acceptodds
Under review as a conference paper at ICLR 2027

DenMark: Robust Semantic Watermarking for Diffusion Language Models

Abstract

Semantic text watermarks encode signals in meaning rather than surface token choices, making them inherently more robust to paraphrasing and other semantic-preserving edits. Existing semantic watermarking methods, however, are primarily designed for autoregressive language models (ARLMs), where candidate semantic units can be completed and evaluated before generation proceeds. This assumption breaks down for diffusion language models (DLMs), whose intermediate states contain incomplete semantic units and whose tokens can be updated in flexible orders. We propose DenMark, a semantic watermarking framework that injects key-dependent signals directly into the DLM denoising process. DenMark partitions the output into fixed token regions and introduces temporary rollouts as semantic lookahead: conditional completions are sampled from an incomplete region to estimate its eventual semantic watermark score, allowing DenMark to favor local denoising updates that are expected to strengthen the watermark. Repeating this procedure throughout generation progressively accumulates semantic watermark evidence in the final output. For detection, DenMark performs calibrated scanning over candidate unit sizes, improving robustness to boundary shifts caused by semantic-preserving edits. Across four DLM backbones, three datasets, and four semantic attacks, DenMark consistently outperforms existing watermarking baselines, achieving the highest TPR in 46 of 48 backbone–dataset–attack settings at the reported operating points. These results demonstrate that semantic watermarking can be effectively integrated into the flexible, non-autoregressive generation process of DLMs while retaining strong robustness to semantic edits.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Diffusion Large Language Models, Semantic Watermarking, Robust Watermarking, Text 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_DZvcOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

DenMark introduces semantic watermarking for diffusion language models. Its central idea is to evaluate local denoising decisions through temporary semantic completions: the method samples candidate token updates, fills the remaining masks within the current semantic block, and scores the resulting embeddings against keyed directions. Only the selected local update is committed; the temporary completion is discarded. Detection scans possible unit sizes using length-stratified empirical calibration and a multiple-testing correction.

The paper evaluates the method on four DLM backbones and three datasets under paraphrasing, token editing, and length-changing attacks. Additional experiments examine the encoder, decoding choices, blockwise semantic baselines, generation quality, and runtime. The results show strong detection robustness, although the generation overhead is substantial.

Critical strengths

A meaningful methodological contribution. DenMark uses temporary semantic completions to guide local denoising updates without committing the full completion. This directly addresses the difficulty of evaluating partially masked text and offers a substantive adaptation of semantic watermarking to DLM generation, rather than simply applying an existing detector to a new backbone. Strong empirical support. Evaluation across four backbones, three datasets, and diverse attacks supports the method’s robustness. The encoder ablation, blockwise semantic comparisons, and calibration checks address important alternative explanations and failure modes, making the evidence sufficiently broad to support acceptance.

Critical weaknesses

The benefit of local steering is not isolated from computational expenditure. The blockwise comparisons do not establish matched total compute: equal candidate counts do not equalize repeated selections and rollout costs. Given the approximately twelvefold slowdown over clean generation, this leaves uncertainty about both the source of the gains and their practical value. Missing artifacts limit independent verification. The submission omits the exact trained encoder and raw experimental outputs, although the encoder ablation shows that fine-tuning materially affects performance. The available code and training recipe are useful, but do not fully resolve uncertainty about reproducing the reported gains.

Questions for authors

Comparison evidence: Does DenMark retain its advantage over the blockwise semantic baselines under a matched compute or wall-clock budget at comparable output quality? This would help distinguish the benefit of local steering from additional computation. Reproducibility clarification: Are the exact trained encoder and per-example detection scores available for verification? If the checkpoint cannot be provided, is there evidence that the main results are stable across independently retrained encoders?

Reviewer_a2mROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a semantic watermarking method, DenMark, specified for mask-decoding-based diffusion language models. DenMark basically generates candidate tokens within a block and estimate the semantic score calculated using a secret key, which eventually outputs tokens with high semantic scores. To make the detector more robust, they use a scanning method to calculate the final score. Experiments show that their method outperforms the baselines under attacks.

My overall assessment: The empirical gains make this a useful contribution. However, I view the novelty as primarily an integration of candidate evaluation and semantic scoring into DLM decoding, with limited theoretical insight into the resulting robustness, distortion, and computational trade-offs.

Critical strengths
  1. The paper addresses a relevant problem and provides a practical way to apply semantic watermarking to DLMs.
  2. The empirical results look reasonable, with consistent robustness improvements across multiple models, datasets, and attacks.
Critical weaknesses
  1. Limited justification and guarantees on generation distortion: Maximizing the watermark score provides no explicit bound on deviation from the original generation distribution or degradation in text quality. Although Tables 12 and 15 report quality and downstream performance comparisons, these do not establish a distortion guarantee. A more systematic comparison of robustness at matched quality degradation would better characterize the trade-off.

  2. Substantial generation overhead and unreported detection cost: Table 2 reports 0.642 seconds per token for DenMark, compared with 0.054 for unwatermarked generation and 0.236 for DGMark—approximately 11.9× and 2.7×, respectively. This is a substantial practical cost. Detection also appears expensive due to semantic encoding across multiple block sizes, but detection time is not reported.

  3. The DIPPER evaluation only varies lexical diversity. Varying order diversity would be particularly relevant, since the watermark depends on block indices.

Questions for authors

Please see weaknesses.

Reviewer_kx2VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DenMark, a semantic watermarking method for diffusion language models. The method temporarily completes the not-yet-fully-generated text regions, assigns semantic scores to candidate local updates, and selects the update that is more conducive to watermark embedding. At the detection stage, it scans over different unit lengths to mitigate the boundary shifts caused by rewriting. Experiments show that the method achieves good detection robustness across multiple models and under various semantic attacks.

Critical strengths
  1. The motivation is sound: targeting the problem of incomplete semantic units in diffusion models, the method integrates semantic scoring with local denoising decisions, yielding a well-targeted design.

2.The detection advantage after attacks is evident, and performance is relatively consistent across multiple models and datasets.

3.The experimental coverage is fairly comprehensive, including semantic attacks, encoder ablations, boundary perturbations, and downstream task evaluation.

Critical weaknesses

1.High generation overhead. In Table 2, the runtime of DenMark is approximately 11.9× that of unwatermarked generation. A comparison under an equal time or equal model-call budget is recommended, to clarify the extent to which the performance gain stems from the additional computation.

2.The cross-domain stability of the false positive rate remains unclear. Both the calibration set and the evaluation negatives are drawn from C4, whereas the theoretical guarantee relies on the exchangeability assumption. It is recommended to fix the calibration set and threshold, and to verify the empirical false positive rate on human-written texts from different domains as well as on unwatermarked model outputs.

3.More direct evidence is needed for the reliability of the rollout approximation. The semantic score obtained from single-step completion may not accurately predict the outcome of the subsequent full denoising process. Proposition 5.1 provides a benefit decomposition but does not guarantee that local selection yields a positive benefit. It is recommended to analyze the correlation between the estimated score and the final score, as well as the candidate ranking accuracy under different masking ratios.

4.The comparison with semantic watermarking baselines can still be strengthened. Appendix H.2 already provides adapted baselines such as SemStamp and PMark, but the main comparison reports only no-attack detection results. It is recommended to further report performance after attacks and to examine the effects of block size and sampling budget.

Questions for authors

1.The generation runtime of DenMark is approximately 11.9× that of the unwatermarked model. Does its robustness advantage still hold under the same inference time or model-call budget?

2.Both the current calibration set and the evaluation negatives are drawn from C4. After fixing the calibration set and detection threshold, can the method maintain the target false positive rate on human-written texts from other domains and on unwatermarked model outputs?

3.How strongly does the semantic score from single-step rollout correlate with the score of the final denoised result? Does this predictive capability vary with the proportion of remaining masks?

4.The semantic watermarking baselines in Appendix H.2 mainly report no-attack results. Could the authors supplement their performance after rewriting attacks and analyze the impact of block size and sampling budget on the comparison results?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.