acceptodds
Withdrawn from ICLR 2027

CROSSMARK: Weaving Cross-Granularity Watermarks for Robust Detection of LLM-Generated Text and Code

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Watermarks for large language model (LLM) defined at a single granularity struggle to simultaneously achieve dense detection evidence, low generation distortion, robustness to local insertion, and resilience to rewriting. We introduce CrossMark, a cross-granularity watermarking framework that combines complementary evidence across token, local-unit, and global-unit levels. Its key idea is to let different granularities specialize in different attack regimes while reinforcing one another during detection. Dense token-level evidence supports reliable clean detection, local-unit keyed preferences expose insertion-induced inconsistencies, and rewrite-stable semantic representations preserve global evidence after substantial paraphrasing or code transformation. CrossMark further couples these signals rather than treating them independently: local consistency dynamically calibrates token statistics and identifies suspicious units for semantic verification, producing a unified joint detection score. Experiments on natural-language and code benchmarks show that CrossMark delivers the strongest or near-strongest clean detection, leads evaluated baselines under local-unit insertion and single-pass LLM rewriting, and preserves competitive text quality and code correctness. Ablations further reveal a clear division of robustness: local evidence is the primary contributor against insertion, whereas global semantic evidence provides the dominant signal under rewriting.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Language Models Watermark, Representation Learning, robust watermark detection

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_TrcqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a new watermarking method called CrossMark that combines token-level logit bias, local-unit hash filtering, and learned global semantic categories within a shared candidate pool. During the watermark detection step, these multi-level statistics are aggregated to enable a more robust detection. Experiments on natural-language and code generation evaluate detection and output quality, with ablations of the generation and detection components.

Critical strengths
  • Combining local hash evidence with token weighting and semantic filtering is an interesting approach to handling insertion and rewriting. The ablations provide useful evidence of the different roles of these components.
  • Table 3 includes detection-only controls evaluated on identical outputs, which help assess the benefit of the proposed interactions beyond simply adding three independent scores.
Critical weaknesses
  • The proposed method is not fully specified. Equation (2) in the Appendix lists six training losses without their mathematical definitions or weights. The main text is somewhat abstract and less explicit about the overall watermarking procedure.
  • The baseline comparisons need more clarification. The application of ACW to natural-language data is unexplained, and the token-only ablation performs substantially worse than WLLM despite the same stated token strength.
Questions for authors
  1. The semantic-clusterer training needs a bit more detail. Since this component is central to rewriting robustness, the authors should provide the exact definitions and weights of the 6 loss components in eq(2), together with the projection/category-head architecture and optimization settings. The verbal descriptions in Appendix A.5 and the training curves in Appendix E are insufficient to reproduce the learned representation.

  2. It is a bit unclear which ACW method and implementation were used. The cited ACW method applies transformations to source code after generation, whereas Tables 1–2 report results on C4 and Figure 2 includes a watermark-strength sweep. Please clarify how it was adapted to natural language and which parameter was varied.

  3. The token-only ablation also needs more clarification. On C4, Table 3 reports a token-only AUROC of 0.884, whereas WLLM achieves 0.9658 in Table 1, although Appendix A.3 describes standard red–green watermarking and Table 4 specifies δ=3.5\delta=3.5. Please explain any differences in decoding, candidate selection, or scoring, since this gap affects how the contribution of the additional components should be interpreted.

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

This paper proposes CrossMark, a cross-granularity watermarking framework for LLM-generated text and code. It integrates token-level red–green watermarking, local-unit keyed hash preferences, and global semantic watermarking through a shared multi-path generation process. During detection, local hash evidence is further used to weight token-level statistics and filter suspicious units before global semantic verification.

Experiments on natural-language and code benchmarks demonstrate improved detection performance under clean generation, local-unit insertion, and single-pass LLM rewriting. The results suggest that different watermark granularities provide complementary robustness against different types of modifications.

Critical strengths

Meaningful problem and well-motivated design. The paper addresses an important limitation of existing watermarking approaches: token-level signals are vulnerable to insertion and rewriting, while semantic-level watermarks may provide insufficient detection evidence. Combining complementary granularities is a reasonable and practically relevant direction.

Interesting cross-level interactions. Beyond combining multiple watermark signals, CrossMark uses local hash evidence to dynamically weight token statistics and construct a filtered view for semantic verification. This is a meaningful design choice that goes beyond simply summing independent detectors.

Comprehensive experimental evaluation. The paper evaluates multiple generation backbones, natural-language and code benchmarks, and different attack settings. The ablation studies provide useful evidence for the complementary roles of the three granularities. In particular, local evidence contributes most strongly under insertion, whereas global semantic evidence is more important under rewriting. The paper also includes compute-matched controls, independent threshold calibration, and uncertainty analyses.

Critical weaknesses

Novelty relative to existing semantic watermarking methods could be better established. Several components, including sentence-level candidate selection, semantic clustering, and contrastive learning for paraphrase robustness, have precedents in NLP watermarking methods such as SemStamp and k-SemStamp. The paper would benefit from a clearer discussion of what is fundamentally new beyond combining these existing techniques. Additional comparisons with recent semantic watermarking approaches would further strengthen the contribution.

Limited evaluation against more targeted rewriting attacks. The robustness experiments primarily consider local-unit insertion and single-pass LLM rewriting. It would be useful to investigate whether the global semantic watermark remains effective under diverse rewriting strategies, repeated rewriting, and combinations of local modifications and global rewriting.

Some implementation details require clarification. The shared candidate pool is repeatedly filtered by local and global constraints, but the paper does not clearly explain how generation proceeds when no candidate satisfies a filtering condition. Clarifying the fallback or resampling strategy, together with its computational overhead, would improve reproducibility.

Questions for authors

What is the main methodological advantage of joint cross-granularity injection over a simpler combination of existing token-level and semantic watermarking methods under matched generation quality and computational budgets?

How does the rewrite-contrastive semantic encoder balance distinguishing different generation paths with maintaining category consistency under meaning-preserving rewriting?

How frequently does candidate-pool exhaustion occur during local and global filtering, and how is it handled in the implementation?

Reviewer_kz8aOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CrossMark, a cross-granularity watermarking framework that combines complementary evidence across token, local-unit, and global-unit levels.

I suggest the work to be rejected, as I seriously doubt that this paper was generated by over-reliance on AI tools or auto-research tools.

Critical strengths

This paper proposes CrossMark, a cross-granularity watermarking framework that combines complementary evidence across token, local-unit, and global-unit levels.

Critical weaknesses

(1) I doubt the paper is AI-generated or auto-researched, based on the following reference having incorrect author names or page numbers.

Yan Wang and et al. Efficient and universal watermarking for llm-generated code detection. arXiv preprint arXiv:2402.07518, 2024.

Qinkai Zheng, Xiao Xia, Xu Sui, Haohua Han, Jiangbin Yang, Kehan Guo, Zhanhui Liu, Zhaoxi Zuo,, Fan Wu, et al. Codegeex: A pre-trained model for code generation with 13 billion parameters. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 5620–5631, 2023.

(2) The work shows limited technical contributions, which combine some ideas in WLLM and SemStamp.

Questions for authors

Please kindly answer the questions in the above weakness (1) and (2).

In suggest the authors to prove that there was no overuse of AI tools.

2 comments
Area_Chair_PHmdOpenReview ↗

Importance of hallucinated references and inquiry of elaboration on the review

Dear Reviewer kz8a,

I agree with the reviewer that hallucinated and incorrect references are crucial and even this can be considered as a desk-reject reason in some other conferences. Those two references have both wrong author names and the second paper title is incorrect. Nevertheless, I was not able to find the explicit reason to reject this paper.

Would you elaborate more on the "limited technical contribution?" Otherwise, after reading author's response, please leave your remained concerns.

AC

Reviewer_kz8aOpenReview ↗

Reason for technical contributions

Limited Novelty:

At its core, CrossMark combines WLLM-style token biasing with SemStamp-style semantic partitioning. The modifications to negative sampling, hashing, and weighting are reasonable engineering refinements but just incremental. Also, evaluating different detection rules on outputs already generated by CrossMark (the independent-sum control) doesn't prove why joint generation is necessary over a straightforward pipeline combining existing methods.

Impractical Overhead:

A 7.12× token overhead for 8-path generation is hard to justify in practice. Scaling WLLM to 8 paths creates a fair compute-matched baseline, not a realistic deployment scenario. The limited timing numbers in Appendix F don't address how this scales to real-world throughput on fixed hardware, nor do they show the robustness–cost curve for smaller, more practical candidate pools.

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

Related papers

Discussion (0)

Sign in to comment.