MorphMark: Flexible Adaptive Watermarking for Large Language Models

Zongqi Wang, Tianle Gu, Baoyuan Wu, Yujiu Yang

Abstract

Large Language Model (LLM) watermarking typically forces a trade-off between detectability and text quality. We introduce MorphMark, a training-free, model-agnostic method that dynamically adjusts watermark strength token-by-token based on the cumulative probability of "green-list" tokens ($P_G$). By adapting strength on the fly, MorphMark achieves high detectability and robustness against paraphrasing without sacrificing text quality or increasing inference latency.

Introduction

As LLMs become ubiquitous, watermarking their outputs is critical for copyright protection and tracing AI-generated text. The standard approach, such as KGW (Kirchenbauer et al., 2023), uses a cryptographic hash of the previous token to split the vocabulary into a "green list" and a "red list," artificially boosting the probability of sampling green tokens. However, a fixed watermark strength often leads to a fundamental dilemma: high strength ruins text quality in constrained contexts, while low strength is easily removed by paraphrasing. MorphMark addresses this by formalizing the effectiveness-quality trade-off and introducing an adaptive mechanism.

The MorphMark Approach

The core insight of MorphMark is the importance of $P_G$, the cumulative original probability of all tokens assigned to the green list at a given step. If $P_G$ is naturally low, forcing a green token choice requires a massive, unnatural shift in the probability distribution, which spikes perplexity.

Extracted figure from page 4 of MorphMark: Flexible Adaptive Watermarking for Large Language Models

Instead of a fixed hyperparameter, MorphMark dynamically calculates a strength value $r$ using a growth function $r = \phi(P_G)$. When $P_G$ is high, a strong watermark is applied; when $P_G$ is low, the watermark is weakened to preserve quality. The authors explored linear, logarithmic, and exponential scaling functions, finding that the exponential variant (MorphMark$_{exp}$) provides the optimal balance by ramping up strength only when it is safe to do so.

Experimental Evaluation

MorphMark was evaluated across several datasets and tasks to test general generation and robustness:

  • General Generation: C4 (Colossal Clean Crawled Corpus).
  • Downstream Tasks: WMT16 (German-to-English translation) and CNN-DM (summarization).
  • Low-Entropy Scenarios: HumanEval (code generation).
  • Models: OPT (1.3B to 6.7B), LLaMA-2-7B (for perplexity), and StarCoder2-3B.

Results and Discussion

MorphMark demonstrates superior robustness compared to fixed-strength baselines. Under a 30% word substitution attack (Word-S/30%), MorphMark$_{exp}$ maintained a True Positive Rate (TPR@1%) of 0.9600 on OPT-1.3B, whereas the KGW baseline dropped to 0.8050.

Extracted figure from page 6 of MorphMark: Flexible Adaptive Watermarking for Large Language Models

Crucially, this robustness does not come at the cost of text quality. MorphMark achieved lower Perplexity (11.35) than KGW (11.49) and maintained or improved performance on semantic metrics like BERTScore. Furthermore, because it relies solely on output logits, MorphMark is training-free and adds zero computational overhead to detection. In low-entropy tasks like code generation, MorphMark synergizes well with detection-side interventions like Entropy-Weighted Detection (EWD), boosting detectability on HumanEval to 0.9207.

Conclusion

MorphMark provides a flexible, adaptive framework for LLM watermarking that breaks the traditional quality-robustness trade-off. By intelligently modulating strength based on the local token distribution, it ensures reliable detection even under adversarial conditions. While challenges remain in highly deterministic settings, MorphMark represents a significant step forward in deployable, model-agnostic AI tracing.

BibTeX
@inproceedings{wang2025morphmark,
  author = {Zongqi Wang and Tianle Gu and Baoyuan Wu and Yujiu Yang},
  title = {{MorphMark: Flexible Adaptive Watermarking for Large Language Models}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure