / 1.5), with the target, key, prompts and target-side sampling held fixed (CNN/DailyMail, n=300):

conditionBAATPSTRANLPPT-UTPR@1%FPR (128 tok)
D0: 1B, T=1.043.77143.630.05810.673
D0: 1B, T=1.084.03545.050.05780.670
D1: 3B, T=1.044.19436.370.05900.673
D1: 3B, T=1.084.38937.140.05860.670
D2: 1B, T=0.543.47040.400.05820.677
D2: 1B, T=0.583.62841.900.05860.677
D3: 1B, T=1.543.62542.490.05880.687
D3: 1B, T=1.583.94444.280.05900.687

The drafter moves efficiency substantially — AATPS spans 3.47–4.39 and token rate 36.4–45.1, with the 3B drafter giving the highest acceptance but the lowest throughput, since its own forward pass costs more than the extra acceptance returns. The detector-visible quantities barely move: across the four conditions ANLPPT-U ranges 0.0009 at B=4 and 0.0012 at B=8, and TPR@1%FPR ranges 1.4 and 1.7 points. For contrast, the submission's Figure 2(a) shows MSE varying by 12–22 points and MSE-Pseudo by 6–11 points under the same substitutions. Efficiency responds to the drafter; the detector-visible signal does not — which is the division of labour the new proposition asserts.

AuthorsOpenReview ↗

Responses to cost accounting and ablations

W3 — cost accounting, the B > 1 unkeyed ablation, and evaluation breadth

We agree that without the cost breakdown the wall-clock claim cannot be assessed, and we have now measured all six quantities the reviewer lists. One book-keeping note first: the reviewer correctly reads Table 1 as 15 of 16 matched settings on AATPS. The grid below is a new measurement — one machine, one harness, n=300 per cell, L and B both swept — and is not a restatement of Table 1; on it MPFR leads 16/16 on AATPS and 14/16 on measured token rate. We report both rather than replacing one with the other, and Appendix D.1's hardware note (see §d) is why the submitted tables could not be compared directly in the first place.

(a) The B > 1 unkeyed ablation

The submission's keying ablation existed only at B=1. We added an MPFR-NoWM decoder: the keyed Poisson source is replaced by an unkeyed fresh-noise source with the draft/verify/cache path untouched, so the comparison isolates keying. Llama pair, L=4, n=300, paired in-process:

datasetBΔAATPSΔTRANLPPT-U: NoWM → MPFR
CNN/DM2+0.03%+0.39%0.0024 → 0.0587
CNN/DM4−0.01%+0.14%0.0025 → 0.0581
CNN/DM8+0.45%+0.59%0.0021 → 0.0578
ELI54+0.57%+0.71%0.0016 → 0.1439
ELI58+0.43%−0.02%0.0020 → 0.1459

Keying is not measurably an efficiency cost at B > 1 — ΔAATPS within −0.01%…+0.57%, ΔTR within −0.02%…+0.71%, four of five positive — while the detector-visible signal rises by 23–90×. The same holds at B=1 (PFR 41.46 vs PFR-NoWM 41.72 tok/s, and equally at every other L).

(b) Component accounting

Measured with the keyed-RNG step timed separately, L=4, 30 instrumented prompts per cell, one A100-80GB, MPFR and the baseline in one process.

Per-block latency (ms/block), MPFR:

pairBblockdraft fwdtarget fwdPFR arrival samplingkeyed RNGcache opslogits
Llama483.2948.82 (58.6%)26.86 (32.3%)1.93 (2.3%)0.347 (0.42%)2.450.88
Llama885.4348.57 (56.9%)28.16 (33.0%)2.51 (2.9%)0.446 (0.52%)2.700.90
Qwen4101.4569.59 (68.6%)24.05 (23.7%)1.75 (1.7%)0.359 (0.35%)2.800.79
Qwen899.3065.83 (66.3%)24.76 (24.9%)2.28 (2.3%)0.476 (0.48%)3.010.76

The keyed clock generation the reviewer asked about costs 0.35–0.48 ms/block, i.e. 0.35–0.52% of block time. Taking the whole sampling path together (arrival sampling + keyed RNG + logits processing), MPFR spends 3.0–3.9 ms/block against the active-set baseline's 1.5–1.8 ms (Gumbel sampling 0.76–1.05 + logits 0.70–0.78), so the selection step is 1.4–2.1 ms/block more expensive — about 2% of block time, and repaid by the target-batch saving below.

Draft tree, visited contexts, target batch (MPFR, per block):

pairBtree nodesvisited contextstarget fwd callstarget batchdraft fwd calls
Llama47.6810.671.0010.673.98
Llama810.8515.841.0015.843.99
Qwen48.7012.001.0012.004.00
Qwen813.3419.281.0019.284.00

Verification is one batched target forward per block over the visited contexts. On the merge rate: an unmerged list of B drafts at lookahead L would expand L·B nodes; context merging brings this to 7.68 of 16 at B=4 and 10.85 of 32 at B=8 (Llama), i.e. 52% and 66% fewer expanded nodes, with the saving growing in B because duplicate prefixes become more frequent as the list widens.

This provides direct timing evidence for the wall-clock advantage. At Llama B=8 the two arms spend almost the same time in the drafter (48.57 vs 50.34 ms) but MPFR's target forward is 4.7 ms/block cheaper (28.16 vs 32.82), consistent with the smaller verification workload produced by context merging. That saving exceeds the 1.4–2.1 ms sampling overhead in this configuration and helps explain the measured throughput gain. The advantage tends to grow with BB in the most favorable configurations (§c), although it is not universally monotone.

Incremental peak memory, for completeness, depends on the model pair rather than on the method:

pairBMPFRactive-set baselineΔ
Llama41076.7 MiB989.6 MiB+8.8%
Llama82046.7 MiB1942.0 MiB+5.4%
Qwen4456.8 MiB912.4 MiB−49.9%
Qwen8839.5 MiB1817.4 MiB−53.8%
MPFR costs 5–9% more on Llama and about half as much on Qwen. Its footprint tracks the retained per-level draft KV caches, and hence the target's KV geometry (Llama-3.1-8B has 8 KV heads × 32 layers against Qwen2.5-7B's 4 × 28, a 2.3× ratio, matching the observed 2.36× in MPFR's peak), whereas the baseline's is nearly pair-independent. Both are a small fraction of the weights at these scales.
Reviewer_ZXzXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a watermarkable multi-draft speculative sampling method based on Poisson processes. The strategy is to extend Poisson Functional Representation (PFR) to multi-draft coupling (MPFR), and to enable efficient speculative decoding while preserving the target distribution and watermark strength. The authors provide theoretical guarantees on acceptance probability and stopping-time drafter invariance. Experiments on Qwen and Llama show that MPFR maintains strong watermark detectability while achieving competitive sampling efficiency and slightly higher throughput than existing Gumbel-max list sampling (GLS) methods.

Critical strengths

The paper presented a technically solid and well-motivated approach. It is a good idea to use Poisson-process-based coupling to extend single-draft sampling to the multi-draft setting while preserving watermark strength.

Critical weaknesses
  1. The paper focuses heavily on improving sampling efficiency (AATPS), but this seems insufficient for evaluating tree-based speculative decoding. Increasing the number of drafts is naturally expected to improve acceptance efficiency, so this result itself is not surprising. A more important question is whether these gains (are large enough) can actually offset the additional computational cost and overhead of constructing and verifying a larger draft tree.

  2. It is noted that the authors report throughput comparisons against GLS, a more meaningful comparison would be against existing single-draft watermark speculative decoding methods (MSE and MWS). In fact, some results in the appendix shows that MPFR can still be slower than single-draft PFR with a shorter lookahead, or even autoregressive watermarked decoding. This raises concerns about how much practical acceleration the proposed method actually provides.

Questions for authors
  1. See the weaknesses.
  2. One result, which is a bit surprising, is that MPFR consistently achieves slightly higher throughput than GLS, even though MPFR additionally incorporates watermarking. Could the authors explain where this improvement comes from?
3 comments
AuthorsOpenReview ↗

Responses on comparison against MSE and MWS

Thank you for recognizing that Poisson-process multi-draft coupling is a promising idea for preserving watermark strength, and for noting the encouraging measured throughput. We have provided detailed responses to your points and substantially revised the manuscript using the permitted additional page, with all changes highlighted in blue. The new theoretical result is stated in Section 5 and proved in Appendix C; the additional efficiency and ablation results are summarized in Section 6 and detailed in Appendix D.

Both weaknesses are well taken, and the second concern proved correct in an important operating regime. We reran the efficiency comparison in a unified harness on an NVIDIA A100-SXM4-80GB, running MPFR and its baseline sequentially in one process on matched prompts. Unless otherwise specified, the new (L,B)(L,B) sweeps use n=300n=300 prompts per cell; the separate L=4L=4 three-seed comparisons use n=1000n=1000 prompts per seed, and the scale and instrumentation studies state their own sample sizes. These are distinct experiments, each compared on matched hardware, rather than one pooled measurement.

W2 — the comparison against MSE and MWS, and the slower-than-AR observation

You ask for two things here: a comparison against the existing single-draft watermarked speculative methods (MSE, MWS) rather than only against GLS, and an explanation of the appendix rows where MPFR is slower than single-draft PFR at a shorter lookahead or than autoregressive watermarked decoding. We take them in that order.

The comparison against MSE and MWS at matched watermark strength

This is the comparison we should have led with, and on the Llama pair it is clearly in our favour. The relevant axis is throughput at comparable watermark strength, since a method that carries a weaker signal is not paying the same price:

Each method is shown at its own best lookahead, so none is handicapped by an operating point chosen for another:

method (Llama-3.1-8B/3.2-1B, same card)token rateANLPPT-Usignal
PFR (single-draft)42.04 (L=3)0.0562strong
MWS (Hu & Huang, strength end)33.62 (L=1)0.0617strong
BASIC-UWM (AR + watermark)35.27 (L-indep.)0.0622strong
MSE (Hu & Huang, speed end)35.25 (L=1)0.0248weak
MSE-Pseudo (He et al.)32.78 (L=2)0.0149weakest

Against the two strong-signal references we are +25.0% (MWS) and +19.2% (AR + watermark). The MSE variants are both weaker and slower, so we are ahead on both axes rather than trading one for the other. Our multi-draft optimum on this pair reaches 48.30 tok/s, i.e. +43.7% over MWS.

The appendix rows where MPFR is slower

The reviewer is right, and on the Qwen pair this is not a measurement artifact. Measured in one harness on one card, Qwen2.5-7B/0.5B:

methodL=1L=2L=3L=4ANLPPT-U
BASIC-UWM (AR + watermark)39.7839.7839.8539.870.0492
VSPS (speculation only, no watermark)38.6834.6430.9127.12~0
PFR (single-draft)38.5433.7829.9225.390.0512
MSE34.8229.9425.4721.770.0173
MWS32.2227.4724.1221.760.0489
MSE-Pseudo31.9628.3226.3124.150.0142

At L=4 — the setting our main table reports — multi-draft MPFR reaches 26.9–32.8 tok/s against the AR watermarked reference's 39.87. Even the best B=8B=8 cell is about 18% slower, and the lower-BB cells are slower still. So the pattern the reviewer identified is real, and we should not have presented L=4 results without this context.

The slowdown is not specific to watermarking or MPFR. The decisive row is VSPS: unwatermarked speculation-only decoding at L=4 also reaches only 27.12 tok/s on this pair, likewise far below the AR reference. Four sequential draft forwards per block are too costly relative to the achieved acceptance gain in the measured Qwen implementation. L=4 is simply a poor operating point for this model pair, and we inherited it from the protocol of the prior single-draft work (Hu & Huang 2024) in order to be comparable, not because it is throughput-optimal.

At the throughput optimum for this pair, MPFR is ahead of both references:

Qwen2.5-7B/0.5BTRvs AR + watermarkvs best single-draft PFR
MPFR at L=4, B=8 (main table)32.75−17.9%−15.0%
MPFR at L=1, B=8 (optimum)43.51+9.1%+12.9%

On the Llama-3.1-8B/3.2-1B pair the picture is different, and more favourable:

Llama-3.1-8B/3.2-1BTRvs AR + watermark (35.27)
MPFR at L=1, B=2 (weakest cell)43.70+23.9%
MPFR at L=4, B=8 (optimum)48.30+36.9%

Here every one of the 16 (L,B) cells is above the AR reference, and the optimum even exceeds the unwatermarked speculation-only reference (48.30 vs 43.36, +11.4%).

The two pairs disagree about the direction of L, and we think that is the substantive finding. Token rate decreases with L on Qwen (39.0 → 26.9 at B=2) and increases with L on Llama (43.4 → 48.3 at B=8). The difference tracks the measured draft-to-target execution-cost ratio, rather than the model-parameter ratio: at L=4,B=8L=4,B=8, draft/target forward times are 65.83/24.76 ms for Qwen and 48.57/28.16 ms for Llama. The throughput-optimal lookahead is therefore model-pair specific and cannot be carried across pairs. The revision will report both grids, name the optimum for each pair, and state plainly that L=4 is not throughput-optimal for the Qwen pair.

One further comparison the submission could not make correctly for the same reason — throughput at matched watermark strength. On the Llama pair, PFR carries ANLPPT-U 0.0562, the same band as MWS (0.0617) and BASIC-UWM (0.0622), at 42.04 tok/s against 33.62 and 35.27: +25.0% and +19.2%. The MSE variants are both weaker (0.0248 / 0.0149) and slower (35.25 / 32.78), so against Hu & Huang's two endpoints and the He et al. variant we are ahead on both axes rather than trading one for the other.

W1: "AATPS alone is insufficient for tree-based speculative decoding; the gains must offset the cost of building and verifying a larger draft tree"

We agree, and the unified grid makes the reviewer's point concretely. Llama pair, token-rate gain of MPFR over the drafter-invariant multi-draft baseline (GLS):

LB=2B=4B=6B=8
1+2.60%+4.63%+6.29%+9.25%
2+3.31%+4.17%+5.15%+8.51%
3−0.47%−1.41%+0.92%+5.60%
4+1.06%+1.49%+1.45%+4.62%

MPFR has higher AATPS in 16/16 cells and higher measured token rate (wall clock time) in 14/16. Both exceptions sit at L=3: B=4 is a reproducible negative (−1.41%, paired t = −3.16 over 300 prompts on a re-run with a dedicated GPU), and B=2 is within noise (−0.47%, t = −0.97). We report them as measured. The L=3 row above comes from that dedicated re-run, which we will note in the revision.

On your underlying point — AATPS is not a sufficient proxy for wall-clock throughput, and our grid demonstrates it rather than contradicting it. On the Llama pair, L=1/B=8 has AATPS 1.93 and reaches 45.35 tok/s, while L=4/B=8 has AATPS 4.04 and reaches 48.30: less than half the AATPS delivers 94% of the throughput, because the shallower configuration spends far fewer sequential drafter forwards per block. Reporting AATPS alone would hide this entirely. The revision therefore reports the full (L,B) grid with measured throughput, and names the operating point, instead of leading with an AATPS figure.

We additionally measured the component accounting this implies (L=4, 30 instrumented prompts per cell, same card, both decoders in one process):

pairBblockdraft fwdtarget fwdarrival samplingkeyed RNGcache ops
Llama483.2948.8226.861.930.347 (0.42%)2.45
Llama885.4348.5728.162.510.446 (0.52%)2.70
Qwen4101.4569.5924.051.750.359 (0.35%)2.80
Qwen899.3065.8324.762.280.476 (0.48%)3.01
pairBtree nodes/blockvisited contexts/blocktarget fwd callstarget batch
Llama47.6810.671.0010.67
Llama810.8515.841.0015.84
Qwen48.7012.001.0012.00
Qwen813.3419.281.0019.28

Verification is one batched target forward per block. Keyed clock generation costs 0.35–0.48 ms/block (0.35–0.52%), but MPFR's whole sampling path (arrival sampling + keyed RNG + logits) is 3.0–3.9 ms/block against the baseline's 1.5–1.8 ms, so it is 1.4–2.1 ms/block more expensive — we state this rather than implying the keyed path is free.

The tree cost you ask about is repaid by context merging. An unmerged list of B drafts at lookahead L expands L·B nodes; merging brings this to 7.68 of 16 at B=4 and 10.85 of 32 at B=8 on the Llama pair (52% and 66% fewer), which shrinks the batch the target must verify. At Llama B=8 both arms spend the same time in the drafter (48.57 vs 50.34 ms) while MPFR's target forward is 4.7 ms/block cheaper (28.16 vs 32.82) — more than enough to cover the sampling overhead and account for the observed gain. The saving grows with B, because duplicate prefixes get more frequent as the list widens, which is why the advantage is especially pronounced at larger BB in several measured settings (although not universally monotone).

AuthorsOpenReview ↗

Question on MPFR's better throughput and Summary

Q2: "why does MPFR have slightly higher throughput than GLS even though it additionally carries a watermark?"

Because the watermark is not where the cost is. We verified this directly with an ablation the submission lacked at B>1: an otherwise identical decoder whose keyed Poisson source is replaced by unkeyed fresh noise, with the draft/verify/cache path untouched. Across five cells (two datasets, B ∈ {2,4,8}) on the Llama pair, ΔAATPS lies in −0.01%…+0.57% and ΔTR in −0.02%…+0.71% — keyed and unkeyed are indistinguishable in efficiency — while ANLPPT-U rises by 23–90×. The same holds at B=1 (PFR 41.46 vs PFR-NoWM 41.72 tok/s, and equally at every other L).

The margin comes from the target batch, not from the sampler. As the component table above shows, MPFR's sampling path is actually 1.4–2.1 ms/block more expensive than the baseline's. What more than repays it is context merging: MPFR expands 52–66% fewer draft-tree nodes, so the single batched target forward per block runs over a smaller batch — 4.7 ms/block cheaper at Llama B=8, against an equal drafter cost.

The same mechanism says where the advantage should be small, and the data agrees: the saving scales with the number of duplicate prefixes, so it shrinks as B narrows. At B=2 there is little to merge and the two methods are within noise of each other (or the baseline is ahead); at B=8 the gap is large and highly significant — on Qwen2.5-72B the paired t-statistic for token rate is +9.94 (0.5B drafter, +11.62%) and +10.85 (1.5B drafter, +10.47%). The crossover depends on the model and workload, and the revision reports both the positive wide-list cases and the small- or negative-gain configurations.

Summary

We sincerely hope that these explanations, corrections, and new results address the reviewer's concerns. We have revised the manuscript accordingly, with changes highlighted in blue. We appreciate the reviewer's time and constructive feedback, and would be grateful if the reviewer could kindly reconsider their evaluation in light of the strengthened theoretical and empirical evidence.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Summary of Revisions and Additional Evidence

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 to each reviewer and substantially revised the manuscript using the permitted additional page, with all changes highlighted in blue. Our main revisions and additions include:

  1. Theoretical clarification and proof: We sharpened the interpretation of Theorem 5.1 and provided a rigorous multi-draft pathwise/stopping-time drafter-invariance proposition and proof. This establishes that the emitted tokens follow the same target-keyed sequence independently of the compatible drafter configuration, while the block length may change.

  2. Measured end-to-end efficiency: We reran matched MPFR/GLS comparisons on unified hardware, reported throughput as well as AATPS over lookahead and draft-count sweeps, and compared MPFR with MWS, MSE, and autoregressive watermarked decoding. The new measurements identify model-pair-dependent optimal lookahead and explain the apparent slowdown in some original configurations.

  3. Overhead and practical scaling: We added detailed measurements of draft/target computation, clock generation, context merging, target verification, and memory. Additional experiments cover 32B and 72B targets, different target–draft gaps, and decoding parameters.

  4. Watermark preservation and robustness ablations: We added a multi-draft keyed-versus-unkeyed ablation, drafter-substitution experiments varying model scale and temperature, and a synthetic context-indexed versus draft-indexed watermark study. These experiments support preservation of the watermark signal while draft count and efficiency change.

  5. Presentation and reproducibility: We clarified the single-draft relation to prior shared-Gumbel schemes, separated theoretical guarantees from measured performance, and corrected the previous hardware description. The distinct measurement protocols are now explicitly identified.

We believe these revisions substantially improve the clarity, theoretical positioning, and empirical support of the paper, and address the main concerns raised during review.

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 ↗

A Gentle Follow-up on Our Responses and Revisions

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 the initial revisions highlighted in blue and the latest revisions highlighted in red. Our main additions include:

  1. Theoretical contributions: Clarified the novelty of MPFR and its relationship to existing work, refined the interpretation of Theorem 5.1, and established a new multi-draft pathwise/stopping-time drafter-invariance proposition with a rigorous proof.
  2. End-to-end efficiency: Conducted extensive comparisons against GLS, MWS, MSE, and autoregressive watermarked decoding on matched hardware. Notably, MPFR achieves 43.7% higher throughput than MWS with comparable watermark strength on the Llama model pair.
  3. Practical scalability: Extended our evaluation to 32B and 72B target models, demonstrating up to 11.62% higher throughput than GLS, and provided detailed computational overhead and memory analyses across different decoding configurations.
  4. Watermark preservation and robustness: Conducted multi-draft keyed-versus-unkeyed ablations, drafter-substitution experiments, and context-indexed versus draft-indexed watermarking comparisons, further supporting watermark preservation without a measurable efficiency penalty from keying.

We particularly thank Reviewer v89d for the constructive follow-up, based on which we have further revised Sections 1, 4, and 5 to better distinguish our novel contributions from existing work and highlight the new empirical evidence. We would also greatly appreciate any further feedback from the other reviewers and Area Chairs, especially if any concerns remain or additional clarification would be helpful.

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 theoretical results, additional experiments, and revised presentation.

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

The Authors

AuthorsOpenReview ↗

One More Gental Follow-up

Dear Reviewers and Area Chairs,

As the discussion period approaches its end (in 1 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 the initial revisions highlighted in blue and the latest revisions highlighted in red.

We sincerely hope these revisions have addressed your concerns. We would be grateful if you could kindly reconsider your assessments based on our strengthened results, additional experiments, and substantially revised presentation.

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

The Authors

AuthorsOpenReview ↗

Summary of Contributions and Rebuttal Revisions for the Area Chair

Dear Area Chair,

As the discussion period concludes, we would like to briefly summarize our main contributions and the additional theoretical and empirical evidence provided during the rebuttal to facilitate your assessment.

1. Main Contributions and Novelty

Our central contribution is to introduce draft multiplicity as an additional dimension of the fundamental watermark–sampling efficiency trade-off identified by Hu and Huang (2024), enabling practical efficiency improvements while preserving watermark strength.

  • Novel Poisson-process-based multi-draft coupling: While our single-draft PFR construction is equivalent to existing shared-Gumbel/MWS schemes, our main contribution, MPFR (Algorithm 4), nontrivially exploits the mapping theorem and order statistics of Poisson processes to construct an exact, naturally watermarkable multi-draft coupling without communication.
  • Novel theoretical acceptance guarantee: We establish an acceptance probability lower bound (Theorem 5.1) that improves upon the corresponding bound of GLS (Rowan et al., 2025). This improvement concerns the theoretical lower bound at a fixed context, rather than universally ordering actual acceptance probabilities or throughput.
  • Context-indexed watermarking and drafter invariance: Unlike draft-indexed GLS, MPFR constructs context-indexed keyed Poisson sources, ensuring that the emitted tokens depend only on the target and keyed randomness. This preserves the target-keyed watermarked text independently of compatible drafter configurations, draft multiplicities, and lookahead lengths.

2. Main Additions and Clarifications During Rebuttal

We have substantially revised the manuscript, with the initial revisions highlighted in blue and the latest revisions highlighted in red.

  • New theoretical result: Established and rigorously proved multi-draft pathwise drafter invariance (Proposition 5.2), including an extension to compatible heterogeneous drafters, and clarified the scope of Theorem 5.1.
  • Measured end-to-end efficiency: Conducted extensive comparisons against GLS, MWS, MSE, and autoregressive watermarked decoding on matched hardware. On the Llama model pair, MPFR achieves 48.30 tokens/s, representing 43.7% higher throughput than MWS and 36.9% higher throughput than autoregressive watermarked decoding, with comparable watermark strength.
  • Scalability to larger models: Extended our evaluation to 32B and 72B target models, demonstrating up to 11.62% higher throughput than GLS. Detailed computational analyses explain how context merging reduces verification costs, with up to 66% fewer expanded tree nodes relative to an unmerged draft tree in the instrumented Llama setting.
  • Watermark preservation: Conducted multi-draft keyed-versus-unkeyed ablations, drafter-substitution experiments, and context-indexed versus draft-indexed watermarking comparisons. The keyed-versus-unkeyed experiments demonstrate no measurable efficiency penalty from keying while increasing watermark evidence by 23–90×.
  • Practical limitations and presentation: Investigated model-dependent optimal lookahead lengths, explicitly acknowledged configurations where speculative decoding is slower than autoregressive decoding, and substantially revised Sections 1, 4, and 5 to distinguish our novel contributions from existing techniques.

We particularly appreciate the reviewers' recognition of the theoretical value of our Poisson-process construction and their constructive suggestions regarding practical efficiency and the presentation of our contributions.

We respectfully emphasize that our contribution is not merely extending single-draft sampling to multiple drafts or increasing AATPS, but developing a novel coupling that translates draft multiplicity into measured wall-clock efficiency improvements while preserving target-keyed watermark strength. The new theoretical guarantees and extensive additional experiments directly support this central contribution.

We sincerely thank the Area Chair and all reviewers for their time, constructive 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.