acceptodds
Under review as a conference paper at ICLR 2027

Watermarking Continuous Diffusion Language Models

Abstract

Existing watermarking methods for text generation focus on autoregressive and discrete diffusion language models, leaving continuous diffusion language models (CDLMs) unexplored. To close this gap, we introduce COMPASS (Continuous-Latent Output Marking for Provenance via Anchor Steering and Scoring), a training-free watermark designed specifically for CDLMs. COMPASS constructs key-dependent target and non-target anchor representations in the latent diffusion space and, during a selected denoising window, steers latent vectors associated with non-target anchors toward the target anchor set. Detection requires only the candidate text. The detector re-encodes the text using the CDLM’s encoder and measures whether the resulting latent representations align with the target anchors defined by the watermark key. We show that this design provides robust detection of content generated by state-of-the-art CDLMs, namely Cola and ELF-L, while preserving their high-quality outputs. Furthermore, COMPASS requires neither model fine-tuning nor additional neural-network evaluations during generation, providing an efficient and practical framework for provenance tracing in CDLMs.

Primary area:
generative models
Keywords:
diffusion, continusous diffusion language models, watermarking

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NVNfOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces COMPASS, a training-free watermarking method for continuous diffusion language models. It steers intermediate latent representations toward key-defined target anchors during denoising. Detection re-encodes the output text and measures its preference for target over non-target anchors. Experiments on Cola and ELF-L show improved detection and robustness to text edits with low computational overhead.

Critical strengths
  1. The paper studies a novel setting. To the best of my knowledge, it is the first work on watermarking continuous diffusion language models.

  2. Although the method builds on existing watermarking ideas, its design is reasonable and well suited to the continuous denoising process.

Critical weaknesses
  1. On page 3, the authors explain that decoding is many-to-one and re-encoding cannot recover the original latents, which weakens the image-watermarking baselines. The same issue also affects COMPASS. Although COMPASS only needs to preserve a bias toward target regions, this advantage needs clearer evidence. Comparing the watermark signal before decoding and after re-encoding would help.

  2. The claim of preserving text quality needs stronger support. The absolute log-perplexity gap on Cola is substantial, but its direction is not reported. Also, the ELF-L examples use weaker steering than the main experiments. Raw perplexity values, human evaluation, and examples generated with the main settings would make the quality claims more convincing.

  3. Detection thresholds are calibrated on unedited, unwatermarked texts and then fixed across attacks. However, the false-positive rate may change after editing, especially when texts become shorter. The authors should apply the same edits to unwatermarked texts and report the resulting false-positive rates.

  4. The ablations mainly vary hyperparameters, but do not clearly justify several design choices. For example, how does random target-anchor selection compare with choosing the nearest target anchor? Does SLERP offer an advantage over normalized linear interpolation?

Questions for authors

See Weaknesses.

6 comments
AuthorsOpenReview ↗

Thank you & watermark survival through decoding and re-encoding

We thank the Reviewer for the positive feedback. We appreciate recognizing our watermarking scheme for CDLMs as novel and well designed. We address each concern one by one below:

W1: On page 3, the authors explain that decoding is many-to-one and re-encoding cannot recover the original latents, which weakens the image-watermarking baselines. The same issue also affects COMPASS. Although COMPASS only needs to preserve a bias toward target regions, this advantage needs clearer evidence. Comparing the watermark signal before decoding and after re-encoding would help.

We compare detection on the final generated latents before decoding with detection after the generated text is re-encoded. For Gaussian Shading and Tree-Ring, the corresponding latent representation is passed through the inversion and keyed detector.

For Cola:

MethodBefore decoding TPRAfter re-encoding TPR
COMPASS (ten test sets, mean)93.92%93.93%
Gaussian Shading69.50%32.00%
Tree-Ring99.67%88.17%

For ELF-L:

MethodBefore decoding TPRAfter re-encoding TPR
COMPASS (ten test sets, mean)99.39%100.00%
Gaussian Shading100.00%53.50%
Tree-Ring100.00%18.20%

The signal for COMPASS remains detectable after re-encoding almost to the same extent as before the decoding. We attribute this robustness of COMPASS to the fact that it relies on a directional bias rather than exact reconstruction of every latent. We provide further details in Appx. J.9 (Figure 18).

AuthorsOpenReview ↗

Quality evaluation: raw perplexity, signed changes, paired multi-judge ratings, and MAUVE

W2: The claim of preserving text quality needs stronger support. The absolute log-perplexity gap on Cola is substantial, but its direction is not reported. Also, the ELF-L examples use weaker steering than the main experiments. Raw perplexity values, human evaluation, and examples generated with the main settings would make the quality claims more convincing.

Raw perplexity and signed change. Perplexity is calculated using the pretrained Llama-3.1-8B base model, with Cola continuation tokens scored conditional on their prompts (Appx. N.1). We note that our method is on the Pareto frontier at any text quality value that a model provider may select (please see Figure 3). We selected the values for Tables 1 and 2 that reflect an average Δ\Delta log PPL among all compared methods. For Cola, the reference-model perplexity increases from 26.34 (clean) to 78.06 (COMPASS), and a signed Δ\Delta log PPL of +1.0865. We include the additional results in Appx. O.3, Table 36. While perplexity measures the reference model's assigned likelihood, it is not by itself a complete measurement of fluency, coherence, or preservation of meaning. This is why we also reported scores for independent language-model judges and add additional results.

Paired quality assessment and interpretable scores. CompassJudger evaluates fluency, coherence, repetition, naturalness, and meaningfulness on a 0-4 degradation scale (lower is better). The previously reported 9.1% degradation on Cola and 4.2% on ELF-L is a relative increase in the mean rating. We now give the absolute values rather than relying on a percentage of an ordinal rating.

Quality evaluationClean mean ↓COMPASS mean ↓Mean paired change ↓
Cola2.1052.296+0.191
ELF-L2.2202.314+0.094

We added a multi-judge evaluation on 600 Cola continuations (Appx. O.1, Table 34). In addition to CompassJudger-2-7B-Instruct, we evaluate outputs with Llama-3.1-8B-Instruct and Mistral-Small-24B-Instruct. Each judge independently scores the same five linguistic dimensions using the same 0-4 degradation rubric. We report the mean paired increase in degradation rating (watermarked minus clean), with lower values indicating better preservation. All eight baselines are evaluated at their strengthened paraphrase-evaluation configurations.

MethodCompassJudger-2 ↓Llama-3.1-8B ↓Mistral-Small-24B ↓
COMPASS+0.264+0.305+0.510
TextSeal+0.304+0.320+0.576
KGW+0.591+0.439+0.761
Unigram+0.984+0.522+0.862
SynthID+0.620+0.434+0.795
DiPmark+1.838+0.690+0.967
GumbelSoft+1.810+0.662+0.962
Tree-Ring+1.632+0.619+0.963
Gaussian Shading+0.672+0.609+0.921

COMPASS has the lowest observed mean degradation under all three judges, with TextSeal second in each case. The remaining baselines are included in Table 34. We emphasize the consistent ordering of the point estimates across independent evaluators, without treating the small gap to TextSeal as proven statistically significant.

We also include an exploratory distributional assessment against human-written C4 continuations using the MAUVE score (Appx. O.2, Table 35). A higher MAUVE score means that the generated text is closer to the human distribution. On the original 600 Cola prompts, COMPASS obtains 0.611, compared with 0.436 for TextSeal. This analysis is restricted to prompt-conditioned Cola: unconditional ELF-L has no corresponding prompt-matched human-reference distribution in our experiment. We present MAUVE as complementary evidence about distributional agreement, not as proof of semantic faithfulness or an individual-text human-quality rating.

We have not conducted a formal blinded human-rating study within the limited time. Instead, we provide reproducible cross-judge ratings, paired uncertainty estimates, reference-model perplexity, and the distributional check above.

To summarize, the additional judges, prompt-paired comparisons, and distributional analysis substantially broaden the original text quality evaluation. The new results further support the conclusion that COMPASS achieves favorable empirical quality–detection trade-off for the studied CDLMs.

W3: Detection thresholds are calibrated on unedited, unwatermarked texts and then fixed across attacks. However, the false-positive rate may change after editing, especially when texts become shorter. The authors should apply the same edits to unwatermarked texts and report the resulting false-positive rates.

In all evaluations, we kept the same fixed threshold (calibrated to 1% FPR on unattacked unwatermarked generations) rather than recalibrating per source or attack. Our goal was to reflect the most realistic deployment scenario: at detection time, the detector does not know in advance whether an arbitrary input text is human-written, unwatermarked model output, edited, truncated, or modified in any adversarial way, and must evaluate any candidate text using that single frozen threshold.

In Tables 1 and 2 we relabelled the attack columns to "TPR at the clean-calibrated threshold" and added the results for both FPRs and recalibrated TPRs to the appendix.

False-positive rate on Cola:

We evaluated all 11 post-generation attacks on unwatermarked outputs from Cola with the same frozen 1%-FPR thresholds from the initial submission. We observe that the resulting FPR increase is very modest (with average around 1.79%), considering that the TPR for COMPASS stays much higher (77.2%-88.3%) than for other methods (e.g., TPR collapsing to 0.0%–5.5% for TreeRing) :

For Cola:

MethodWD10WD30WD50Span30First50Last50Swap20Typo5ShuffleCP25CP50
COMPASS (n=600)1.7%2.2%3.2%1.7%2.2%3.5%0.8%1.0%1.0%1.2%1.2%
KGW (n=600)1.0%1.3%1.2%0.5%0.7%0.7%0.7%1.0%1.3%1.0%1.0%
Unigram (n=600)0.5%0.7%0.3%0.7%0.7%0.5%0.8%2.5%0.8%0.8%0.8%
GumbelSoft (n=600)0.3%0.3%0.3%0.3%0.3%1.3%0.7%0.3%0.7%1.0%1.0%
TextSeal (n=600)1.3%0.8%1.0%0.5%0.3%0.2%1.0%0.8%0.3%1.3%1.3%
SynthID (n=600)1.7%0.7%0.3%0.3%1.0%0.5%0.7%1.0%1.0%1.0%1.0%
DiPmark (n=600)2.0%1.7%2.2%1.3%1.5%1.2%1.0%2.2%0.8%1.0%1.0%
Tree-Ring (n=600)1.0%0.2%0.0%0.0%0.0%0.0%2.0%2.3%1.0%1.0%1.0%
Gaussian Shading (n=600)1.8%0.8%2.2%1.3%2.3%3.3%1.3%0.7%1.5%1.0%1.0%

We also report the corresponding TPRs:

MethodDeg. (%) ↓|Δ log PPL| ↓Unattacked ↑WD10 ↑WD30 ↑WD50 ↑Span30 ↑First50 ↑Last50 ↑Swap20 ↑Typo5 ↑Shuffle ↑CP25 ↑CP50 ↑
COMPASS (ours)9.11.0994.286.888.382.386.277.286.089.778.788.277.354.3
KGW10.90.7635.217.810.24.716.311.715.814.711.828.823.014.7
Unigram11.60.8935.224.820.513.217.810.317.527.822.329.822.713.5
GumbelSoft22.41.5049.521.84.71.325.216.821.214.39.839.732.823.8
TextSeal9.20.7276.847.822.26.252.537.546.544.729.370.065.350.2
SynthID15.01.0772.026.57.01.835.325.329.517.212.055.551.330.8
DiPmark15.91.068.36.74.72.36.54.33.54.23.57.75.33.5
Tree-Ring25.92.4189.22.00.00.05.50.00.079.05.214.738.020.2
Gaussian Shading34.21.2931.532.545.045.843.842.750.031.537.333.726.823.0

On surface edits (Swap20, Typo5, Shuffle) and dilution (CP25, CP50), COMPASS remains its FPR close to 1% (0.8%-1.2% FPR). Under deletion and truncation, shortening further reduces the number of averaged latents and increases the variance of the mean margin, raising COMPASS's FPR to 1.7%-3.5%, comparable to Gaussian Shading (up to 3.3%) and DiPmark (up to 2.2%), while COMPASS still retains 77.2%-88.3% TPR on those same edits. By contrast, Tree-Ring's 0.0% FPR under heavy deletion/truncation coincides with its TPR collapsing to 0.0%–5.5%.

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

The paper proposes COMPASS, a training-free watermarking method for continuous diffusion language models. A secret key defines target and non-target regions through an anchor codebook. During selected denoising steps, selected latent directions are steered toward target anchors while preserving their norms. Detection re-encodes the final text and aggregates keyed margins without recovering the generation trajectory or initial noise. Experiments on Cola and ELF-L compare eight adapted baselines and evaluate text edits and DIPPER paraphrasing.

Critical strengths
  1. CDLM watermarking solution. COMPASS operates directly on continuous generation states and detects the signal from final text, without requiring the original trajectory. This is consequential because it offers a plausible solution where token-decoder modification and initial-noise recovery are difficult.

  2. Empirical effectiveness. Tables 1-2 report strong clean detection and robustness to several edits, particularly on ELF-L. Figure 3 additionally examines the detection-quality trade-off on Cola. These results provide substantive support for the feasibility of the method and make the work potentially useful beyond a conceptual proposal.

Critical weaknesses
  1. Calibration scope. Section 3.5 allows variable candidate length LxL_x, whereas the evaluation predominantly scores 128-token texts. Keeping a clean-calibrated threshold fixed under attacks is reasonable, but calibration/test separation and the actual FPR across lengths and edits are not clearly established. Since the mean-margin distribution may depend on length and positional correlations, these uncertainties limit confidence in the scope of the advertised low-FPR detection and robustness claims.

  2. Quality-robustness evidence. Tables 1-2 and Figure 3 provide useful evidence of a favorable detection-quality trade-off. However, the strengthened baseline configurations used in Table 3 are not accompanied by their corresponding quality costs, leaving the robustness comparison insufficiently connected to output quality. Table 3 also shows that token-level methods can achieve high clean detection: Unigram reaches 96.5% on Cola, versus 94.2% for COMPASS. Thus, the claims that token-level methods "fail" and that COMPASS achieves the highest detection rate require qualification. Since improving robustness while maintaining quality is central to the contribution, the main text needs a more consistent comparison linking these dimensions to substantiate the breadth of its claimed advantage.

Questions for authors
  1. Calibration independence. Section 3.5 specifies held-out calibration text, but its relationship to the negative test set remains unclear. Were the clean-score statistics and detection thresholds estimated on data separate from the negative evaluation samples? Please clarify the sample sizes, data split, and whether the reported FPR was independently verified in the main text.

  2. Diffusion-configuration applicability. How does COMPASS transfer when the number of denoising steps, guidance scale, or sampler changes? In particular, how is the watermarking window mapped across different schedules, and does maintaining the reported detection-quality trade-off require retuning? Please clarify the supporting evidence and limits of applicability.

  3. Threat model and provenance claims. Are the provenance claims restricted to the evaluated attacks without model, key, or detector access, or do they also encompass adaptive removal, watermark forgery, and false attribution? Please distinguish detecting a keyed watermark from establishing reliable source attribution, and identify which claims are experimentally supported.

  4. Key security and multi-user scalability. What PRNG, seed space, and key entropy are used, and what security properties are intended? Since codebooks can be reconstructed from the key and shared reference distribution, what per-key state must actually be stored or cached, and how do setup, calibration, and detection costs scale with the number of users? If multi-user attribution is intended, how are wrong-key matches, attribution-search costs, and the overall false-positive rate across multiple keys handled?

10 comments
AuthorsOpenReview ↗

Thank you & Clarifying calibration data, length-dependent detection, and false positives on human-written text

We thank the Reviewer for the positive feedback. We appreciate recognizing our watermarking scheme for CDLMs, its empirical effectiveness, robustness results, and practicality. We address each concern one by one below:

W1 & Q1: Calibration scope. Section 3.5 allows variable candidate length LxL_x, whereas the evaluation predominantly scores 128-token texts. Keeping a clean-calibrated threshold fixed under attacks is reasonable, but calibration/test separation and the actual FPR across lengths and edits are not clearly established. Since the mean-margin distribution may depend on length and positional correlations, these uncertainties limit confidence in the scope of the advertised low-FPR detection and robustness claims. Q1: Calibration independence. Section 3.5 specifies held-out calibration text, but its relationship to the negative test set remains unclear. Were the clean-score statistics and detection thresholds estimated on data separate from the negative evaluation samples? Please clarify the sample sizes, data split, and whether the reported FPR was independently verified in the main text.

In all evaluations, we kept the same fixed threshold (calibrated to 1% FPR on unattacked unwatermarked generations) rather than recalibrating per source or attack. Our goal was to reflect the most realistic deployment scenario: at detection time, the detector does not know in advance whether an arbitrary input text is human-written, unwatermarked model output, edited, truncated, or modified in any adversarial way, and must evaluate any candidate text using that single frozen threshold.

In Tables 1 and 2 we relabelled the attack columns to "TPR at the clean-calibrated threshold" and added the results for both FPRs and recalibrated TPRs to the appendix.

Variable candidate length. We additionally evaluated continuations up to 512 tokens, calibrating a 1%-target threshold on 600 length-matched clean continuations. The table below reports TPR (%). Appendix I gives the full curves in Figure 13. FPR is measured on each length independently.

Method\Token length3264128256512
COMPASS54.876.894.297.094.8
KGW6.58.835.551.562.7
Unigram1.07.824.351.762.0
GumbelSoft11.718.062.855.332.8
TextSeal30.557.578.782.074.5
SynthID17.037.773.082.282.7
DiPmark0.31.35.88.314.5
Tree-Ring1.77.788.344.059.8
Gaussian Shading25.322.030.033.317.2

Longer continuations do not guarantee higher TPR: the amount of available text, score distribution, and length-specific thresholds can all change. Overall, we observe that our COMPASS achieves the highest TPR across all the evaluated token lengths.

Sample usage clarification. In the main Cola experiments (Section 4.1 and Appendix J), 600 prompts from C4 realnewslike yield 600 paired watermarked and unwatermarked 128-token continuations. The 600 unwatermarked continuations calibrate the 1%-FPR thresholds and serve as the clean reference in hyperparameter ablations and main evaluation. The original human-written C4 continuations tested below were not used during calibration or hyperparameter selection.

Human-written text (unattacked original C4 continuations). Under the frozen model-calibrated thresholds, COMPASS achieves 0.5% FPR for Cola and 0.3% for ELF-L (below the 1% target) on human-written text, whereas, for example, Tree-Ring (1.8%) and Unigram (2.3%) exceed it:

FPR for human written text on Cola:

MethodFPR, FP/600 [95% Wilson CI, %]
COMPASS0.5% (3/600) [0.2, 1.5]
KGW0.8% (5/600) [0.4, 1.9]
Unigram2.3% (14/600) [1.4, 3.9]
GumbelSoft0.5% (3/600) [0.2, 1.5]
TextSeal0.8% (5/600) [0.4, 1.9]
SynthID0.8% (5/600) [0.4, 1.9]
DiPmark1.0% (6/600) [0.5, 2.2]
Tree-Ring1.8% (11/600) [1.0, 3.3]
Gaussian Shading0.3% (2/600) [0.1, 1.2]

FPR for human written text on ELF-L:

MethodFPR, FP/600 [95% Wilson CI, %]
COMPASS0.3% (2/600) [0.1, 1.2]
KGW0.0% (0/600) [0.0, 0.6]
Unigram0.3% (2/600) [0.1, 1.2]
GumbelSoft0.0% (0/600) [0.0, 0.6]
TextSeal0.0% (0/600) [0.0, 0.6]
SynthID0.0% (0/600) [0.0, 0.6]
DiPmark0.5% (3/600) [0.2, 1.5]
Tree-Ring0.0% (0/600) [0.0, 0.6]
Gaussian Shading4.3% (26/600) [3.0, 6.3]
AuthorsOpenReview ↗

FPR across edits: we evaluated perturbed clean (unwatermarked) generations from Cola under the same fixed threshold

We evaluated all 11 post-generation attacks on unwatermarked outputs from Cola with the same frozen 1%-FPR thresholds from the initial submission. We observe that the resulting FPR increase is very modest (with average around 1.79%), considering that the TPR for COMPASS stays much higher (77.2%-88.3%) than for other methods (e.g., TPR collapsing to 0.0%–5.5% for TreeRing) :

For Cola:

MethodWD10WD30WD50Span30First50Last50Swap20Typo5ShuffleCP25CP50
COMPASS (n=600)1.7%2.2%3.2%1.7%2.2%3.5%0.8%1.0%1.0%1.2%1.2%
KGW (n=600)1.0%1.3%1.2%0.5%0.7%0.7%0.7%1.0%1.3%1.0%1.0%
Unigram (n=600)0.5%0.7%0.3%0.7%0.7%0.5%0.8%2.5%0.8%0.8%0.8%
GumbelSoft (n=600)0.3%0.3%0.3%0.3%0.3%1.3%0.7%0.3%0.7%1.0%1.0%
TextSeal (n=600)1.3%0.8%1.0%0.5%0.3%0.2%1.0%0.8%0.3%1.3%1.3%
SynthID (n=600)1.7%0.7%0.3%0.3%1.0%0.5%0.7%1.0%1.0%1.0%1.0%
DiPmark (n=600)2.0%1.7%2.2%1.3%1.5%1.2%1.0%2.2%0.8%1.0%1.0%
Tree-Ring (n=600)1.0%0.2%0.0%0.0%0.0%0.0%2.0%2.3%1.0%1.0%1.0%
Gaussian Shading (n=600)1.8%0.8%2.2%1.3%2.3%3.3%1.3%0.7%1.5%1.0%1.0%

We also report the corresponding TPRs:

MethodDeg. (%) ↓|Δ log PPL| ↓Unattacked ↑WD10 ↑WD30 ↑WD50 ↑Span30 ↑First50 ↑Last50 ↑Swap20 ↑Typo5 ↑Shuffle ↑CP25 ↑CP50 ↑
COMPASS (ours)9.11.0994.286.888.382.386.277.286.089.778.788.277.354.3
KGW10.90.7635.217.810.24.716.311.715.814.711.828.823.014.7
Unigram11.60.8935.224.820.513.217.810.317.527.822.329.822.713.5
GumbelSoft22.41.5049.521.84.71.325.216.821.214.39.839.732.823.8
TextSeal9.20.7276.847.822.26.252.537.546.544.729.370.065.350.2
SynthID15.01.0772.026.57.01.835.325.329.517.212.055.551.330.8
DiPmark15.91.068.36.74.72.36.54.33.54.23.57.75.33.5
Tree-Ring25.92.4189.22.00.00.05.50.00.079.05.214.738.020.2
Gaussian Shading34.21.2931.532.545.045.843.842.750.031.537.333.726.823.0

On surface edits (Swap20, Typo5, Shuffle) and dilution (CP25, CP50), COMPASS remains its FPR close to 1% (0.8%-1.2% FPR). Under deletion and truncation, shortening further reduces the number of averaged latents and increases the variance of the mean margin, raising COMPASS's FPR to 1.7%-3.5%, comparable to Gaussian Shading (up to 3.3%) and DiPmark (up to 2.2%), while COMPASS still retains 77.2%-88.3% TPR on those same edits. By contrast, Tree-Ring's 0.0% FPR under heavy deletion/truncation coincides with its TPR collapsing to 0.0%–5.5%.

AuthorsOpenReview ↗

FPR across edits: we evaluated perturbed clean (unwatermarked) generations from ELF-L under the same fixed threshold

We also evaluated all 11 post-generation attacks on unwatermarked outputs from ELF-L with the same frozen 1%-FPR thresholds from the initial submission. We observe that the resulting FPR increase is very modest (from 0% to 3.7%), considering that the TPR for COMPASS stays much higher (93%-100%) than for other methods (e.g., TPR decreasing to less than 26.2% for TreeRing):

For ELF-L:

MethodWD10WD30WD50Span30First50Last50Swap20Typo5ShuffleCP25CP50
COMPASS (n=600)2.4%1.7%2.5%1.9%3.7%3.3%2.2%0.0%1.5%2.2%2.2%
KGW (n=600)1.0%1.2%0.8%1.2%0.8%0.8%0.7%0.7%1.3%1.0%1.0%
Unigram (n=600)1.3%1.5%1.5%1.7%1.0%1.0%1.0%2.2%1.0%1.0%1.0%
GumbelSoft (n=600)1.0%1.0%2.0%1.2%2.2%1.0%0.7%0.2%1.2%1.0%1.0%
TextSeal (n=600)1.8%1.3%0.8%1.2%1.7%1.8%1.3%1.5%1.7%1.0%1.0%
SynthID (n=600)0.8%0.2%0.0%0.2%0.3%0.0%1.0%3.0%1.0%1.0%1.0%
DiPmark (n=600)2.2%1.7%1.2%1.0%2.0%1.2%1.2%1.7%1.8%1.7%1.7%
Tree-Ring (n=600)1.5%5.2%4.8%3.3%3.3%3.3%2.3%4.2%2.5%1.3%1.3%
Gaussian Shading (n=600)1.0%1.0%2.8%1.2%2.7%2.5%0.7%0.0%1.2%0.8%0.8%

We report the corresponding TPR results for ELF-L as well:

MethodDeg. (%) ↓|Δ log PPL| ↓Unattacked ↑WD10 ↑WD30 ↑WD50 ↑Span30 ↑First50 ↑Last50 ↑Swap20 ↑Typo5 ↑Shuffle ↑CP25 ↑CP50 ↑
COMPASS (ours)4.30.62100.0100.0100.099.7100.099.8100.0100.093.0100.099.570.0
KGW8.30.8275.565.844.520.559.246.045.357.823.374.049.323.3
Unigram6.51.1183.280.873.360.873.761.261.583.270.783.261.531.5
GumbelSoft8.80.7283.768.736.715.073.061.867.560.334.580.271.748.2
TextSeal7.50.6777.861.533.510.567.356.059.852.029.574.764.241.5
SynthID9.70.9136.728.819.57.317.77.87.836.757.536.717.06.2
DiPmark5.22.5192.059.313.32.376.363.758.335.07.087.379.369.7
Tree-Ring31.90.1718.29.226.011.526.216.812.815.06.512.012.311.2
Gaussian Shading2.42.7653.553.558.263.758.264.763.353.875.553.530.714.7
Reviewer_6BCkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors of the paper propose COMPASS. COMPASS is a training-free watermark for continuous diffusion language models. A secret key is used to determine target and non-target anchors in the model’s latent space. Then during a selected denoising window, the method steers a subset of non-target latent directions toward target anchors while preserving their norms.

Detection re-encodes candidate text with the model’s frozen encoder and measures its relative alignment with the two anchor sets. The paper then analyzes source-cell survival under an idealized random-anchor experiment and evaluates the complete method on Cola and ELF-L datasets. Experiments compare some adapted watermarking baselines, examine editing and paraphrasing robustness, ablate the steering parameters, and measure text-quality proxies and computational overhead.

Critical strengths

(1) The paper develops a practical response to a diffusion architectural challenge: In Appendix A, the authors discuss the highly concentrated output-token distributions in the evaluated models and explain why interventions at final token selection have limited room to operate. Steering intermediate representations provides a useful alternative that works within the existing denoising framework. Detection from re-encoded text also addresses the information loss introduced by converting continuous states into discrete outputs. This seems to be nice contribution for continuous diffusion language generation but this might have already been done before in the literature.

(2) The empirical detection and robustness evidence is good: Tables 1–3 show strong separation from multiple adapted baselines, including baseline-strength adjustments for the paraphrasing comparison. COMPASS reports unattacked TPRs of 94.2% on Cola and 100% on ELF-L. Under DIPPER 40/40 paraphrasing, the corresponding TPRs are 46.3% and 94.5%. The hyperparameter sweeps and ablations help explain the operating tradeoffs. Appendix I additionally reports generation overheads of only 0.34% and 1.63%, without additional neural-network evaluations during watermark injection.

Critical weaknesses

(1) It seems the technique of steering has been evaluated previously in the literature (e.g., see [1]). How different is the use of steering compared to previous works in the literature (not just [1])?

(2) Issues with robustness results and preserving useful content: Section 4.1 calibrates thresholds on unedited, unwatermarked model outputs and then keeps them fixed under attacks. The attacked columns therefore measure retained detection at a clean-calibrated threshold; I do not actually think these imply a 1% FPR on edited or shortened negatives. Deletion and truncation can change the variance of the mean-margin statistic and paraphrasing can change its distribution. I also could not identify independent human-negative testing and sufficiently explicit separation of calibration, parameter selection, and final negative evaluation. For ELF-L, splitting a 1,024-token generation into eight segments additionally requires accounting for dependence when constructing splits and uncertainty estimates. I believe that these limit the interpretation of the results. Also, the evidence for preserving useful content is less complete than the detection evidence: Quality evaluation relies on a reference-model perplexity measure and one LLM judge. The reported percentage degradation is a relative change in an average ordinal judge score, which is difficult to translate into terms. On Cola, the main configuration also has an absolute log-perplexity change of 1.09. These measurements are useful, but I believe that paired assessment of the existing outputs would make the claim of preserved quality more convincing since the steering strength is close to a full move toward the selected anchor.

[1] https://arxiv.org/abs/2605.05220

Questions for authors

(1) The authors discuss some related works on discrete diffusion. But there's one work that has pretty good empirical evaluation and that does not have hyperparameter tuning. Can the authors compare to [2]?

(2) Also, there are issues with false-positive calibration: After freezing all settings and thresholds, what FPR is observed on independent unwatermarked model outputs, human-written text, and corresponding edited or truncated negatives? Please report confidence intervals and clarify which samples were used for hyperparameter selection, calibration, and final evaluation.

(3) Key and sampling variability: How do detection, FPR, and quality vary across independently sampled keys? For ELF-L, are calibration/evaluation splits and resampling performed at the original 1,024-token generation level, so correlated segments do not cross splits or inflate the effective sample size?

(4) Quality validation: I wonder if the favorable detection–quality ordering persists under an assessment beyond the current judge and perplexity proxies?

[2] https://arxiv.org/abs/2511.02083

12 comments
AuthorsOpenReview ↗

Thank you & COMPASS’s novelty lies not in representation steering itself, but in a keyed watermarking mechanism for continuous diffusion language models

We thank the Reviewer for recognizing the architectural motivation, empirical detection, robustness results, and low generation overhead. We address each concern one by one below:

W1: It seems the technique of steering has been evaluated previously in the literature (e.g., see [1]). How different is the use of steering compared to previous works in the literature (not just [1])?

COMPASS’s novelty lies not in representation steering itself, but in a keyed watermarking mechanism for continuous diffusion language models (CDLMs). Unlike prior steering methods, COMPASS uses a keyed partition of latent space rather than a fixed direction, without labels or training. Our design addresses two central challenges: preserving the watermark through denoising and through text decoding and re-encoding. We provide a comparison with prior work, theoretical and empirical survival analyses, and an evaluation of the detection–quality trade-off.

While steering of intermediate representations has been shown to be applicable, for example, to vision diffusion models [1] and to continuous text-diffusion latents for controllable generation [8], we are the first to use it for watermarking of CDLMs. The approach from [1] changes the intermediate representations of a generative model during generation by adding a steering vector that encodes a target concept. Instead, we add many such steering vectors that form a watermark (we add a hidden key signal that can be detected). Thus, while our method can be seen as a form of intermediate representation steering, it substantially changes its form to obtain a distinct watermarking mechanism. The key contribution is not the steering of representation(s) themselves but a method to embed a keyed preference that survives denoising and the continuous-to-discrete decoding and discrete-to-continuous encoding. We additionally provide a comprehensive comparison with other methods that leverage steering in the Table below:

WorkModelSteering localizationSteering target obtained fromPurpose
MidSteer [1]AR LLM; image diffusionSelf-attention; cross-attention, respectivelyConcept-labelled activation covariancesConcept switching
ActAdd [3], CAA [4], ITI [5], RepE [6], HiDRA [7]AR LLMActivationsContrastive or labelled examplesControl
Diffusion-LM [8]Continuous text diffusionLatentsTrained-classifier gradientsControl
MDLM steering [9], ILRR [15]Masked DLMActivationsContrastive prompts; reference textControl
Guidance Watermarking [10]Image diffusionDenoising statesPretrained watermark-decoder gradientsImage watermarking
DenMark [11]Discrete DLMToken updates, selected via rolloutsKeyed directions in an external sentence encoderDiscrete-DLM watermarking
SLAM [12]AR LLMResidual streamKeyed subset of SAE-derived structural directionsAR-LLM watermarking
Pletenev (concurrent) [13]AR LLMResidual streamMean difference or random directionWatermark test; negative result
AWM [14]AR LLMHidden statesKey-derived direction embedded by fine-tuningSafety monitoring
COMPASS (ours)Continuous DLMLatents at a few denoising stepsKey + unlabelled reference distribution; no labels or trainingContinuous-DLM watermarking

If there is any other work on representation steering that the Reviewer would like us to additionally consider, we are happy to do so.

AuthorsOpenReview ↗

Beyond Fixed-Direction Steering: Keyed Watermarking That Survives Denoising and Re-encoding

We also provide our distinct contributions beyond the application of the steering:

  1. A keyed partition instead of a fixed direction. ActAdd [3], CAA [4], and SLAM [12] add the same vector independent of the current activation, and MidSteer [1] folds one affine map into the attention weights. COMPASS splits the latent space into keyed target and non-target anchor cells. Among the latents whose nearest anchor is non-target, it moves a fixed fraction toward a sampled target anchor and restores their norm. Latents already in a target cell are left unchanged. Detection measures the re-encoded text's preference for target over non-target anchors, at a threshold calibrated on unwatermarked text: a continuous counterpart of a green-list test [16] over tokens.

  2. Survival through the remaining denoising steps. In an autoregressive LLM, a steered activation acts directly on the next-token distribution. In a CDLM, the denoising steps after the intervention can undo it. Particularly for larger codebooks, target occupancy approaches chance when the watermark is applied too early (Appx. F, Figure 9). For an idealized move onto an anchor, Appendix F derives in closed form the probability that the final latent is still in that anchor's cell. It compares this parameter-free prediction with measured survival (Figures 9-10). The analysis motivates the watermarking window, which trades detection against quality: extending the window from [4, 6] to [4, 9] raises TPR@1% from 94.33% to 99.83%, but judged degradation also rises from 9.07% to 58.62% (Table 21).

  3. Survival through decoding and re-encoding. COMPASS steers latents toward a key-defined set of target anchors, creating an aggregate directional bias that its detector can test after the generated text is re-encoded. This design does not require recovery of each original latent. To evaluate whether the bias survives in the reencoded latents, we compare detection before decoding and after re-encoding, including the Gaussian Shading and Tree-Ring baselines.

AuthorsOpenReview ↗

Fixed-Threshold Robustness and False Positives on Human-Written Text

W2 & Q2: Issues with robustness results (...): Section 4.1 calibrates thresholds on unedited, unwatermarked model outputs and then keeps them fixed under attacks. The attacked columns therefore measure retained detection at a clean-calibrated threshold; I do not actually think these imply a 1% FPR on edited or shortened negatives. Deletion and truncation can change the variance of the mean-margin statistic and paraphrasing can change its distribution. I also could not identify independent human-negative testing and sufficiently explicit separation of calibration, parameter selection, and final negative evaluation. & Q(2) Also, there are issues with false-positive calibration: After freezing all settings and thresholds, what FPR is observed on independent unwatermarked model outputs, human-written text, and corresponding edited or truncated negatives? Please report confidence intervals and clarify which samples were used for hyperparameter selection, calibration, and final evaluation.

In all evaluations, we kept the same fixed threshold (calibrated to 1% FPR on unattacked unwatermarked generations) rather than recalibrating per source or attack. Our goal was to reflect the most realistic deployment scenario: at detection time, the detector does not know in advance whether an arbitrary input text is human-written, unwatermarked model output, edited, truncated, or modified in any adversarial way, and must evaluate any candidate text using that single frozen threshold.

In Tables 1 and 2 we relabelled the attack columns to "TPR at the clean-calibrated threshold" and added the results for both FPRs and recalibrated TPRs to the appendix.

  1. Human-written text (unattacked original C4 continuations): Under the frozen model-calibrated thresholds, COMPASS achieves 0.5% FPR for Cola and 0.3% for ELF-L (below the 1% target) on human-written text, whereas, for example, Tree-Ring (1.8%) and Unigram (2.3%) exceed it:

FPR for human written text on Cola:

MethodFPR, FP/600 [95% Wilson CI, %]
COMPASS0.5% (3/600) [0.2, 1.5]
KGW0.8% (5/600) [0.4, 1.9]
Unigram2.3% (14/600) [1.4, 3.9]
GumbelSoft0.5% (3/600) [0.2, 1.5]
TextSeal0.8% (5/600) [0.4, 1.9]
SynthID0.8% (5/600) [0.4, 1.9]
DiPmark1.0% (6/600) [0.5, 2.2]
Tree-Ring1.8% (11/600) [1.0, 3.3]
Gaussian Shading0.3% (2/600) [0.1, 1.2]

FPR for human written text on ELF-L:

MethodFPR, FP/600 [95% Wilson CI, %]
COMPASS0.3% (2/600) [0.1, 1.2]
KGW0.0% (0/600) [0.0, 0.6]
Unigram0.3% (2/600) [0.1, 1.2]
GumbelSoft0.0% (0/600) [0.0, 0.6]
TextSeal0.0% (0/600) [0.0, 0.6]
SynthID0.0% (0/600) [0.0, 0.6]
DiPmark0.5% (3/600) [0.2, 1.5]
Tree-Ring0.0% (0/600) [0.0, 0.6]
Gaussian Shading4.3% (26/600) [3.0, 6.3]

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the rebuttal

Dear Reviewers and Area Chair,

COMPASS is the first watermarking scheme for continuous diffusion language models with a design well suited to the continuous denoising process (Reviewer NVNf) and a practical answer to the near-one-hot output distributions that limit token-level methods (Reviewer 6BCk). The detection works from the final text alone, without access to the generation trajectory or initial noise (Reviewers ejgK & 6BCk). The empirical evidence is strong: COMPASS clearly outperforms the adapted baselines, reaching 94.2% TPR on Cola and 100% on ELF-L, stays robust to edits and paraphrasing (Reviewers 6BCk & ejgK), and comes with ablations that clarify its trade-offs (Reviewers 6BCk). Finally, COMPASS is a training-free method with negligible overhead (0.34% and 1.63%) and no extra network evaluations (Reviewers 6BCk & NVNf).

We summarize the main revisions below.

  1. We compare COMPASS with prior representation steering methods and explain how COMPASS’ keyed partition of latent space serves watermarking. The revised paper explains why the related work by Bagchi et al. for discrete diffusion watermark does not directly transfer to continuous diffusion (Appendix S).
  2. We provide additional ablation studies for our method and compare: random and nearest target selection, spherical and normalized linear interpolation, and sensitivity to guidance scale, denoising steps, and solver choice (Appendices J.6–J.8).
  3. Calibration and false positives (6BCk, ejgK, NVNf). We report false positives on original C4 continuations and on unwatermarked generations subjected to all 11 text edits. With the clean-calibrated threshold held fixed, the largest observed COMPASS false-positive rates across these edits were 3.5% on Cola and 3.7% on ELF-L and comparable to other watermarks (e.g., 3.3% and 5.2%, respectively). We also evaluate different continuation lengths and watermark keys, and document how the ELF-L statistical validation keeps segments from each full generation together (Appendices H, I, and K.2).
  4. Text quality and paraphrase robustness (6BCk, ejgK, NVNf). Table 3 now reports the quality costs of the strengthened baselines alongside their detection rates, showing that matching COMPASS’s clean detection forces baselines to incur substantially higher degradation. COMPASS has the lowest observed mean degradation under all three model judges. We also report paired quality scores, signed perplexity changes, a MAUVE comparison, and more watermarked ELF-L examples at the main watermark setting (Appendices N, O, and Q).
  5. We report the detection of watermarks in latents before decoding into text, showing that COMPASS survives re-encoding better than baselines adapted from image diffusion models.
  6. We evaluate removal, forgery, and wrong-key matches. Across the tested regeneration settings with access to the base model and without verifier feedback.
  7. We introduce a multi user setup for COMPASS achieving 95.10% joint detection and correct user attribution for 2,048 users (Appendix E).

Once again, we thank the Reviewers for their careful reading and constructive feedback, which has helped us improve our work. We have done our best to address the concerns through detailed explanations and additional experiments. We are also happy to address 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.