acceptodds
Under review as a conference paper at ICLR 2027

MuxMark: Multi-Channel Bit Multiplexing for LLM Watermarking

Abstract

Multi-bit LLM watermarking embeds recoverable messages for text provenance. With a fixed token budget and per-token bit load, larger payloads reduce coverage per message bit. Increasing this load improves coverage, but joint encoding based on disjoint vocabulary partitions concentrates the bias on exponentially smaller target regions, potentially weakening recovery signals at a fixed bias strength. We propose MuxMark, a multi-channel bit multiplexing framework that assigns selected message positions to separately keyed binary channels. During generation, channel contributions are summed to form a token-level logit bias that increases with the number of matched target bits. During decoding, the channels are reconstructed from the secret key and observed prefix, and weighted token votes are aggregated to recover the message bit by bit. Experiments across payloads, generation lengths, model families, and text perturbations demonstrate improved message recovery over existing multi-bit watermarking methods while maintaining comparable text quality, with particularly strong gains when recovering long messages from short texts.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Watermarking, text watermarking

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_beUXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • Through multi-channel bit multiplexing (MuxMark), each token is designed to provide independent evidence for multiple message locations.
  • At generation, channel-specific logit biases are summed to form a token-level logit bias, and during decoding, bits are recovered through channel-wise voting.
Critical strengths
  • The paper is generally well written.
  • The topic of robust multi-bit watermarking for LLM-generated text is interesting and important.
Critical weaknesses
  1. The proposed MuxMark method does not demonstrate substantial performance improvements over the baselines, making it unclear why MuxMark should be preferred over existing methods. In particular, although MuxMark achieves higher Exact Accuracy than the baselines, the absolute values remain relatively low, making it difficult to justify its practical usefulness based on this metric alone.
  2. The results presented in Table 1 under clean conditions seem largely expected. Moreover, the Exact Accuracy remains relatively low even in the absence of attacks, raising concerns about the effectiveness of the proposed method.
  3. In a previous study [1], the TPR under the DIPPER paraphrasing attack was approximately 50%, despite a relatively high AUC. This suggests that AUC alone may not sufficiently characterize watermark detection performance. Therefore, reporting additional evaluation metrics alongside AUC would provide a more comprehensive assessment of the proposed method.

[1] AliMark: Enhancing Robustness of Sentence-Level Watermarking Against Text Paraphrasing

Questions for authors
  1. Could the authors provide a more detailed justification for choosing MuxMark over existing methods? In particular, could they provide specific scenarios or use cases in which MuxMark offers the greatest advantages?
  2. Although MuxMark achieves higher Exact Accuracy than other methods, it is difficult to determine whether this improvement is sufficient for practical applications. Could the authors clarify what level of Exact Accuracy would be considered acceptable for practical payload recovery and provide specific application scenarios or criteria to justify the practical significance of the reported results?
  3. Could the authors additionally report Exact Accuracy under attack conditions? Furthermore, it would be helpful to include TPR at 1% FPR (TPR@FPR=1%) to better assess detection performance under a low false positive rate.
Reviewer_HYxgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MuxMark, a multi-bit LLM watermarking method designed to improve message recovery when the payload is long and the generated text is short. MuxMark assigns each selected bit to a separate binary channel. The biases from these channels are combined during token generation, so one token can provide evidence for multiple message bits. During decoding, the same channels are reconstructed from the secret key and token prefix, and weighted votes are aggregated to recover each bit independently. Experiments on two LLMs show that MuxMark generally improves bit recovery over several recent baselines, especially under limited token budgets, while maintaining comparable text quality and reasonable robustness to text editing, copy-paste, translation, and paraphrasing attacks.

Critical strengths
  1. The paper targets a timely and important problem. The proposed MuxMark is a distortion-based multi-bit watermarking method with a technically sound design.

  2. The experimental results show that MuxMark achieves strong decoding accuracy across many settings and generally outperforms recent baselines, especially under limited token budgets.

Critical weaknesses
  1. MuxMark applies watermark bias only to the top-10 candidate tokens. This design (only perturbing top-K tokens) is applied without explanation, and this may reduce the diversity of the LLM's response. Although the appendix studies the effect of different top-K values, the fairness of the comparison with other baselines remains unclear. Restricting the bias to high-probability tokens may itself improve text quality, while the baselines appear to follow their original implementations without the same restriction. This makes it difficult to separate the benefit of multi-channel multiplexing from the benefit of top-K biasing.

  2. The motivation for pseudo-randomly selecting both the number of injected bits and their positions is not sufficiently clear. The manuscript does not fully explain why this strategy is preferable to conventional block-based allocation, where each token also carries multiple bits. In addition, both the bit positions and channel assignments depend on the token prefix. Text edits may therefore change the reconstructed assignments during decoding. The paper reports strong robustness under editing and paraphrasing attacks, but the reason for this robustness is not well explained.

Questions for authors
  1. Could the authors explain why MuxMark remains robust to text editing and paraphrasing attacks? Decoding requires reconstructing both the selected bit positions and the corresponding channel assignments from the token prefix. When the prefix is modified, these assignments may also change and introduce incorrect votes. However, the reported DIPPER paraphrasing results show only a modest decrease in Bit Accuracy. What mechanism makes MuxMark robust to this?

  2. Figure 3 is difficult to read. The caption states that watermark strength is varied for methods with tunable strength. However, StealthInk and BiMark are distortion-free methods and how does δ\delta apply to them? Also, since MuxMark is a distortion-based method while StealthInk and BiMark are designed to preserve the original sampling distribution (distortion-free), the surprisingly favorable PPL of MuxMark deserves more discussion.

  3. How much of MuxMark's text-quality advantage comes from restricting watermark bias to the top-K candidate tokens? Why does this design make sense and seem reasonable? Will this design reduce the diversity of the LLM's response?

Reviewer_Acf3OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces MuxMark, a multibit watermarking scheme that can encode multiple bits in a single token more efficiently than prior works. For each bit in the message, MuxMark independently pseudo-randomly partitions the vocabulary into a 0 set and a 1 set, and boosts the tokens from the set corresponding to the message bit. To encode multiple bit in a single token, MuxMark sums each bit bias and normalize it. For detection, MuxMark relies on majority voting to detect the message. The results suggest than MuxMark has a better detectability-quality trade-off than prior works.

Critical strengths
  • The paper is well-written and easy to follow. The method is well explained and illustrated, with all key components in the main part of the paper.
Critical weaknesses
  • The proposed method appears to be very similar to a prior work [1] where they also sum bits contribution from independent binary channels.
  • The evaluation misses several important ablations (or interpretation) to understand the design choices of the method. The number of samples is very small (200) and only text completion is evaluated whereas relevant LLM usage is with chat template (and most recent watermarking works now evaluate watermarks with chat tasks and not completion tasks).
    • The number of bits per token is stochastic and uniform with an upper bound kmaxk_{max}. Table 4 suggests it might work better than a fix amount. Why is this the case is unclear, could a smaller/higher fix amount work better, and could another distribution rather than uniform work better. It is also unclear how all those parameters scale with the payload length.
    • The choice of wtw_{t} in equation 9 is not ablated.
    • Table 1 results are hard to relate given that the quality is not normalized across the scheme. Only Figure 3 allows to rigorously compare schemes, but at a fixed payload size.

Minor weaknesses

  • Most of the figures and some tables are so small that they are simply unreadable in the printed manuscript.
  • For Figure 3, 6 and 7 I believe it is not indicated which payload size is evaluated.
  • Zerobit detection uses an empirical calibration but could instead use 2 sided binomial/z-tests. The choice of opting for empirical calibration is not-well justified.
  • The proposed scheme is not distortion-free whereas all (zerobit) watermarking schemes used in the industry are distortion-free (i.e. SynthID, TextGrain and TextSeal).

[1] Every Bit, Everywhere, All at Once: A Binomial Multibit LLM Watermark, Gloaguen et al., NeurIPS 2026

Questions for authors
  • MuxMark uses a stochastic number of bits per token with a fixed upper bound. Could the authors better motivate this choice? It seems very unclear to me why this would be better than a fixed choice, and what drives the value of kmaxk_{max}.
  • Given the similarity to [1], what are the novel contributions of MuxMark? Does it outperform [1]?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.