- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper introduces TIDEmark, which jointly watermarks tokens in diffusion language models through multi-marginal optimal transport, using tree structure to reduce computational cost. Experiments demonstrate competitive detection–quality trade-offs in selected settings.
Critical strengths
- Joint watermarking is well motivated by the parallel generation capabilities of diffusion language models.
- Combining tree structure with Sinkhorn and belief propagation avoids exponential enumeration of token combinations.
- The evaluation covers two models and multiple tasks, reporting both detection performance and generation quality.
Critical weaknesses
- Differences in temperature and decoding settings across methods, together with the lack of controlled ablations, make it difficult to attribute the observed gains specifically to the joint watermarking mechanism.
- Detection thresholds are selected and evaluated on the same clean cohort, without independent validation of false-positive rates. Consequently, the generalizability of the reported TPR at fixed FPR remains unclear.
- Robustness is evaluated only for the Chain configuration on a single model under random edits, without matched baseline comparisons. Resistance to more realistic attacks, such as meaning-preserving paraphrasing, remains insufficiently established.
Questions for authors
See above.
3 comments
Response to Reviewer FDnm (2/2)
W3
Robustness is evaluated only for the Chain configuration on a single model under random edits, without matched baseline comparisons. Resistance to more realistic attacks, such as meaning-preserving paraphrasing, remains insufficiently established.
We expanded the experiments beyond one topology, one model and random edits. The new comparisons include LR-DWM, DMark and dgMARK on Finance QA and Multi-News, all three TIDEmark topologies on Finance QA, and a comparison with DMark on a second model, LLaDA2.1-mini.
We also added BERT word substitution, English-German-English back-translation and Qwen rewriting. The severity sweeps cover random deletion, insertion and substitution at 10%, 20% and 30%, and BERT word substitution at 10%, 20%, 30% and 50%.
These additions let us compare robustness across methods, topologies and models, including attacks beyond random edits. Tables 11 and 12 report detection rates and the corresponding generation-quality costs, respectively. Each detector's cutoff is selected on its unedited clean texts and kept fixed under attack. Quality changes are reported separately for each tested configuration, including those using post-fit tilting.
Manuscript changes: .
Detection under text attacks on LLaDA-Instruct (Table 11)
Detection rates (%) at cutoffs targeting 5% FPR, held fixed under attack. Deletion, insertion, substitution and BERT substitution use 20% editing.
Co denotes token-commit; Ch denotes block-ready chain. All TIDEmark configurations use previous-block MinHash. For token-commit and block-ready chain, +tilt uses strengths of 0.25 and 0.5, respectively. Rows marked use stronger tilting, with strengths of 0.5 and 1.0, respectively.
Finance QA
| Method | Deletion | Insertion | Substitution | BERT substitution | Back-translation | Rewriting |
|---|---|---|---|---|---|---|
| LR-DWM | 21.0 | 31.0 | 25.0 | 26.5 | 23.5 | 2.5 |
| DMark | 91.5 | 90.5 | 82.5 | 86.0 | 71.0 | 15.5 |
| dgMARK | 21.0 | 30.5 | 33.0 | 32.5 | 30.0 | 10.5 |
| TIDEmark (Co) | 55.0 | 57.0 | 55.0 | 61.5 | 36.5 | 17.0 |
| TIDEmark (Co+tilt) | 58.5 | 62.5 | 61.0 | 65.5 | 47.0 | 19.5 |
| TIDEmark (Co+tilt) | 61.0 | 59.5 | 64.5 | 69.5 | 47.0 | 24.0 |
| TIDEmark (Ch) | 22.5 | 21.5 | 24.5 | 31.0 | 16.0 | 4.5 |
| TIDEmark (Ch+tilt) | 38.0 | 41.5 | 39.0 | 51.0 | 22.5 | 9.5 |
| TIDEmark (Ch+tilt) | 63.0 | 61.5 | 58.5 | 71.0 | 35.5 | 10.5 |
Multi-News
| Method | Deletion | Insertion | Substitution | BERT substitution | Back-translation | Rewriting |
|---|---|---|---|---|---|---|
| LR-DWM | 20.5 | 29.5 | 22.0 | 28.5 | 20.5 | 1.0 |
| DMark | 82.5 | 83.0 | 79.5 | 81.0 | 70.0 | 11.0 |
| dgMARK | 30.0 | 35.5 | 46.5 | 48.0 | 26.0 | 14.5 |
| TIDEmark (Co) | 33.0 | 34.0 | 38.5 | 37.5 | 20.0 | 8.0 |
| TIDEmark (Co+tilt) | 47.0 | 48.0 | 58.0 | 56.5 | 27.0 | 13.0 |
| TIDEmark (Co+tilt) | 55.0 | 65.0 | 67.5 | 64.0 | 27.5 | 10.0 |
| TIDEmark (Ch) | 28.0 | 26.5 | 31.0 | 34.0 | 15.5 | 9.5 |
| TIDEmark (Ch+tilt) | 57.0 | 56.5 | 59.5 | 57.5 | 24.5 | 8.0 |
| TIDEmark (Ch+tilt) | 75.5 | 78.5 | 83.5 | 78.5 | 34.5 | 9.5 |
Detection under text attacks across topologies: Finance QA (Table 16)
Detection rates (%) at cutoffs targeting 5% FPR. All three configurations use block-index states without tilting.
| Attack | chain | balanced tree | middle-star |
|---|---|---|---|
| Unedited | 74.0 | 49.5 | 41.5 |
| Deletion 20% | 33.0 | 13.0 | 9.5 |
| Insertion 20% | 33.5 | 9.0 | 7.0 |
| Substitution 20% | 55.0 | 29.5 | 30.0 |
| BERT word substitution 20% | 54.0 | 32.5 | 22.0 |
| Back-translation | 30.5 | 8.5 | 9.5 |
| Rewriting | 12.0 | 7.0 | 7.0 |
Thank you again for your thoughtful comments and helpful feedback. We have responded to the points you raised and updated the manuscript. We would be happy to know if there is anything left we can further clarify.
Response to Reviewer FDnm (1/2)
Thank you for your suggestions. We have added temperature sweeps and a resampling control, tested whether detection thresholds transfer across datasets, and expanded robustness comparisons to additional baselines, topologies, models and rewriting attacks. Below, we explain what these experiments show and their limitations.
W1
The resampling control shows that quality loss occurs even without watermark rewards or coupling fitting. On LCC with LLaDA-Instruct, we redraw each token independently from the saved predictions after its block is complete. Code similarity decreases by 2.94 points under this control, compared with 2.69 points for block-ready TIDEmark. Thus, the observed loss cannot be attributed entirely to the joint watermarking mechanism.
We also compared KGW, Gumbel, HW and TIDEmark at temperatures of 0.3, 0.4, 0.5 and 0.6 under token-commit and block-ready insertion, giving 32 configurations. Other settings remain fixed within each method's sweep. Temperature changes the distributions supplied to fitting for HW/TIDEmark, but changes sampling for KGW/Gumbel. Higher temperatures increased detection rates, but also increased cross-entropy across all four methods. These sweeps show how detection and quality change with temperature; they do not make the methods' decoding procedures identical.
Manuscript changes: Appendices C.3 and C.6, Table 7 and Figure 6.
Deferred-resampling control on LCC (Table 7)
CE is cross-entropy in nats per token; lower values are better.
W2
We added a test with separate calibration and evaluation datasets. It shows that a cutoff targeting 5% false positives on one dataset does not necessarily retain that rate on another.
For block-ready TIDEmark with chain and block-index states on LLaDA-Instruct, we select the cutoff using one dataset's clean texts, then apply it unchanged to another dataset's clean and watermarked texts. Across these transfers, false-positive rates range from 1.5% to 10.1%. The largest increases occur on LCC. For example, the Finance QA cutoff gives 10.1% false positives on LCC, compared with 4.8% using LCC's own cutoff.
These results provide independent cross-dataset evaluation, but do not establish a universal 5% operating point. Entries using the same dataset for calibration and evaluation, and the combined-dataset row, reuse calibration texts and are not independent validation.
Manuscript changes: Appendix C.4, Table 8.
Detection-threshold transfer (Table 8)
Entries are false-positive rate / true-positive rate (%).