acceptodds
Under review as a conference paper at ICLR 2027

WaterMoE: Expert-Routing-based Watermarking for High Fidelity and Efficiency

Abstract

Large language models (LLMs) have achieved remarkable success but raise growing concerns about content provenance and misuse, motivating the need for reliable watermarking techniques. However, these techniques have rarely been adopted in practice mainly for two reasons: i) severely degraded model performance, and ii) additional inference overhead. To confirm the problem, we construct a comprehensive benchmark spanning different generation tasks to systematically evaluate 9 representative watermarking methods. We found almost all existing methods are designed for text fluency, but not for restricted and complicated tasks, and their overhead prevents them from deployment in latency-critical systems. To address i) and ii), we propose an LLM watermarking scheme WaterMoE for the growingly popular Mixture-of-Experts (MoE) LLMs. WaterMoE embeds watermarking signals through controlled perturbation into the expert selection at each router, which accumulates to token selection shift at the final output. In contrast to watermarking as a post-processing token-sampling approach, WaterMoE embeds watermark within the inference loop incurring negligible quality degradation and computational overhead. Extensive experiments demonstrate that our method achieves a fidelity performance close to the unwatermarked and consistently outperforms state-of-the-art watermarking methods on the benchmark, with up to speedup, incurring merely 1% additional inference latency compared to native generation. The results demonstrate the capability of WaterMoE to be deployed in real-world tasks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM watermarking

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_zDNJOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents WaterMoE, a watermarking method for the Mixture-of-Experts (MoE) LLMs. It embeds watermark through the expert routing process. The experiments demonstrate that WaterMoE is efficient, can achieve high fidelity over datasets with different complexity levels.

Critical strengths

The proposed watermarking mechanism for MoE LLMs is novel and interesting.

Critical weaknesses

As noted by the authors, WaterMoE requires access to the model and its routing configuration, limiting its applicability to provider-side verification.

WaterMoE modifies the expert routing paths, causing deviations in the generated texts compared to those of the clean model. The proposed approach relies on the inherent redundancy of MoE models, which allows experts with similar routing scores can be dynamically selected without significantly affecting output quality. However, it remains unclear whether such rich redundancy is a general property of MoE models, with each expert having sufficiently many alternatives exhibiting similar routing scores to enable reliable selection. The evaluation is limited to only two MoE models, which appears to be insufficient to establish whether such redundancy is a general property or is specific to the models evaluated. A broader evaluation across diverse MoE architectures and model scales would help assess the generalizability of the proposed approach.

Questions for authors

How are the routing scores defined and computed? What is the rationale for selecting 0.2 as the threshold for the difference in routing scores between experts?

Reviewer_Q7D3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes WaterMoE, a zero-bit text watermarking method that applies small biases to expert-routing scores in MoE language models. Detection compares token likelihoods under biased and reference routing configurations and uses a calibrated statistic to identify the watermark. The evaluation covers detection performance, generation quality, insertion overhead, and robustness across text generation, summarization, coding, reasoning, and instruction-following tasks on Mixtral and Qwen.

Critical strengths

(1) WaterMoE exploits expert selection and mixture weights through a precomputed map and lightweight routing biases, without additional training. Combining inexpensive insertion with model-based verification gives the method a clear design rationale.

(2) The evaluation considers both detectability and generation quality across several task types and includes encouraging results. On Qwen/CodeContests, for example, WaterMoE achieves 92% TPR@1%, compared with a maximum of 64% among the reported baselines, while maintaining competitive PASS@1. Disjoint calibration and evaluation sets, incorrect Expert Map tests, and negatives from multiple sources provide useful support for the detector’s validity.

Critical weaknesses

(1). Further discussion of when the method is most effective.

The gains vary across models and tasks. On Mixtral/C4 and Mixtral/Booksum, detection is already saturated, and WaterMoE has higher perplexity than some baselines. The reported task scores also show some differences from native generation: APPS PASS@1 changes from 65.4% to 61.9%, and GSM8K accuracy changes from 72% to 68%.

(2). Details about the SynthID comparisons .

Table 5 could specify whether the non-distortionary or distortionary mode was used in each setting. On MultiNews, SynthID achieves 8% TPR@1% with Mixtral and 98% with Qwen. Clarifying the configurations and discussing possible reasons for this difference.

Questions for authors

(1) Could the authors summarize the tasks and operating conditions in which WaterMoE offers the most useful quality–detectability trade-offs, and explain how these benefits relate to the method’s generation and verification requirements?

(2) Could expert redundancy, the extent of routing changes, or output distributions help explain the variation across tasks? It would be interesting to understand whether the two evaluated models suggest a consistent explanation.

(3) Could the authors provide the SynthID watermarking mode and main decoding settings, together with some discussion of the 8% versus 98% detection rates on MultiNews?

Reviewer_wu6ROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a watermarking method for MoE language models, named WaterMoE, which embeds watermarks by adding small, keyed biases to expert-routing scores. Watermark detection is performed by comparing token likelihoods under the biased and original routing configurations. Therefore, this paper needs access to the corresponding watermarked model and is not yet applicable to black-box models. This paper also evaluates detection, generation quality, insertion efficiency, and robustness across several language generation tasks.

Critical strengths
  • Using the flexibility of expert routing to embed watermarks is an interesting idea, and the proposed mechanism is simple and requires no additional training.
  • The evaluation considers code correctness, reasoning accuracy, and instruction following. The reported combination of high detection rates and performance close to the unwatermarked model on tasks such as CodeContests and IFEval is promising.
Critical weaknesses
  • The experimental comparisons need a bit more clarification. Table 5 increases the watermark strength of several baselines on high-complexity tasks, but Table 2 does not show their detection–quality tradeoffs, and the common decoding settings are not specified.
  • The comparison with existing structural watermarking methods can also be potentially enriched. GaussMark (ICML 2025) embeds watermarks through model-weight perturbations without additional generation latency, and it is currently missing from the discussion; a comparison would help establish the specific benefit of perturbing expert routing.
Questions for authors
  1. Some numerical details need further clarification. For example, the common decoding settings is not explicitly specified. And it would be more informative to extend the comparison in Figure 1(b) to a representative high-complexity task, with comparisons at similar detection rates, to better illustrate the fidelity advantage.

  2. It would be useful to clarify the benefit of expert-routing perturbations over the weight perturbations used by GaussMark (ICML 2025). Such a comparison would help establish whether routing provides a better detection-quality tradeoff.

  3. As acknowledged in Section 5, the likelihood-based detector requires access to the generating model. It would be useful to discuss whether routing statistics, such as the empirical frequency of selecting green experts under the reference configuration, could provide a simpler detection rule. Although this would still require model access, it may avoid evaluating token likelihoods under both configurations and reduce the verification cost.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.