- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes WorldCup, a multi-bit watermarking framework that embeds message bits directly into tournament sampling during autoregressive generation. It uses complementary g-value functions to guide token selection, entropy-aware modulation to adapt watermark strength to the model’s local uncertainty, and confidence-aware decoding to aggregate evidence across tournament layers. The paper evaluates message recovery, detectability, text quality, downstream task performance, decoding efficiency, and robustness under editing, copy-paste, and paraphrasing attacks.
Critical strengths
The paper builds on SynthID-Text’s zero-bit tournament sampling and existing multi-bit watermarking research, and uses tournament sampling and hash- or pseudorandom-function-based g-value generation. On this basis, it proposes complementary g-values conditioned on message bits and a complementary-distribution subtraction mechanism, q − λq̄. It further extends the method to k-bit encoding, incorporates entropy-aware modulation, and proposes a vectorized computation of the tournament distribution and confidence-aware decoding.
Critical weaknesses
-
The authors claim in the introduction section that static logit perturbations interfere with natural generation “especially at higher payloads,” degrading text quality and fluency. However, they provide no evidence that increasing the number of watermark bits causes this effect when watermark strength is held constant. In Table 1, PPL does not systematically increase with bit length, so the results do not show the claimed degradation. The authors should support or qualify this claim.
-
The notation in the core method is ambiguous. Section 3.1 uses m for the message, whose length is b, while Algorithm 1 uses m for the number of tournament layers. The g-function families contain m layer-specific functions, and Figure 1 shows a tournament depth matching the largest subscript of g, suggesting that message length and tournament depth are linked, although the layer count is a hyperparameter independent of message length. The paper also uses boldface and regular-weight g to distinguish a function family from an individual function, a distinction that is visually easy to miss. Together, these ambiguities make the indexing and function selection difficult to follow.
-
Appendix G.1 establishes score-level separation but notes that the distinguishability of the induced token distributions depends on the base distribution and tournament parameters. The paper does not provide this distribution-level analysis. Section 5.1 reports how removing the subtraction of complementary g-value distributions and replacing complementary g-value functions with randomly sampled ones affect decoding accuracy, but does not quantify the statistical distance between the induced distributions q and q̄. The link between score-level separation and the reported decoding gains therefore remains unclear.
Questions for authors
- Could the authors clarify the notation for the message, tournament layers, and g-function families issue mentioned in “weakness” section?
- Can the authors quantify the statistical distance between q and q̄ under complementary and randomly sampled g-value functions (e.g., using symmetric KL divergence or MAE), and analyze how it relates to decoding accuracy?
- The paper’s decision-margin argument for complementary g-value functions appears to address stochasticity during generation and sampling. It does not explain why WorldCup is robust to the evaluated post-generation attacks, which affect watermark detection and message extraction through different mechanisms. Could the authors provide an attack-specific mechanistic analysis and identify which components account for WorldCup’s advantage over the baselines? In addition, could they report the attack-induced drops in AUC or TPR@1% FPR relative to the matched clean condition for WorldCup and the baselines? This would clarify whether better post-attack performance arises because a method performs better before the attack or because it is more resistant to the attack.
4 comments
As the rebuttal deadline approaches, we welcome further discussion, which would help us promptly address any remaining questions. If our responses have addressed your concerns, we hope you will take them into account in your final assessment, and we sincerely appreciate your time and feedback.
Thank you for addressing the questions raised in the review. I would like to keep the rating unchanged.
Thank you for taking the time to consider our responses. We sincerely appreciate your thoughtful feedback and support for our work. We wish you all the best!
Response to Reviewer X5is
We thank the reviewer for careful reading and constructive suggestions. We address each concern below.
W1: Payload and text quality
We agree Table 1 shows no consistent increase in perplexity with message length. The phrase “especially at higher payloads” left the role of watermark strength implicit. At a fixed generation length, a longer message has fewer tokens per bit on average. Maintaining or improving bit accuracy may then require stronger logit perturbations, which can reduce fluency and overall text quality. We will make this condition explicit in the introduction.
W2 & Q1: Notation
In Section 3.1, bold m denotes the message, while plain m denotes the tournament depth. The message length b and the depth m are chosen independently. As the reviewer notes, we use bold g for the ordered family of scoring functions, following the convention for vectors, and plain gℓ for the individual function at layer ℓ.
For clarity, we will change the message symbol to M and keep m for the tournament depth. We will write the function family for bit value v as Gv=(g1(v),…,gm(v)) and use this notation consistently in Section 3.1, Algorithm 1, and Figure 1.
W3 & Q2: Distribution-level analysis
We measured symmetric KL and MAE between q and qˉ, the token distributions for the two bit values. At each of 512 contexts per model, we fixed the base distribution, candidate set, and g0, then compared g1=1−g0 with an independently sampled g1.
Complementary functions increased both distance measures relative to independently sampled functions, which coincided with gains in bit accuracy of 5.08 percentage points on Llama3 and 4.98 points on Gemma2. Therefore, greater separation between q and qˉ was associated with higher decoding accuracy on both models.
Q3: Robustness analysis
Moreover, we measured each method’s AUC drop in percentage points from its aggregate clean AUC to its AUC under attack. The settings were the same throughout (Llama3-8B, C4, 16 bits, 256 tokens).
Although WorldCup and BiMark had the same reported clean AUC of 1.000, WorldCup lost less AUC under every attack, averaging 6.14 points versus 8.73 points for BiMark. Therefore, WorldCup’s post-attack advantage reflects greater robustness to all eight attacks, not simply better pre-attack performance.