Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
SAEMark: Steering Personalized Multilingual LLM Watermarks with Sparse Autoencoders
Authors: Zhuohao Yu, Xingru Jiang, Weizheng Gu, Yidong Wang, Qingsong Wen, Shikun Zhang, Wei Ye
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments across 4 datasets demonstrate strong watermarking performance on English, Chinese, and code while preserving text quality. |
| Researcher Affiliation | Academia | Peking University EMAIL, EMAIL |
| Pseudocode | Yes | Figure 2: Pseudocode for SAEMARK: generation and detection. |
| Open Source Code | Yes | Justification: We report hyperparameters and we also open-source everything in the linked mentioned in abstract. ... Justification: We open-source everything in the linked mentioned in abstract. https://zhuohaoyu.github.io/SAEMark |
| Open Datasets | Yes | We evaluate on 4 diverse datasets as shown in Table 1. ... C4 [22] LCSTS [23] MBPP [24] Panda LM [25] |
| Dataset Splits | Yes | From test split of sanitized version of MBPP. |
| Hardware Specification | No | The paper discusses leveraging highly optimized inference backends like TGI with parallel candidate generation and custom CUDA kernels, and mentions using backbone LLMs like Qwen2.5-7B-Instruct, Llama-3.2-3B-Instruct, gemma-3-4b-it. However, it does not specify concrete hardware components such as GPU models (e.g., NVIDIA A100), CPU models, or memory details used for the experiments. General terms like 'Modern inference engines' or 'inference backends' are used without specific hardware details. |
| Software Dependencies | No | The paper mentions using 'Qwen2.5-7B-Instruct' as the base model and notes parameters like 'Temperature: 0.7' and 'Max New Tokens: 20'. It also refers to 'Mark LLM [26] toolkit and Waterfall [13]' as baselines. However, it does not list specific software dependencies with their version numbers, such as Python, PyTorch, or CUDA versions, which are necessary for full reproducibility. |
| Experiment Setup | Yes | B.1 Hyperparameters (SAEMARK): The following hyperparameters were used for the SAEMARK: Candidate Number (N): 50. Unit Number (M): 10. Attempt Number (K): 15. B.2 Model Configuration: Base Model: Qwen2.5-7B-Instruct. Sampling: do_sample is set to True. Temperature: 0.7. Max New Tokens: 20. |