acceptodds
Under review as a conference paper at ICLR 2027

ClockMark: Adaptive and Host-Dependent Watermarking for Large Language Models

Abstract

Large language model (LLM) watermarking operates on the next-token distribution that evolves during generation. In streaming generation, each sampled token becomes part of the text and affects subsequent distributions. Embedding therefore proceeds with a changing host, unlike classical watermarking, which adapts embedding to an existing host signal. To accommodate this difference, we formulate a novel embedding framework for LLM watermarking based on two principles. We first argue that embedding should be adaptive, retaining the state of previous sampling decisions to better guide subsequent choices. We further argue that embedding should be host-dependent, using the current LLM distribution to constrain both token selection and state updates. Guided by these principles and the two requirements of distribution preservation and robustness to text editing, we propose ClockMark. It combines adaptive and host-dependent embedding with message spreading, embedding the full payload in each watermarked token. Theoretically, we prove that ClockMark preserves the original LLM sequence distribution, thereby maintaining generation quality. Empirically, extensive experiments against 12 baselines demonstrate improved robustness to text editing, with gains of up to 16.08 percentage points in bit accuracy over the strongest tested baseline.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Text Watermarking, Multi-bit Watermarking

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_5dvSOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies a watermarking method named ClockMark. For each one token context cc, each candidate token vv has a countdown Rc,vR_{c,v}, which decreases at rate pt(v)p_t(v). The first countdown to reach zero determines the next token: xt=arg⁡min⁡v:pt(v)>0Rc,vpt(v).x_t=\arg\min_{v:p_t(v)>0}\frac{R_{c,v}}{p_t(v)}. The selected pair then receives a fresh interval, while the others retain their updated remaining values and resume when the context reappear. The watermark decoder calculates the sum of the vectors reconstructed from distinct adjacent token pairs and recovers each message bit from the sign of the corresponding component.

The authors provide theoretical arguments and conduct experiments on Qwen2.5-7B-Instruct and Gemma-2-9B. Robustness is evaluated against deletion, substitution, and paraphrasing.

Critical strengths

The proposed method is intuitive and easy to understand.

The experiments seem to show good robustness to deletion, substitution, and paraphrasing compared with the baseline. The method in this paper is supported with both theoretical analysis of distribution preservation and false-positive control, and experimental results.

Critical weaknesses

For the robustness evaluation, the authors should also consider cropping attacks, for example by removing the first 10%, 30%, or 50% of the generated text and evaluating message recovery and watermark detection on the remaining text. This additional evaluation is needed because the detector in the current paper uses only the initial vector zc,v,0z_{c,v,0} for each distinct token pair. It seems to be unclear how reliably the method can detect the watermark after the beginning of the text is removed.

Questions for authors

None

Reviewer_K1AJOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

ClockMark embeds multi-bit messages through a stateful exponential race that retains residual clocks across repeated contexts. A model-free receiver aggregates distinct token-pair signals. The paper provides theoretical analyses and experiments on message recovery and editing robustness.

Critical strengths
  1. The exponential-race analysis gives a coherent treatment of retained state across repeated contexts.
  2. Deduplicating token pairs accounts for repeated evidence in the stated random-key detection null.
  3. The receiver requires no access to the source model, and the reported recovery results cover multiple models and editing conditions.
Critical weaknesses
  1. The experiments do not isolate the contribution of retained clocks. A matched variant that resets clocks while preserving the payload mapping and decoder would show whether state retention explains the gains. Results grouped by context repetition would make this comparison more informative.
  2. Average bit accuracy does not establish reliable recovery of a complete identifier. Even the reported 96.35% accuracy at 64 bits leaves message-level success unresolved. Exact-message recovery, error-correction overhead, and false-attribution rates are needed for the proposed attribution use case.
  3. The uncertainty in the reported detection gains is unclear. The paper does not adequately specify independent key counts, repeated runs, or the actual denominators for attacked TPR. These details, together with paired confidence intervals, are needed to judge the reliability of small differences between methods.
Questions for authors
  1. How does recovery change when clocks are reset while the payload mapping and decoder remain fixed, particularly under repeated contexts?
  2. What are the exact-message recovery and false-attribution rates at 32 and 64 bits, including editing attacks?
  3. What independent key counts, run counts, and sample denominators support the reported TPRs and confidence intervals?
2 comments

Dear Reviewer K1AJ,

Thank you again for your valuable comments! We have carefully considered your comments and have provided our responses.

We would greatly appreciate it if you could let us know if our replies have satisfactorily addressed your issues. If you have any further questions or need additional clarification, please do not hesitate to reach out.

Thank you very much for your time and consideration.

Best regards,

Authors

As the discussion period is coming to an end, we would like to gently follow up and check whether our response has addressed your concerns. If so, we would appreciate it if you could consider updating your score. We are happy to answer any remaining questions.

Reviewer_JsjZOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents a heuristics for watermarking the text generated by LLMs and via comparing with some baselines demonstrates some performance gain.

The paper is poorly written. The proposed algorithm is not specified with adequate clarity and it is not clear how exactly the algorithm works. The cartoon figure (Figure 1) does seem to suggest that this is a simple algorithm, and I in fact doubt there is anything substantially clever.

Critical strengths

The paper is on an interesting topic.

Critical weaknesses

Poor presentation makes the paper hard to evaluate.

Questions for authors
  1. What is the algorithm?
  2. How accurate are the modeling assumptions in line 194-195 (i.e, Gaussianity and independence)?
  3. Is the watermark detection algorithm principled (for example, optimal in some sense)?
2 comments

Dear Reviewer JsjZ,

Thank you again for your valuable comments! We have carefully considered your comments and have provided our responses.

We would greatly appreciate it if you could let us know if our replies have satisfactorily addressed your issues. If you have any further questions or need additional clarification, please do not hesitate to reach out.

Thank you very much for your time and consideration.

Best regards,

Authors

As the discussion period is coming to an end, we would like to gently follow up and check whether our response has addressed your concerns. If so, we would appreciate it if you could consider updating your score. We are happy to answer any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.