acceptodds
Under review as a conference paper at ICLR 2027

CORE-BREW: LLR-Based Soft Decoding for Robust Multi-Bit LLM Watermarking

Abstract

Reliable provenance for LLM outputs requires multi-bit watermarks that remain robust under editing while maintaining low false-positive rates. Existing ECC-based LLM watermarks rely on hard-decision decoding, discarding token-level reliability information and limiting robustness under post-generation edits. We propose **CORE-BREW**, a **CO**nstant-hit-**R**ate **E**mbedding extension of BREW for multi-bit watermarking. CORE-BREW calibrates the watermark channel by targeting a fixed hit rate , yielding closed-form per-token log-likelihood ratios (LLRs) for soft-decision decoding. It incorporates entropy-aware erasures to limit perturbations in low-entropy contexts and combines likelihood-based scoring with soft-decision list decoding to exploit soft evidence. Experiments on open-source LLMs under token-level edits and paraphrasing demonstrate that CORE-BREW generally improves detection robustness and payload recovery over the BREW baseline while maintaining low observed false-positive rates. Despite higher conditional perplexity, BLEU and BERTScore remain close to those of unwatermarked text, indicating comparable reference-based translation quality.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
large language models, text watermarking, multi-bit watermarking, error-correcting codes, BCH codes, soft-decision decoding, log-likelihood ratios, AI safety and provenance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_R2HhOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper extends BREW’s block-wise designated-codeword watermarking framework with constant-hit-rate embedding, distortion-aware erasures, and Chase-style list decoding. At generation time, an adaptive logit offset targets a prescribed probability of sampling from the token partition corresponding to the embedded bit. At detection time, constant-magnitude log-likelihood scores, erasure handling, and shifted block searches are combined to verify designated BCH codewords. Experiments on OPT-1.3B and Mistral-7B cover clean generation, token edits, and paraphrasing, with additional analyses of payload size, text length, quality, and computational cost. The reported results show improvements over BREW in several settings, although the extent to which these improvements arise from calibrated soft information than other changes to the detector or operating point, remains unclear.

Critical strengths
  1. A coherent connection between embedding and decoding.

    The paper explicitly connects generation-time probability shaping with detection-time error correction. The constant-hit-rate identity is valid when the intended bias is applied without clipping or subsequent distribution changes. This provides a reasonable starting point for investigating how the watermark channel and decoder should be designed together.

  2. Meaningful empirical improvements in some settings.

    Several results support the practical promise of the construction. For example, under 10% substitution on OPT-1.3B/C4, Table 9 reports an increase in TPR from 0.6550 for BREW to 0.7017, with no observed false positives for either method. The evaluation also goes beyond clean detection to include paraphrasing, short texts, payload scaling, and runtime.

Critical weaknesses

My concerns affect the interpretation of the central mechanism and the comparability of the main results. Addressing them would require additional calibration diagnostics, matched detector baselines, and comparisons at controlled false-positive and quality levels, instead of only textual clarification.

  1. The implemented channel does not consistently satisfy the calibration used to interpret its scores and the soft-decoding contribution is not isolated.

    In Section 4.1 and Algorithm 1, clipping can leave a position non-erased whenever its target-list mass remains above 1/2, even if it falls substantially below the target hit rate. Appendix E.5 additionally states that calibration precedes top-p sampling with p=0.9p=0.9. Nevertheless, the detector assigns the same nominal LLR magnitude to every non-erased position. The paper acknowledges some of these qualifications, but does not quantify how often the default implementation operates outside the calibrated regime or how much this affects detection. This gap is consequential because calibrated reliability is central to the proposed explanation of the method.

    All non-erased positions have the same reliability magnitude. Consequently, for a candidate block with neffn_{\mathrm{eff}} non-erased positions and ee mismatches, the designated-codeword score is proportional to neff−2en_{\mathrm{eff}} - 2e. It therefore contains no reliability ordering among non-erased positions. Table 2 simultaneously changes erasure treatment, candidate search, and score filtering: the hard baseline fills erased positions with PRF bits, whereas the proposed detector preserves zero reliability. This does not isolate the benefit of calibrated soft information. A BCH errors-and-erasures baseline and a matched candidate-budget baseline are needed to determine whether the improvement primarily comes from erasure handling or a larger search region.

  2. The main comparisons do not establish an advantage at matched false-positive and quality levels and the incremental contribution and payload-recovery protocol require clearer positioning.

    Table 1 evaluates methods at their default or recommended thresholds, with very high FPRs for MPAC and Qu et al. This does not establish their relative performance under a common low-FPR requirement. Some improvements over BREW also accompany higher FPR: under OPT/C4 insertion, Table 11 reports TPR increasing from 0.4567 to 0.5250 while FPR increases from 0 to 0.0083. Quality comparisons are similarly difficult to interpret jointly: Table 6 reports PPL on OPT/C4 but BLEU and BERTScore on NLLB/WMT16 translation. The latter cannot establish quality preservation in the former setting, where PPL rises from 8.454 without watermarking to 15.306. Comparisons at matched FPR and matched quality are necessary to establish the claimed trade-off improvement.

    BREW (Kim et al., 2026) already provides the relationship between target-list probability and logit bias in Appendix C.5, from which the constant-hit-rate offset follows directly. Its public description also includes blind message estimation, whereas the submitted CORE-BREW detector takes designated codewords as input. The paper should clarify what is newly contributed beyond applying this relationship with erasures and list decoding, and whether its payload-recovery results are obtained without access to the true payload. In addition, the stated lack of a public implementation for Chao et al. should be reconciled with the currently available official GitHub repository (Chao et al.), with attention to version and release timing.

Questions for authors
  1. Calibration under default sampling settings.

    Under the default clipping and top-p settings, what fractions of positions are fully calibrated, clipped but retained, and erased? What are their realized hit rates, and does the improvement persist when using scores consistent with the actual sampling distribution?

  2. Comparison with matched decoding baselines.

    How does CORE-BREW compare with errors-and-erasures BCH decoding and a baseline using the same number of candidate trials, on identical generated texts and at matched FPR? These comparisons would materially affect whether I attribute the gain to soft reliability.

  3. Evaluation at matched FPR and generation quality.

    Can the authors provide comparisons at a common FPR and comparable generation quality, together with the number of independent null samples and confidence intervals? For the translation quality results, what detection performance and effective embedding rate are achieved on those same translations?

  4. Blind payload extraction and recovery.

    Is the true payload available to any part of the evaluated extraction procedure, including erasure construction or designated-codeword selection? If not, what is the complete blind extraction procedure and its message-level exact recovery rate?

3 comments
AuthorsOpenReview ↗

Response to Reviewer R2Hh (1/2)

Thank you for the detailed questions. We report the additional measurements below, keeping three things distinct: nominal calibration, detector operating points, and verification-conditioned recovery.

Q1 — Calibration and realized reliability. The target-mass identity holds before clipping and the subsequent sampling transforms. Our additional OPT-1.3B/C4 diagnostic reports these generation-time fractions and hit rates:

Position categoryFractionObserved hit rate [95% CI]
Nominally calibrated, unclipped, non-erased88.2%0.873 [0.863, 0.883]
Clipped but retained0.5%0.756 [0.725, 0.784]
Erased11.3%0.551 [0.544, 0.557]

Clipped-but-retained positions are rare: 99.4% of non-erased positions are unclipped. Their hit rate nonetheless shows that a retained position need not reach p∗=0.9p^*=0.9. The table separates the nominal target from the realized hit frequency, but it does not isolate the causal effect of each sampling transform. With this diagnostic we will report the token/text counts, interval construction, sampling order and safeguard defaults, and reconcile Appendix E.5 with the corresponding run configuration.

The implemented ±ln⁡9\pm\ln 9 values are therefore nominal scores. An exact bit LLR requires the transition probabilities under both hypotheses, log⁡[P(Y=y∣z=1)/P(Y=y∣z=0)]\log\big[P(Y=y\mid z=1)/P(Y=y\mid z=0)\big].

One replacement can be settled exactly. Since 63−f−2e63-f-2e is an integer, the default gate τ=5\tau=5 accepts iff 63−f−2e≥363-f-2e\ge 3, for every homogeneous scale λ∈[5/3,,5/2)\lambda\in[5/3,\\,5/2). That range corresponds to hit rates p∈[σ(5/3),,σ(5/2))≈[0.8411,,0.9241)p\in\big[\sigma(5/3),\\,\sigma(5/2)\big)\approx[0.8411,\\,0.9241), where σ(x)=1/(1+e−x)\sigma(x)=1/(1+e^{-x}). Any positive scale also leaves the Chase ordering unchanged. Both p∗=0.9p^*=0.9 and the realized 0.8730.873 (λ^=ln⁡(0.873/0.127)≈1.93\hat\lambda=\ln(0.873/0.127)\approx 1.93) lie in this range, so replacing the nominal scale by the realized homogeneous one changes no decision. Position-specific two-hypothesis scores would define a different detector, and we will evaluate them in the revision.

Q2 — Additional decoding baselines on identical texts. The General Response compares separately calibrated configurations on the same 371 positives and 1,800 held-out nulls, with calibration nulls disjoint from the test nulls.

  • Same nominal trial cap. Four PRF fills match order-2 Chase's maximum of four trial assignments per block/shift, but not its candidate vectors or total computation. Against this baseline, CORE-BREW detects 200 vs. 179 insertion positives (false positives 7 vs. 6) and 257 vs. 280 substitution positives (2 vs. 12 false positives).
  • Errors-and-erasures (E&E). E&E reaches 76.01%/0.61% under substitution and 53.10%/0.39% under insertion. CORE-BREW's net insertion difference from E&E is three detections at the same false-positive count.
  • Score-only verification has the same aggregate rates as the E&E row. Analytically, the two predicates coincide at τ/λ∈(32,33]\tau/\lambda\in(32,33].

These rows are separately calibrated configurations without a paired uncertainty analysis. They are therefore neither single-factor nor matched-FPR comparisons, and none relies on graded non-erased reliability. Within that limit, both configurations that treat erasures explicitly detect more insertion positives than four PRF fills (197 and 200 vs. 179 of 371) at a similar false-positive count (7 vs. 6 of 1,800). This is consistent with a benefit from erasure handling, but it does not isolate one. A gate-only ablation with all other settings fixed can only remove detections (AChase+gate⊆AChaseA_{\mathrm{Chase+gate}}\subseteq A_{\mathrm{Chase}}); we will report it in the revision.

AuthorsOpenReview ↗

Response to Reviewer R2Hh (2/2)

Q3 — FPR, independent samples and quality. The main tables contain 600 null trials per setting. If these trials are independent, zero false positives gives a 95% two-sided Clopper-Pearson upper endpoint of 0.613%. We will specify the source/prompt/seed structure when reporting uncertainty; pooling attacks that reuse null sources does not create independent samples.

  • Substitution: zero observed false positives for both methods is equal observed FPR, not equal population FPR.
  • OPT/C4 insertion: TPR rises from 45.67% to 52.50% while FPR rises from 0% to 0.83%, so we claim no matched-FPR advantage at this default-threshold point.
  • Table 1: likewise does not rank MPAC, Qu et al. and CORE-BREW at low FPR. In the revision we will select thresholds on calibration nulls, freeze them, and evaluate on disjoint test nulls, reporting counts and intervals.

On OPT/C4, PPL is 8.454 unwatermarked, 13.643 for BREW and 15.306 for CORE-BREW. The safeguards reduce PPL from 18.631 to 15.306 (17.85%). This is a safeguard benefit, not a matched-quality advantage over BREW. Detector thresholds cannot change PPL on fixed text, so we will add an embedding-strength sweep for matched-PPL comparison.

NLLB/WMT16 BLEU and BERTScore come from a separate translation run that logged only reference-quality metrics without block-detection or erasure outputs, and we will remove the inference that those cross-model scores show quality preservation for OPT/C4 detection. In the revision we will instead evaluate quality and detection jointly on the same watermarked continuations, using the embedding tokenizer and block schedule.

Q4 — Targeted verification vs. blind payload extraction. In the submitted experiments, yes: as documented in Algorithm 3 and App. F.1, our evaluated pipeline implements targeted provenance verification of a hypothesized payload rather than blind decoding. Specifically, Algorithm 3 takes the candidate payload's designated codewords c(j)\\{c^{(j)}\\} as an explicit input, and observation construction also uses them in the feasibility erasure check. Because we evaluated this verifier at the embedded payload, the reported metric is verification-conditioned: an accepted designated chunk is known by construction, so crediting it does not measure unknown-message recovery. Calling this "payload recovery" obscured that distinction; we will relabel the metric and state its exact accounting rule.

Blind search also carries a multiple-testing cost that the per-hypothesis analysis does not cover. Under an ideal null with independent fair bits, accepting any of the 128 BCH(63,7) codewords within radius 15 has per-block/shift probability exactly 128×1.8803×10−5≈2.41×10−3128\times1.8803\times10^{-5}\approx2.41\times10^{-3} for hard decoding (the radius-15 balls are disjoint because dmin⁡=31d_{\min}=31), and at most 128×4.4188×10−5≈5.66×10−3128\times4.4188\times10^{-5}\approx5.66\times10^{-3} for the ungated two-position list, by a union bound with payload-independent selected positions. Blind thresholds must therefore be recalibrated, not inherited from targeted verification. These are ideal-null values, not natural-text FPRs.

Neither the submission nor this response reports blind-extraction numbers; the blind evaluation we will add in the revision uses the following protocol.

  • The detector receives the text, the key and the scoring model, without the payload or designated codewords.
  • The message is sampled independently of the key; key-dependent placement or permutation is allowed.
  • Erasures use only payload-independent conditions, Ht<Hmin⁡H_t < H_{\min} or max⁡vpt(v)≥qmax⁡\max_v p_t(v)\ge q_{\max}.
  • For each block and shift, the full codebook is searched, and decoded chunks are combined by the block schedule.
  • Thresholds are selected on calibration nulls and frozen before evaluation on disjoint test nulls.
  • We will report exact-message success over all positives, counting abstentions as failures, together with the abstention rate.

Positioning and corrections. BREW already provides the logit-bias relationship and a blind estimation procedure. Our extension is the target-mass/safeguard configuration coupled to model-aware erasure handling and bounded verification. Chao et al.'s official repository is public; we will correct Section 6.1 and include an empirical comparison under our matched low-FPR protocol in the revision (not yet completed here).

The diagnostic quantifies how far realized reliability departs from nominal and shows that a homogeneous correction leaves the default decisions unchanged. The same-text results add E&E and four-trial comparators. The matched-FPR, matched-quality and blind evaluations described above will be reported separately in the revision.

Thank you for the candid response and additional measurements. While the clarifications resolve several ambiguities regarding nominal calibration, detector operating points, and verification-conditioned recovery, they also confirm that central claims in the submission are not supported by the evaluated protocol. I therefore maintain my recommendation of clear rejection (1).

1. The reported “payload recovery” evaluates known-payload verification, not unknown-message extraction. The response confirms that the detector receives the designated codewords corresponding to the true embedded payload, which are also used in the feasibility-based erasure check. Crediting an accepted designated chunk therefore measures verification conditioned on a known payload, as opposed to recovering an unknown message. While targeted provenance verification is a valid task, it represents a substantially narrower contribution than the payload-recovery claims in the paper, and comparing it directly against baselines that must blindly extract the payload creates an unequal evaluation. Relabeling the metric does not supply evidence for blind recovery, and the proposed blind-extraction protocol remains an unexecuted plan with no reported results.

2. The new baselines show no clear gain from the proposed soft-decoding machinery. Under insertion, CORE-BREW detects 200/371 positives compared with 197/371 for simple errors-and-erasures decoding, with an identical 7/1,800 false positives. A three-sample difference without paired uncertainty analysis does not establish a meaningful advantage. Under substitution, CORE-BREW yields both lower TPR and lower FPR than the added baselines, leaving the trade-off inconclusive. Moreover, the score-only verifier matches the aggregate rates of the errors-and-erasures baseline and is analytically equivalent over the stated threshold range. Thus, while explicitly handling erasures is beneficial, the core algorithmic additions—calibrated LLR scoring and Chase-style list decoding—demonstrate no clear improvement over simpler erasure-aware alternatives.

3. Homogeneous-scale invariance underscores the limited role of calibration. The observation that replacing the nominal homogeneous scale with the realized scale leaves default decisions unchanged addresses the scalar-replacement question, but it also highlights that all non-erased positions share a uniform magnitude. Because the detector does not differentiate per-position reliability, this invariance does not show that calibrated reliability drives the observed gains. Furthermore, nominally calibrated, unclipped positions achieve a hit rate of 0.873 (whose confidence interval excludes the 0.9 target), showing that clipping alone does not explain the calibration gap, while position-specific two-hypothesis scores remain untested. Consequently, the evidence supports only a nominal-score verifier with erasure handling, leaving the claimed benefit of calibrated soft information unsubstantiated.

4. Principal comparisons at matched FPR and matched quality remain incomplete. The response acknowledges that the default-threshold insertion results do not represent a matched-FPR advantage, and that low-FPR rankings against MPAC and Qu et al. are not established; the proposed independently calibrated evaluations have not been conducted. Similarly, CORE-BREW’s PPL on OPT/C4 remains worse than BREW’s, with the embedding-strength sweep needed for a matched-quality comparison deferred to future work. Meanwhile, the NLLB/WMT16 experiment reports quality metrics without corresponding detection or erasure measurements, leaving the joint quality–detection trade-off unsupported across both setups.

Overall, while the rebuttal improves transparency, it substantially shrinks the demonstrated contribution: the actual task is targeted verification, the complex soft-decoding machinery shows no clear edge over basic erasure handling, and the critical evaluations—blind extraction, matched-FPR detection, and matched-quality comparisons—remain prospective plans rather than realized evidence. Accordingly, the empirical support is insufficient to warrant acceptance.

Reviewer_wD98OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CORE-BREW, extending BREW with Constant Hit-Rate embedding, entropy-aware erasures, and LLR-based soft-decision decoding to improve multi-bit watermark robustness against text editing and paraphrasing. Experiments across multiple models, datasets, and attack settings demonstrate improvements in detection robustness.

Critical strengths
  1. Clear methodological presentation with a well-organized embedding and detection pipeline.
  2. Comprehensive experimental evaluation across multiple models, datasets, and attack settings, demonstrating improvements over BREW in several scenarios.
Critical weaknesses
  1. Limited justification of soft-information benefits. The paper motivates soft decoding by arguing that hard decisions discard token-level reliability. However, all non-erased positions receive identical absolute LLR values. Thus, the method does not exploit fine-grained token-level reliability, and the reported gains may primarily come from erasure handling and Chase-style candidate search rather than genuinely informative soft decoding.
  2. Insufficient ablation and controlled comparisons. The hard-decision ablation simultaneously removes LLR scoring and Chase-style decoding while replacing erased positions with PRF-generated bits. This makes it difficult to isolate the actual contribution of soft decoding.
Questions for authors

Given that non-erased positions have identical LLR magnitudes, what additional reliability information does the proposed soft decoder exploit beyond erasure indicators? Can the authors separately evaluate LLR scoring, Chase-style decoding, and erasure-aware hard decoding, and report comparisons at matched FPR?

1 comment
AuthorsOpenReview ↗

Response to Reviewer wD98

Thank you for the precise questions. We agree that our original interpretation attributed too much to soft reliability. The decision-rule analysis and the additional measurements below support a more specific account.

Q1 — Reliability information beyond erasure indicators. The nominal decoder uses a ternary observation—the keyed bit sign (±λ\pm\lambda) plus the erasure indicator (00)—with no graded reliability ordering among non-erased positions, since all non-erased magnitudes equal λ=ln⁡9\lambda=\ln 9. A binary symmetric erasure channel is an idealized probabilistic model for this ternary rule, not an exact description of the implementation.

For a designated codeword, S=λ(63−f−2e)S=\lambda(63-f-2e), where ff counts erasures and ee counts non-erased disagreements. At τ/λ∈(32,33]\tau/\lambda\in(32,33], score-only verification is exactly the bounded errors-and-erasures (E&E) criterion 2e+f≤302e+f\le 30. In our comparison, score-only verification has the same aggregate rates as the E&E row of the General Response (GR §1) table. Showing that the sample-wise decisions are also identical would require the recorded thresholds and predictions.

Chase adds bounded candidate search. Suppose FF holds the two selected positions and all four assignments are tested. A designated radius-15 candidate then exists iff dH(BFc,cFc)≤15d_H(B_{F^c},c_{F^c})\le 15. Erasures outside FF keep their raw token bits in BB, even though they contribute zero to the score. With erasures selected first, every designated list match satisfies

S≥{29λ,f=0,30λ,f=1,(33−f)λ,f≥2.S\ge\begin{cases}29\lambda, & f=0,\\\\ 30\lambda, & f=1,\\\\ (33-f)\lambda, & f\ge 2.\end{cases}

So at λ=ln⁡9\lambda=\ln 9 and τ=5\tau=5, the default gate cannot reject a match when f≤30f\le 30, since 3ln⁡9≈6.59>53\ln 9\approx 6.59>5. This is a mathematical condition; we do not infer the gate's empirical rejection frequency from the mean erasure count.

Q2 — Separate components and compare at matched FPR. GR §1 adds operating points on identical texts: 371 positives and 1,800 held-out nulls, with disjoint calibration nulls. Entries are TPR/FPR.

  • Erasure-aware verification vs. multiple hard completions. E&E gives 76.01%/0.61% (Sub) and 53.10%/0.39% (Ins). Four PRF fills give 75.47%/0.67% and 48.25%/0.33%. On insertion, E&E has a net 18 more detections with one more false positive. These are different decision procedures, not an equal-compute comparison.
  • Same nominal trial cap. Four PRF fills match order-2 Chase's maximum of four trial assignments per block/shift; candidate vectors and total computation are not matched. Full CORE-BREW gives 69.27%/0.11% (Sub) and 53.91%/0.39% (Ins). Relative to four PRF fills, that is 21 more insertion detections with one more false positive, and 23 fewer substitution detections with ten fewer false positives.
  • LLR scoring matches the E&E row in aggregate, as noted above.

These rows are separately calibrated configurations without a paired uncertainty analysis. They are therefore neither single-factor nor matched-FPR comparisons, and none relies on graded non-erased reliability. They do make E&E an essential comparator. Both configurations that treat erasures explicitly detect more insertion positives than four PRF fills (197 and 200 vs. 179) at a similar false-positive count (7 vs. 6). This is consistent with a benefit from erasure handling, but it does not isolate one.

Gate-only ablation. Fix the observations, erasures, Chase candidates, shifts and text aggregation. Then AChase+gate⊆AChaseA_{\mathrm{Chase+gate}}\subseteq A_{\mathrm{Chase}}, so toggling only the gate cannot raise TPR or FPR. The comparison above is not that ablation, because rows were calibrated separately. We will report the nested gate-only toggle in the revision, together with held-out counts and intervals, since a shared calibration target does not ensure matched held-out FPR.

Revision. We will keep these operating points and give their full configurations. We will replace the soft-reliability attribution with the analysis above, and add the gate-only ablation. The evidence supports examining erasure-aware designated verification directly. It does not support a claim that constant-magnitude scores supply graded token confidence.

Reviewer_JpD2OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CORE-BREW, a multi-bit LLM watermarking framework built upon BREW. The method introduces constant hit-rate embedding, entropy-aware erasures, and LLR-based soft-decision decoding to improve robustness against token-level editing and paraphrasing attacks. The experimental results show improvements over BREW in several settings while maintaining low observed false-positive rates.

Critical strengths
  1. The paper addresses an important problem in robust multi-bit LLM watermarking, particularly under text editing and paraphrasing.

  2. The combination of channel calibration, entropy-aware erasures, and soft-decision decoding is practically motivated, and the experimental results show improvements over the original BREW framework.

Critical weaknesses
  1. My main concern is the limited novelty of the proposed method. The central idea is to improve watermark detection by incorporating token-level reliability information into the decoding procedure, rather than relying entirely on hard decisions. However, soft-decision decoding and LLR-based reliability estimation are well-established in coding theory, and recent work on multi-bit LLM watermarking has already explored alternative encoding, extraction, and reliability-improvement strategies. In Section 4, CORE-BREW largely combines constant hit-rate calibration (Equations (1)–(2)), LLR computation (Section 4.2), and Chase-style list decoding (Section 4.4) within the existing BREW framework. These components are individually based on established techniques. While their integration may be useful, the paper does not sufficiently explain which aspects constitute a genuinely new methodological contribution or why this particular combination represents a substantial advance.

  2. Section 5 presents a theoretical analysis of the watermark channel and false-positive behavior, with additional derivations in Appendix C. However, it is not entirely clear what new theoretical insights these results provide or how they explain the empirical improvements. For example, the constant hit-rate property follows from the adaptive logit calibration in Equation (1), while the LLR formulation in Section 4.2 is derived under an approximately binary symmetric channel model. These results appear to rely primarily on standard probability and coding-theoretic arguments. More importantly, the connection between this theoretical model and the actual robustness of the proposed method is not sufficiently established. The authors acknowledge in Section 4.2 that, under attacks, the computed LLRs represent nominal pre-attack reliability rather than exact post-attack likelihoods. Since robustness against token-level editing and paraphrasing is one of the main objectives, this raises a question about how well the theoretical model characterizes the settings in which the method is evaluated. Similarly, the false-positive analysis in Section 5 and Appendix C relies on assumptions that may not fully reflect practical language-model outputs and dependencies among shifted decoding candidates. Although such simplifying assumptions can be reasonable, the paper should more clearly explain their implications and limitations.

  3. I also find Figure 1 somewhat confusing. While it provides a high-level overview of CORE-BREW, several important steps described are not clearly represented. In particular, the relationship among the observed text, the designated payload/codeword, LLR computation, and block-wise soft decoding is difficult to follow.

  4. Although Section 6 reports improvements over BREW, the experimental evaluation does not sufficiently cover recent multi-bit watermarking approaches. The main baselines are MPAC, Qu et al., and BREW. Other relevant multi-bit methods are not included in the comparison. In addition, the reported gains from soft decoding are relatively modest in some settings. For example, Table 2 shows an improvement in TPR from 0.9067 to 0.9383 under clean generation and from 0.6617 to 0.7017 under 10% substitution, compared with the matched hard-decision variant. These results support the usefulness of soft decoding, but do not by themselves establish a substantial methodological contribution.

Questions for authors
  1. What is the key methodological novelty of CORE-BREW beyond combining established constant hit-rate calibration, LLR-based scoring, and Chase-style soft-decision decoding within BREW?

  2. What new theoretical insights are provided by Section 5 and Appendix C? In particular, how do the theoretical results explain robustness improvements under editing and paraphrasing when the LLRs are calibrated for the pre-attack channel?

  3. Can the authors provide stronger comparisons with recent multi-bit watermarking methods

1 comment
AuthorsOpenReview ↗

Response to Reviewer JpD2

Thank you for asking us to distinguish the proposed design from its established components. We will make that distinction explicit and narrow the soft-reliability claim.

W1/Q1 — Methodological contribution. BREW already provides the logit-bias/target-list relationship and designated verification, and Chase and errors-and-erasures (E&E) decoding are established. Our extension couples target-mass embedding and distortion safeguards to model-aware erasure handling and bounded designated verification. The contribution is this specific design and its measured trade-offs, not a new coding algorithm.

The submission and the new same-text comparison (General Response, GR §1) give three measurements:

  • Safeguards. They lower OPT/C4 PPL from 18.631 to 15.306 (17.85%), relative to the configuration without them.
  • Erasure handling. On insertion, E&E reaches 53.10% TPR / 0.39% FPR, versus 48.25% / 0.33% for four PRF fills.
  • Same nominal trial cap. Four PRF fills share order-2 Chase's cap of four trial assignments per block/shift, but not its candidates or total computation. Against them, CORE-BREW detects 21 more insertion positives with one more false positive, and is more conservative under substitution.

These configuration-level operating points delimit the trade-offs. They are not causal or matched-FPR results, and none relies on graded reliability.

W2/Q2 — What the theory explains. The nominal score is exactly S=λ(63−f−2e)S=\lambda(63-f-2e), a weighted error/erasure count. At τ/λ∈(32,33]\tau/\lambda\in(32,33] it equals bounded E&E verification, 2e+f≤302e+f\le 30, so a constant magnitude cannot create a graded reliability ranking.

For order-2 Chase with selected positions FF, testing all four assignments yields a designated radius-15 candidate exactly when dH(BFc,cFc)≤15d_H(B_{F^c},c_{F^c})\le 15. One assignment makes both selected bits agree with the target, and no assignment changes the other positions. The full detector intersects this event with the score gate. We offer this as an explanation and an implementation diagnostic, not as a new decoding theorem.

Under an ideal uniform-bit null for a fixed target, assume the retained bits are independent and fair given FF. The ungated list probability is then 4.4188×10−54.4188\times10^{-5}, versus 1.8803×10−51.8803\times10^{-5} for raw-bit hard verification, a 2.35×2.35\times expansion.

The default-gate conclusion is conditional. Fix the raw token bits BB, the target cc and the tested shift. Let H={dH(B,c)≤15}H=\{d_H(B,c)\le 15\} be raw-bit hard acceptance (not the PRF-filled ablation), AA full acceptance, and Pr⁡0\Pr_0 the null probability. With erasures flipped first, λ=ln⁡9\lambda=\ln 9 and τ=5\tau=5,

H∩{f≤30}⊆A,Pr⁡0(A)≥Pr⁡0(H)−Pr⁡0(H∩{f>30}).H\cap\{f\le 30\}\subseteq A,\qquad \Pr_{0}(A)\ge \Pr_{0}(H)-\Pr_{0}\big(H\cap\{f>30\}\big).

This inclusion needs no fair-bit assumption. Without controlling f>30f>30, however, it gives no unconditional ordering, and it does not compare BREW's different representation. Repeated tokens, data-dependent erasures and the actual number of blocks and shifts tested also affect text-level FPR.

The analysis explains how extra designated matches arise and what they ideally cost. It gives neither an exact post-attack likelihood model nor a paraphrase guarantee. Scaling λ\lambda and τ\tau by the same positive factor preserves every nominal decision. Scaling λ\lambda alone preserves the Chase ordering but can change the gate. These statements will replace the operational interpretation of Proposition C.7.

W3 — Figure 1. The new figure will show:

  • model computation on the observed prefix and the payload-dependent feasibility check;
  • hard bits, erasure flags and block/shift candidates;
  • the Chase search and designated verification, with the payload under test as an explicit input.

It will also separate verification from blind extraction.

W4/Q3 — Evidence and baselines. Table 9 gives substitution TPR gains over BREW of 4.67, 6.34 and 4.16 percentage points, and a 0.83-point decrease in the fourth setting. Both methods have zero observed false positives there, which does not establish equal population FPR. Insertion and paraphrase gains at default thresholds are likewise not matched-FPR results.

Chao et al.'s official repository is public; we will correct Section 6.1 and include an empirical comparison with Chao et al. under our matched low-FPR protocol in the revision (not yet completed for this response).

The revised contribution is a concrete safeguard/verification extension, supported by measured operating points and a transparent decision rule. In the revision we will add disjoint-null calibrated low-FPR operating points for all evaluated methods and a matched-PPL embedding-strength sweep, reported separately from the present default-threshold results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.