An End-to-End Model for Logits-Based Large Language Models Watermarking
Authors: Ka Him Wong, Jicheng Zhou, Jiantao Zhou, Yain-Whar Si
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that our method achieves superior robustness, outperforming distortion-free methods by 37 39% under paraphrasing and 17.2% on average, while maintaining text quality on par with the distortionfree methods in terms of text perplexity and downstream tasks. |
| Researcher Affiliation | Academia | 1State Key Laboratory of Internet of Things for Smart City, Department of Computer and Information Science, Faculty of Science and Technology, University of Macau, China 2Department of Computer and Information Science, Faculty of Science and Technology, University of Macau, China. |
| Pseudocode | No | The paper describes methods in regular paragraph text and uses diagrams (e.g., Figure 1, Figure 2) but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/KAHIMWONG/E2E LLM WM. |
| Open Datasets | Yes | We use samples from the Wiki Text-103 dataset (Merity et al., 2017) as prompts for training and C4 (Raffel et al., 2020) for evaluation. |
| Dataset Splits | Yes | To train our end-to-end model, we choose OPT-1.3B as the online LLM to reduce training cost. We use samples from the Wiki Text-103 dataset (Merity et al., 2017) as prompts for training and C4 (Raffel et al., 2020) for evaluation. The first 30 tokens from the C4 dataset (Raffel et al., 2020) as prompts, and generate 200 clean watermarked tokens as a watermark sample, with original human-written text serving as non-watermarked samples. |
| Hardware Specification | Yes | All experiments are conducted on one single NVIDIA RTX A6000 48G GPU. |
| Software Dependencies | No | The paper mentions several tools and algorithms like 'Adam optimizer', 'OPT-1.3B', 'MGDA', and 'Gumbel-Softmax sampling (GSS)' but does not provide specific version numbers for software libraries such as Python, PyTorch, or CUDA that would be needed to replicate the experiment. |
| Experiment Setup | Yes | Table 15 provides detailed hyperparameters for the end-to-end model training, including Learning rate (1e-4), Batch size (8), Training step (35k), Encoder context size (10), Top-k candidate (20), Gumbel-softmax temperature (0.1), Watermark strength (1), and weights for Ldec (10) and Lsem (1). |