acceptodds
Under review as a conference paper at ICLR 2027

CoSignature: Cross-Modality Watermarking with Verifiable Binding

Abstract

The rapid deployment of multimodal models has enabled the seamless generation of interleaved text and images. However, existing watermarking schemes treat these modalities in isolation, leaving systems vulnerable to misinformation campaigns where malicious actors manipulate context or swap single-modality content. To address this, we propose CoSignature, a cross-modality watermarking scheme that establishes a verifiable binding between visual and textual outputs via a shared content ID, which we hope to embed into both modalities. To overcome the limited capacity of text watermarking, we employ a novel “zero-bit" binding mechanism: rather than embedding the content ID as a payload, we use it to seed the text generation algorithm itself, converting binding verification into a statistical detection task that disambiguates the true content ID from noisy candidates retrieved by the image decoder. We provide theoretical guarantee that our binding error probability decays exponentially with text length. Empirical evaluation on four distinct architectures—native unified models (Janus-Pro-7B), composite pipelines (Qwen3-8B + Qwen-Image), Qwen3-VL-8B, and BAGEL-7B-MoT—demonstrates the efficacy of our approach, achieving binding accuracy on clean data, accuracy in detecting swap attacks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
cross-modality watermark, AI provenance, text and image binding watermark

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_Zn4KOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem of verifying whether AI-generated images and their accompanying text were generated together. Existing watermarking methods typically verify image and text provenance independently, but cannot establish a reliable connection between the two modalities, making them vulnerable to content-swapping attacks.

To address this problem, the authors propose CoSignature, a cross-modal watermarking framework that establishes verifiable image-text binding through a shared Content ID. The framework uses a dual-key strategy in which the Content ID is embedded directly into the image watermark, while the same ID is used to seed a statistical text watermark alongside a global System Key. During verification, a joint decoder combines Soft-Ordered Statistics Decoding (Soft-OSD), CRC-based candidate filtering, and a Text Oracle to recover the correct Content ID from noisy image predictions and verify its consistency with the accompanying text. The authors also provide a theoretical analysis showing that the Text Oracle's candidate-selection error decreases exponentially with text length under specified assumptions.

The framework is evaluated on 1,000 generated image-text pairs across four multimodal model pipelines: Janus-Pro-7B, Qwen3-8B with Qwen-Image, Qwen3-VL-8B, and BAGEL-7B-MoT. The authors report binding detection rates of 96.7%–99.0% at a 1% false-positive rate and average swap detection accuracies of 95.6%–99.3% on clean data. Additional experiments investigate robustness to image distortions, text paraphrasing, decoding configurations, and watermarking components, demonstrating improvements in Content ID recovery over image-only decoding under several evaluated conditions.

Critical strengths
  1. Important and well-motivated problem: The paper addresses a relevant limitation of existing multimodal watermarking systems: independently verifying image and text provenance does not establish whether they were generated together. The proposed cross-modal binding framework provides a useful approach to detecting mismatched image-text pairs and verifying their shared generation provenance.

  2. Novel cross-modal binding mechanism: The proposed dual-key watermarking strategy is an interesting contribution. Instead of embedding a long Content ID directly into text, the authors use it to seed a statistical text watermark. Combining this mechanism with image-based ID recovery enables binding verification without requiring multi-bit payload extraction from the text.

  3. Effective joint decoding framework: The integration of Soft-Ordered Statistics Decoding (Soft-OSD), CRC filtering, and a Text Oracle is a meaningful technical strength. By using statistical evidence from the text to distinguish candidate image IDs, the method improves recovery under several image distortions. The ablations further demonstrate the benefits of CRC filtering and increased search depth.

  4. Theoretical analysis and informative ablations: The paper provides a theoretical error bound showing that the Text Oracle's candidate-selection error decreases exponentially with text length under stated assumptions. Additional experiments investigate text length, paraphrasing, CRC filtering, decoding depth, and the trade-off between checksum redundancy and recoverable Content ID capacity. These analyses help explain both the effectiveness and limitations of the proposed framework.

Critical weaknesses
  1. Limited direct comparison with existing cross-modal watermarking approaches: Although the paper discusses related work such as VLA-Mark and Jovanović et al. (2025), it does not provide direct quantitative comparisons with these approaches. The experimental evaluation primarily focuses on image-only decoding, alternative decoding configurations, and an idealized channel-code baseline. While differences in the underlying tasks may complicate direct comparisons, a more detailed evaluation or justification of non-comparability would help establish the proposed framework's advantages over existing cross-modal approaches.

  2. Incomplete robustness evaluation and sensitivity to severe image distortions: The authors evaluate several image-processing attacks, including resizing, cropping, JPEG compression, Gaussian blur, noise, and photometric changes. However, rotation, perspective distortion, and salt-and-pepper noise are not included. Furthermore, the results show substantial performance degradation under certain severe distortions. For example, CoSignature achieves only 0.0% Content ID recovery under Gaussian noise (σ=10) and 8.1% under JPEG compression (Q=50) on Janus-Pro. These findings suggest that the demonstrated robustness is strongly dependent on the underlying image watermarker and attack severity. Additional experiments would help clarify the framework's robustness boundaries.

  3. Lack of evaluation against targeted semantic-changing text manipulations: The paper evaluates synonym replacement, moderate paraphrasing, and heavy rewriting, with instructions intended to preserve the original meaning. However, it does not systematically investigate targeted semantic modifications, such as replacing entities, objects, actions, or factual statements. Since CoSignature verifies watermark-based provenance rather than semantic consistency, an attacker might alter the meaning of text while retaining sufficient statistical watermark evidence for binding verification. Evaluating such scenarios would help clarify the security guarantees and limitations of the framework, particularly given its motivation around misinformation.

  4. Limited diversity of the evaluation dataset: The experiments use 1,000 generated image-text pairs based on ImageNet-1K class labels and a common prompt template. Although this provides a controlled evaluation, it does not fully reflect the diversity of real-world multimodal documents, such as news articles, reports with multiple images, or content involving complex relationships between text and visual elements. Evaluation on more diverse prompts, text lengths, and document structures would strengthen the evidence for practical generalization.

  5. Dependence on image candidate recovery and sufficient text length: The theoretical analysis assumes that the correct Content ID is present in the candidate set generated by the image decoder. Consequently, the Text Oracle cannot recover the correct identifier when severe distortions eliminate it from that set. Furthermore, binding verification depends on accumulating sufficient statistical evidence from the text. The authors report approximately 80% binding detection at 300 tokens on the Qwen pipeline, increasing to 97.4% at 500 tokens. These dependencies limit the demonstrated applicability to short captions, brief descriptions, and severely distorted images.

  6. Insufficient clarification of the underlying image watermarking configurations: CoSignature relies on existing image watermarking systems, including VideoSeal and PixelSeal, but the manuscript does not provide a detailed account of the training configurations and augmentation policies associated with the checkpoints used. Since the robustness of the joint framework depends partly on these underlying models, clarifying their pretrained configurations and whether the evaluation distortions overlap with their training augmentations would improve reproducibility and interpretation of the results.

  7. Presentation and experimental reporting inconsistencies: Several details require clarification. Section 4 contains an unresolved citation placeholder for Qwen3-VL-8B at line 350. The main text and Appendix C.4 also provide different descriptions of the model used to generate paraphrasing attacks. Furthermore, clean binding results differ between Table 2 and Table 7, including 97.4% versus 97.9% for the Qwen pipeline, 99.0% versus 99.9% for Qwen3-VL, and 96.7% versus 99.1% for BAGEL. These differences may arise from distinct experimental subsets or protocols, but the manuscript should explicitly explain them to ensure consistency and reproducibility.

Questions for authors
  1. Comparison with existing cross-modal watermarking methods: The paper discusses related approaches, including VLA-Mark and Jovanović et al. (2025), but does not provide direct quantitative comparisons. Could the authors clarify the methodological advantages of CoSignature over these approaches and provide comparisons under a common evaluation protocol where feasible? If direct comparisons are not possible because of differences in their objectives, could these differences be explained more explicitly?

  2. Robustness against additional image transformations: The authors evaluate resizing, cropping, JPEG compression, Gaussian blur, noise, and photometric changes. However, rotation, perspective distortion, and salt-and-pepper noise are not considered. Could the authors evaluate these additional attacks? Furthermore, given that Content ID recovery drops to 0.0% under Gaussian noise (σ=10) and 8.1% under JPEG compression (Q=50) on Janus-Pro, could they explain the practical robustness limitations of CoSignature and how these failures might be mitigated?

  3. Semantic-changing text manipulations: The paraphrasing experiments primarily consider meaning-preserving transformations. What happens when an attacker deliberately changes the semantic meaning of the text while preserving much of its original token structure? For example, replacing "The dog is sitting on the bench" with "The cat is sitting on the bench" changes the factual content but may retain part of the statistical watermark. Could the authors evaluate such targeted semantic modifications and clarify whether CoSignature can detect them? This distinction seems particularly important given the paper's motivation around misinformation.

  4. Generalization to diverse image-text datasets: The experimental evaluation uses 1,000 generated image-text pairs based on ImageNet-1K class labels and a common prompt template. How well would CoSignature generalize to more diverse multimodal datasets involving news articles, detailed descriptions, complex relationships, and multiple images associated with a single document? Could the authors provide additional experiments using different prompt structures or realistic multimodal content?

  5. Dependence on text length: The authors report approximately 80% binding detection at 300 tokens on the Qwen pipeline, increasing to 97.4% at 500 tokens. How would CoSignature perform in practical scenarios involving short captions, brief social media posts, or image descriptions containing fewer than 100 tokens? Could the authors discuss the minimum text length required for reliable binding verification at the specified false-positive rate?

  6. Relationship between theoretical guarantees and practical recovery: Theorem 3.2 establishes exponentially decreasing Text Oracle error under the assumption that the correct Content ID is included in the candidate set. However, severe image distortions may prevent the correct ID from appearing among the candidates. Could the authors quantify candidate-set recall across different attack severities and distinguish failures caused by image candidate generation from failures caused by the Text Oracle? This would clarify how the theoretical guarantees relate to end-to-end binding accuracy.

  7. Consistency of experimental results and model descriptions: Several details appear inconsistent across the manuscript. For example, the clean binding rates differ between Table 2 and Table 7 (97.4% versus 97.9% for Qwen, 99.0% versus 99.9% for Qwen3-VL, and 96.7% versus 99.1% for BAGEL). Additionally, the paraphrasing adversary is described differently in Section 4.1 and Appendix C.4, and line 350 contains an unresolved citation placeholder for Qwen3-VL-8B. Could the authors clarify these discrepancies and specify whether different experimental configurations or subsets explain the numerical differences?

  8. Representative paraphrasing examples: Appendix C.4 provides prompts for synonym replacement, moderate paraphrasing, and heavy rewriting, along with token-overlap statistics. Could the authors include representative original and modified text examples for each category? This would make it easier to evaluate the nature of the attacks and interpret the reported robustness results.

Reviewer_KR9LOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The submission proposes a joint image and text watermarking scheme. The scheme first hides the text watermarking key in the image. Then it generates and watermarks the text with this secret key. Moreover, the text watermarking also has a system key in case the image is not present. This allows verification that the text and image were generated jointly (same user, same prompt), or that the image alone was generated, or that the text alone was generated, or that the text and image were generated but separately.

Critical strengths

S1: Originality 1

Very few watermarking papers address joint watermark embedding across multiple modalities. That being said, this topic is quite niche. I am not convinced of any practicality. If a pair of image-text is detected as generated but not jointly, or jointly generated, why does that matter to the final user/reader?

S2: Originality 2

Watermarking generated text with a double key is novel. I regret that this innovation isn't described in more detail. I feel like much space is devoted to justifying the joint watermarking of image and text, whereas the key invention (as far as I am concerned) lies in this new text watermarking.

Critical weaknesses

I am only interested in the proposed text watermarking scheme. The image watermarking is standard. The protocol is very simple. Binding the watermarking secret key to a ContentID is standard industry practice.

W1 Not universal

The proposal is a clever adaptation of the KGW Green/Red watermarking scheme. But, it cannot be extended to other text watermarking schemes. The authors claim that the appendix also covers the Aaronson scheme. This is not exactly true. The reader is left with ideas for adapting Aaronson that are neither fully detailed (especially detection) nor fully tested. See app. C6 and Tab. 12.

W2 Z-statistic

The initial KGW allows assessing the probability of false alarm (see paper « Three Bricks to Consolidate Watermarks for Large Language Models») by replacing the Z-statistics with the true binomial variables (and avoiding the use of the same token key - ie. not scoring text repetitions). We are clearly losing this advantage here. We are back with a Z-statistic which only offers an estimate of the false positive.

W3 Th. 3.2

Theorem 3.2 is not precise enough.

  • What is k⋆k^\star? (I suppose it is kcontk_{cont})
  • What is a detection error? Since the authors consider 3 detection scenarios, there are several ways to be wrong.
  • What is the value of Δmin\Delta_{min}? How does it relate to γ,psys,δ\gamma, p_{sys}, \delta … I understand this is an assumption (there exists Δmin>0\Delta_{min}>0), but still, as a watermark designer, I would like to know how this quantity is impacted by the parameters of the watermarking scheme.

Moreover, I do not understand the proof of the theorem. It bounds the probability that a wrong key produces a score greater than the true key. But this is not a detection error. In practice, scores (with the true key, resp. false keys) are compared to a threshold and two probabilities are bounded (false negative, resp. false positive). The factor (M−1)(M-1) implicitly means that the true key is in the candidate set, yet another key yields a bigger score. This is not a decoding error if indeed the text is truly bound to the image.

Questions for authors
  • Why are 224 bits allocated for the ContentID? This sounds way too much.

  • I would like to see a fair comparison between the double-key and the original single-key watermarking scheme in terms of quality and detectability.

1 comment

Thank you for the detailed comments on the text construction and its statistical interpretation.

W1 — Modularity and the evaluated Gumbel construction. Appendix C.7 identifies and evaluates the max-superposition Gumbel variant. On Qwen3-VL (n = 300), Table 12 reports 95.3% binding TPR and 93.3% independent text TPR. Thus, the paper includes an empirical non-KGW instantiation. We will distinguish this evaluated construction from the alternatives discussed in C.6 and clarify its sampling, scoring, repeated-context handling, and calibration protocol.

Each instantiation requires a compatible detector and its own analysis. Our contribution combines separate source and instance signals with image-derived candidate IDs and text-assisted disambiguation, retaining fixed-key source verification alongside verification of a specific image–text association.

W2 — Calibration of the weighted statistic. The reported z-score uses a Gaussian approximation. An exact weighted null is available analytically under an independent random-partition model.

Consider a score retaining one position per distinct context, selected without consulting the tested content key. Condition on the observed text, system-green memberships, and image-candidate construction. If a false candidate’s partitions remain uniform and independent across retained contexts under this conditioning, then the score has distribution E(k) = 2U + V, where U ∼ Binomial(n₁, γ) and V ∼ Binomial(n₀, γ) are independent; n₁ and n₀ count retained system-green and other positions.

The convolution provides exact inclusive tail probabilities under these assumptions. Bonferroni at α/M controls false acceptance across M candidates when these p-values remain valid conditional on candidate construction; independence between candidate tests is unnecessary. Removing identical context–token pairs alone is insufficient when different tokens reuse a fixed-size partition.

We will distinguish this analytical construction from the reported Gaussian-calibrated results and make the independence assumptions explicit.

W3 — Meaning of the theorem’s error event. The token score is Xₜ(k) = (1 + Aₜ)Bₜ(k), where Aₜ and Bₜ(k) indicate system-green and candidate-content-green membership. The margin is the conditional expected score difference between the true and false candidates. The theorem bounds candidate-selection error, conditional on the true ID being in the candidate set.

The margin depends on token probabilities, mask overlap, temperature, and top-p truncation; positive bias alone does not guarantee it. We will replace that assertion with an explicit assumption that accumulated conditional drift exceeds a deterministic positive bound almost surely, in the experiment including candidate construction. Please also refer to our response to Reviewer vVG3.

Content-ID length. The 224-bit ID follows from the selected 256-bit image payload with 32 checksum bits. It is a configuration choice. Table 15 already studies reallocating the fixed payload between ID and checksum bits.

Single-key comparison. Our objective is to support both direct text-only source verification with a fixed system key and instance-specific image–text binding. Standard single-key KGW with a fixed key supports source detection but does not distinguish generation instances. Using a per-instance key enables verification against a candidate ID, but text-only source verification then requires that ID or a search over candidate keys. Our dual-key construction retains both verification paths.

Reviewer_5sLpOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a novel framework called "co-signature", which watermarks AI-generated images and their accompanying texts in a way that the decoder/verifier can validate that the two were generated together. This enables us to detect cases where someone swaps or modifies one of the modalities. The idea is to generate a 224-bit content identifier which is added to image and also used for generating the corresponding text watermarking component as a seed. This makes the validation a simple yes/no statistical evaluation. The experiments on four image-text models shows high detection accuracy at low 1% false-positive rates.

Critical strengths
  1. Well-motivated problem with vast practical implications for misinformation, fraud and other related problems on the rise with the introduction of highly realistic AI content creation solutions.

  2. Simple and intuitive solution. Seeding the text watermark with the image ID avoids hiding 224 bits in text. The method works with any existing watermark (tested with 3+ approaches).

  3. Extensive experimentation. 4 architectures, vast image distortions, 3 paraphrase levels and combined attacks.

Critical weaknesses
  1. Missing a simple but important baseline. The platform already keeps a record for each content (id and message pairs), so it could store a fingerprint of the text and compare at verification time. No text watermark is necessary.

  2. The approach works primarily with long text. Binding works about 80% of the time at 300 tokens and 97% at 500. Short captions and social media posts (most real image-text pairs) are not covered.

  3. The text quality loss is understated. Qwen3-VL perplexity rises from 3.65 to 4.60. Janus-Pro alignment drops from 26.6 to 23.5. No human evaluation.

  4. Dataset limitations. 1k prompts, one template about ImageNet classes (not real articles).

Questions for authors

The paper is well-motivated and written. If the mentioned weaknesses are addressed, the paper may increase its impact (and also would increase the rating accordingly). There are other relatively minor questions below.

  1. Is robustness headline coming from cross-modal watermarking capability? At the default 32-bit checksum, the text contributes nothing. The gain comes from the image decoder saving 24 bits from checksum. Is this correct?

  2. The theorem assumes every token adds positive signal. Is this assumption always true?

  3. Binding drops to 64–86% after a heavy paraphrase. How should the reader interpret this loss? Is it acceptable?

  4. Overclaiming and polish. "First formal definition" and "cryptographic" overstate the contribution.

Tthere is a missing citation "(?)".

1 comment

Thank you for the concrete suggestions on alternative verification mechanisms and the attribution of our empirical gains.

Stored text fingerprints. An image ID paired with a stored text fingerprint is a valid alternative when that reference is retained. Our verifier in Eqs. (2)–(5) tests the observed text directly against candidate IDs, without retrieving the original text or a stored fingerprint. This remains compatible with maintaining a registry for other provenance information. Exact hashes verify unchanged or canonically normalized text but do not generally tolerate paraphrasing.

What the Text Oracle contributes. The recovery and capacity effects should be distinguished. In the Qwen3-VL ablation in Appendix C.8 (n=300), the default 32-bit checksum yields essentially identical Content-ID recovery with and without the Text Oracle. The recovery gain over hard decoding in this setting comes from CRC-filtered soft search. Text still performs the separate binding test: a uniquely recovered image ID alone does not establish that the supplied text carries that ID’s watermark.

Table 15 separately isolates candidate disambiguation on Qwen3-VL (n=200). With an 8-bit check, CRC-only ID recovery is 0.0%, 0.0%, and 0.5% on clean, JPEG Q=50, and crop 80% inputs; adding the Text Oracle gives 98.0%, 81.0%, and 62.5%, achieving approximately 97.6–98.0% of the corresponding candidate-recall ceilings. Reducing the checksum from 32 to 8 bits frees 24 bits of payload capacity through text-assisted disambiguation. This capacity benefit is distinct from the recovery gain at the default 32-bit setting.

Text length, dataset, and quality. Figure 2 evaluates 100–500-token outputs specifically on Qwen3-8B + Qwen-Image (Qwen Pipeline), reaching 97.4% binding TPR at 500 tokens. We will make the model and length associated with each reported result explicit. We will replace the blanket “negligible degradation” description with the corresponding model-specific trade-offs.

As stated in Section 4, we follow the ImageNet-derived prompting protocol of Jovanović et al. (2025, Appendix H). Our evaluation requires freshly generated text and images watermarked under a shared Content ID. ImageNet supplies prompt topics; the evaluated image–text pairs are newly generated by each pipeline. We will clarify this distinction between the prompt source and the generated evaluation outputs.

Heavy rewriting and theoretical scope. Under heavy rewriting, binding TPR is 76.2%, 63.8%, 85.6%, and 65.8% for Janus-Pro, Qwen Pipeline, Qwen3-VL, and BAGEL+PixelSeal, respectively (Table 7). These quantify robustness to aggressive rewriting. A failed binding test leaves the association unverified, while surviving source evidence can be reported separately. Whether these rates suffice depends on the application’s tolerance for unverified pairs.

Theorem 3.2 assumes a positive conditional expected margin. We agree that positive bias alone does not establish that assumption. We will clarify this point and present the accumulated-drift generalization (see our response to Reviewer vVG3).

Reviewer_vVG3OpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CoSignature verifies whether generated text and images share the same generation identifier, extending provenance verification beyond independent modality detection.

Critical strengths
  1. The paper addresses a meaningful limitation of independent watermarking: verifying each modality’s source does not establish whether the two outputs belong to the same generation. The shared Content ID and dual-key construction provide a useful approach to this problem.

  2. The combination of Soft-OSD and a Text Oracle is technically interesting. The experiments distinguish the contribution of soft-decision image decoding from cross-modal candidate selection and show that text information can compensate for reduced checksum redundancy. Evaluations across several architectures and image watermarkers further support the framework’s modularity.

Critical weaknesses
  1. Practical validation remains limited by the use of ImageNet-derived prompts and predominantly 500-token outputs. Although the paper already evaluates shorter text, paraphrasing, and combined attacks, these results expose important operating limits. For example, binding verification falls to 33.3% under heavy rewriting combined with JPEG Q=75 on Qwen3-VL. This concerns binding rather than complete provenance loss, but it limits the strength of the practical binding claim. Furthermore, shared provenance does not itself establish semantic consistency or factual correctness, so the broader misinformation-related claims should be qualified.

  2. The justification of Theorem 3.2’s signal-margin assumption is incomplete. The concentration argument assumes a deterministic, uniformly positive conditional expected score margin over every false candidate. A positive logit bias alone does not establish this condition across all generation contexts, particularly with highly concentrated token distributions or top-p truncation. The conditional theorem may remain valid, but its applicability to the implemented sampling procedure requires stronger justification.

Questions for authors

Q1. Beyond the existing paraphrasing and combined-attack experiments, could the authors evaluate realistic multimodal articles and short image captions? Please clarify the text lengths and transformation levels over which binding verification remains reliable.

Q2. Could the authors formally establish sufficient conditions for the uniformly positive signal margin in Theorem 3.2 under the implemented temperature and top-p sampling procedure? If a uniform per-step bound is too strong, could the analysis instead use an accumulated or average signal-margin condition?

1 comment

Signal margin and Theorem 3.2. we agree that positive logit bias alone does not establish a uniformly positive margin, particularly under concentrated token distributions or top-p truncation. We will clarify this distinction and generalize the analysis to an accumulated predictable-drift condition.

Condition on the prompt, keys, and image-side information defining a candidate set K\mathcal{K} of size MM containing the true key k∗k^\ast. All probabilities and expectations below use this conditional law, and Ft\mathcal{F}_t contains the initial information and the first tt text tokens. For each false candidate jj, let Yt(j)=Xt(k∗)−Xt(j)Y_t(j)=X_t(k^\ast)-X_t(j). Assume there exists a deterministic aN>0a_N>0 such that, for every false candidate,

\ge a_N \quad\text{almost surely}.$$ Then the candidate-selection error satisfies $$\Pr(\widehat{k}\ne k^\ast) \le \min\left\\{1,(M-1)\exp\left(-\frac{a_N^2}{8N}\right)\right\\},$$ where $\widehat{k}$ is the maximum-score candidate in Eq. (4), before threshold-based binding acceptance. To see this, center each $Y_t(j)$ by its conditional expectation. Since $Y_t(j)\in[-2,2]$, the resulting martingale differences have conditional range length at most $4$. Conditional Hoeffding–Azuma and a union bound over the $M-1$ false candidates give the stated bound. This formulation permits zero or negative margins at individual positions. For fixed $M$, the bound decreases exponentially when $a_N\ge cN$ for a constant $c>0$. The drift assumption must hold under the conditional law above, including the effects of temperature and top-p sampling; positive bias alone does not guarantee it. We will also explicitly distinguish this conditional candidate-selection bound from threshold-based binding FPR/FNR. **Evaluation protocol.** We follow the ImageNet-derived prompting protocol used for interleaved text–image generation in Jovanović et al. (2025, Appendix H), as stated in Section 4. Our evaluation requires freshly generated text and images watermarked under a shared Content ID, so the inputs are generation prompts; pre-existing image–text pairs cannot directly substitute for these generated outputs. ImageNet-1K supplies the class labels used to construct prompts, while both modalities are generated anew by each evaluated pipeline. This provides a common, reproducible protocol for evaluating binding across architectures. We will clarify this role of ImageNet in the experimental description.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.