acceptodds
Under review as a conference paper at ICLR 2027

ReDiffMark: Post-Hoc Text Watermarking by Local Re-Diffusion

Abstract

Post-hoc watermarking embeds detectable signals into existing text, a capability crucial for closed-API outputs, RAG-retrieved passages, and published documents. However, such methods must minimally alter the original text while remaining computationally efficient for large-scale deployment. Existing sparse-editing schemes preserve the text but suffer from poor signal stability and low efficiency due to rule-based byte signaling; conversely, full-rewriting schemes carry strong signals but rewrite most tokens, leading to significant text alteration and slow speeds. We observe that discrete diffusion language models (dLLMs) can bridge this gap. To this end, we propose **ReDiffMark**, a watermarking paradigm based on *localized re-diffusion* for text that is stored or served verbatim or with light edits. We address challenges regarding available contextual entropy and the difficulty of identifying sparse edit locations through entropy-gated editing with key-seeded exponential-minimum (Gumbel) sampling, employing a *replay-based*, surprisal-weighted detector that re-runs the same dLLM on the received text under a calibrated null hypothesis. Extensive experiments across dLLM families, text sources, and passage lengths show that **ReDiffMark** attains high single-passage detection at low edit rates with a constant two-forward-pass cost.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Watermark, Diffusion, dLLMs, Post-hoc watermark

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_8vTgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies post-hoc watermarking, which adds a watermark to text that already exists. Sparse-editing schemes replace a small portion of the tokens, but each edit carries only a weak signal. Full-rewriting schemes regenerate the passage with a watermarked LLM. They carry a strong signal, but they are slow and can change the text in arbitrary ways instead of modifying only a few tokens. The paper proposes ReDiffMark, which follows the sparse-editing idea but uses a discrete (masked) diffusion language model to make the edits. A secret key selects the positions to replace, and these positions are masked. For each masked position, the diffusion model gives a distribution over tokens, and key-seeded Gumbel-max sampling selects the token that carries the watermark. A position is changed only where the model is uncertain enough. The detector re-runs the same steps on the received text and tests the keyed scores against a calibrated null. On clean 200-token news passages, the authors report detection close to that of full-rewriting schemes (TPR at 1% FPR of 0.94, against 0.95 to 1.00) while changing about 8% of tokens instead of 65–78%, at a much lower latency.

Critical strengths
  1. I think designing more efficient post-hoc watermarking is an important research question, with many applications. The method reaches a useful operating point: on clean news passages, TPR@1% is 0.94 while about 8% of tokens change, at 0.37 s per passage (Table 1).

  2. The paper includes sufficient baselines, covering both sparse editing and full rewriting, and the comparison between them is quite fair. The controlled ablation of the scoring rule is also informative.

Critical weaknesses

1. Apart from changing fewer tokens, the benefit over full rewriting is not obvious.

  • The clear gain is that far fewer tokens change. But one of the paper's main arguments against rewriting is its cost. The proposed method does embed faster, but its total compute is similar to that of the rewriting baselines, and its detection is slower than for most of them, because it must re-run the diffusion model (Table 11). The fast key-only detector loses much of the detection power. Overall, the cost remains similar.
  • The claim to outperform all baselines rests on the latency term of the composite score. Without that term, PostMark ranks first.
  • The method is also less robust than full rewriting, under both word substitution and back-translation, although it is much more robust than the sparse-editing baselines.
  • The diffusion model is not what carries the signal: a RoBERTa infiller gives about the same detection, though with less fluent edits (Table 17). The method is therefore close to earlier sparse substitution schemes (Yang et al., 2022; Yoo et al., 2023), with Gumbel sampling (Kuditipudi et al., 2024) as the main change.

2. The fidelity metrics do not show that the watermarked text stays useful.

  • The paper does measure semantic similarity, but with a sentence-embedding score averaged over passages. Such scores can stay high when a single word changes the meaning, such as a negation or a modal verb. The paper's own first example, a health-advice passage, changes "will help" to "may help", the same kind of edit the paper criticizes in a baseline. The paper also notes that edits may change nuance.
  • This matters more here because the signal depends on edits that an external language model finds unlikely: when a verifier keeps only the edits that barely change perplexity, detection largely collapses (Table 6). The edits that carry the watermark are the ones most likely to shift meaning.
  • The right test is whether the watermarked text still serves its purpose. For example, when post-hoc watermarks are used to detect benchmark contamination, models should rank the same on the original and the watermarked benchmark. For RAG documents, a system should still give the same answers. The paper reports no such evaluation.
Questions for authors
  1. Does the watermarked text stay useful for its intended purpose? For example, if a benchmark is watermarked post hoc to detect contamination, do models keep the same accuracy and ranking on the watermarked version? For RAG documents, do the answers stay correct? If utility is preserved, my concern about the fidelity metrics is resolved.

  2. Could you report results for a targeted removal attack and for spoofing? For uses such as contamination detection, adversarial robustness may matter less, but readers still need to know the limits. For example, an attacker could flag the tokens that an external language model finds least likely and refill them without the key. Or the attacker could collect many watermarked passages with their originals and learn which tokens the key favors at each position, as in watermark stealing (Jovanović et al., 2024).

  3. How do embedding and detection costs compare with the rewriting baselines under batched serving? The composite score uses single-passage latency, which favors a two-pass method, while detection requires re-running the diffusion model. This comparison would show whether the cost advantage holds in practice.

minor:

  • The default position keys break after a single insertion or deletion, which conflicts with the goal of surviving light edits. Seeding keys from a wider context of unedited tokens would avoid this. The detector should then score each repeated (context, token) pair only once, to keep the p-values valid (Fernandez et al., 2023).
  • Please state whether detection runs on decoded strings re-tokenized by the diffusion model or on stored token ids, and report detection after word-level edits on decoded text. With position keys, a single tokenization mismatch shifts all later indices.
  • The RAG-attribution use case needs detection after an LLM rewrites the passage. After one paraphrase, every compared scheme, this one included, loses most of its signal. I would state the claim as attribution of verbatim or lightly edited copies.
1 comment
AuthorsOpenReview ↗

Kind Request for Your Feedback

Dear Reviewer 8vTg,

Thank you for your constructive feedback. We believe we have addressed your concerns to a large extent in our response.

In short, we added downstream utility results-RAG QA and retrieval stability-to resolve your fidelity concern. We also added batched-serving costs on efficiency concerns, and the targeted-removal and watermark-stealing experiments you asked for.

Given that the deadline is approaching, could you please take a moment to review our updates? If you have any further questions, we would be more than happy to provide additional clarification.

Thank you again for your time and effort in reviewing our paper. We hope these updates help clarify the manuscript.

Best regards,

The Authors

Reviewer_48zoOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

ReDiffMark watermarks existing text through two passes of local masked refilling with keyed sampling. A replay detector uses entropy filtering and surprisal weights. Experiments assess detection, edit rate, fidelity, runtime, and robustness, with component ablations.

Critical strengths
  1. The method adds a detectable watermark to existing text with relatively few edits and low reported embedding latency.
  2. Matched-position refilling and detector ablations help identify which components produce the gains.
  3. The paper reports paraphrase failures and limitations of its tail approximation, giving a useful account of where the method works.
Critical weaknesses
  1. The left-context variant needs a clearer null-distribution argument. If repeated predecessor/token pairs reuse the same random field, their scores are correlated, whereas Proposition 1 relies on independent position-specific randomness. The PRF inputs and treatment of repeated scoring units need to be explicit.
  2. The precision of the 0.1% FPR results is difficult to assess with 1,000 negative passages per key. It is also unclear whether the five keys reuse the same documents. Per-key false-alarm intervals and an account of document reuse are needed to interpret the pooled results.
  3. The editing gate does not establish semantic equivalence. Top-k/nucleus membership can admit changes in modality or factual meaning, as illustrated by the “will” to “may” substitution. A factual-change audit would support the preservation claim more directly than embedding similarity and entity retention.
Questions for authors
  1. What are the exact PRF inputs in left-context mode, and how does the null account for repeated scoring units?
  2. Are negative documents reused across keys, and what per-key confidence intervals support the 0.1% FPR results?
  3. How often do edits change facts, modality, or negation in a held-out sample?
1 comment
AuthorsOpenReview ↗

Kind Request for Your Feedback

Dear Reviewer 48zo,

Thank you for your constructive feedback. We believe we have addressed your concerns to a large extent in our response.

In our response, we have tried to address each of your concerns: the exact left-context PRF inputs, and a de-duplication rule under which Proposition 1 holds exactly (W1); per-key false-positive counts with Clopper–Pearson intervals, together with an account of the shared negative set (W2); a per-edit NLI audit and a downstream QA measure factual change directly (W3).

Given that the deadline is approaching, could you please take a moment to review our updates? If you have any further questions, we would be more than happy to provide additional clarification.

Thank you again for your time and effort in reviewing our paper. We hope these updates help clarify the manuscript.

Best regards,

The Authors

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

ReDiffMark uses two diffusion infilling passes to watermark selected positions in existing text, then replays the model for detection. It achieves a useful speed and fidelity trade-off on prose. The strongest open questions concern ordinary text handling, repeated key use, and semantic preservation.

Critical strengths
  1. Local infilling is a good fit for watermarking finished text with few edits.
  2. The replay and weighting ablations make clear where the detection gain comes from.
  3. The evaluation reports empirical null calibration, fixed-threshold attack results, and explicit failures on low-entropy text, which helps define the method's useful scope.
Critical weaknesses
  1. Absolute-position keys make extraction boundaries part of the protocol. RAG commonly returns a middle span or prepends a heading, even when the underlying prose is unchanged. The left-context variant helps with edits, but random-offset crops and single-token prefix changes are not isolated. These are more relevant to the claimed use case than 10 to 40 percent random edits.
  2. Equation 4 reuses the same position-specific random field across passages under one key. The experiments do not test whether many observed outputs reveal favored tokens or enable spoofing. A multi-document key-reuse test is needed before treating key secrecy as sufficient for provenance.
  3. Candidate membership is not a semantic equivalence check. The qualitative example changes will to may while retaining similarity 0.996. Entity retention and sentence similarity can miss changes in modality or relations. Audit edited spans for contradiction and unsupported factual changes at the operating points used for detection.
Questions for authors

What detection survives random-offset excerpts with one fixed threshold? How does false attribution change when an attacker observes many passages marked with the same key?

1 comment
AuthorsOpenReview ↗

Kind Request for Your Feedback

Dear Reviewer pXpb,

Thank you for your constructive feedback. We believe we have addressed your concerns to a large extent in our response.

We believe we have addressed each of your concerns as fully as we could, with new experiments on random-offset crops and prepended headings (Tables R1–R3), multi-document key reuse and spoofing (Table R4), and NLI-based and end-to-end factuality audits (Tables R5–R7).

Given that the deadline is approaching, could you please take a moment to review our updates? If you have any further questions, we would be more than happy to provide additional clarification.

Thank you again for your time and effort in reviewing our paper. We hope these updates help clarify the manuscript.

Best regards,

The Authors

Comments on the paper

1 comment

Dear Reviewers,

As the discussion deadline (Nov 18, AoE) approaches, please review the authors' rebuttal if you have not done so, and post a follow-up indicating whether your main concerns have been addressed or whether any issues remain.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.