mlsanityData to 7 Oct 2026

GumbelSoft: Diversified Language Model Watermarking via the GumbelMax-trick

arXiv 2402.12948 ↗First posted 20 Feb 2024cs.CLcited by 31

On the map: LLM safety and security → Watermarking generated content and models

Abstract (arXiv)

Large language models (LLMs) excellently generate human-like text, but also raise concerns about misuse in fake news and academic dishonesty. Decoding-based watermark, particularly the GumbelMax-trick-based watermark(GM watermark), is a standout solution for safeguarding machine-generated texts due to its notable detectability. However, GM watermark encounters a major challenge with generation diversity, always yielding identical outputs for the same prompt, negatively impacting generation diversity and user experience. To overcome this limitation, we propose a new type of GM watermark, the Logits-Addition watermark, and its three variants, specifically designed to enhance diversity. Among these, the GumbelSoft watermark (a softmax variant of the Logits-Addition watermark) demonstrates superior performance in high diversity settings, with its AUROC score outperforming those of the two alternative variants by 0.1 to 0.3 and surpassing other decoding-based watermarking methods by a minimum of 0.1.

Key facts. Cited by 31 papers in the map (data through 7 Oct 2026). Placed in Watermarking generated content and models, one of 12 directions in LLM safety and security. Builds on 31 of its 51 references that are in the map. First posted on arXiv 20 Feb 2024.

Builds on

31 of 51 references.

Closest ancestors · 9

Same direction, newest first

  1. 2023A Survey of Text Watermarking in the Era of Large Language Models13 Dec
  2. 2023A Robust Semantics-based Watermark for Large Language Model against Paraphrasing15 Nov
  3. 2023Publicly-Detectable Watermarking for Language Models27 Oct
  4. 2023Robust Distortion-free Watermarks for Language Models28 Jul
  5. 2023Provable Robust Watermarking for AI-Generated Text30 Jun
  6. 2023Robust Multi-bit Natural Language Watermarking through Invariant Features3 May
  7. 2023Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense23 Mar
  8. 2022CATER: Intellectual Property Protection on Text Generation APIs via Conditional Watermarks19 Sep
  9. 2021Tracing Text Provenance via Context-Aware Lexical Substitution15 Dec

Other references · 22

  1. 2019Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformercited by 14,491
  2. 2019RoBERTa: A Robustly Optimized BERT Pretraining Approachcited by 13,266
  3. 2023Llama 2: Open Foundation and Fine-Tuned Chat Modelscited by 12,865
  4. 2021Evaluating Large Language Models Trained on Codecited by 8,053
  5. 2023DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvaturecited by 705
  6. 2019Defending Against Neural Fake Newscited by 553
  7. 2019Release Strategies and the Social Impacts of Language Modelscited by 505
  8. 2019GLTR: Statistical Detection and Visualization of Generated Textcited by 490
  9. 2023Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvaturecited by 278
  10. 2023DetectLLM: Leveraging Log Rank Information for Zero-Shot Detection of Machine-Generated Textcited by 207
  11. 2023DNA-GPT: Divergent N-Gram Analysis for Training-Free Detection of GPT-Generated Textcited by 127
  12. 2023Intrinsic Dimension Estimation for Robust Detection of AI-Generated Textscited by 111
10 more
  1. 2023On the Possibilities of AI-Generated Text Detectioncited by 88
  2. 2023A Survey on Detection of LLMs-Generated Contentcited by 70
  3. 2023Multiscale Positive-Unlabeled Detection of AI-Generated Textscited by 59
  4. 2023CoCo: A Coupled Contrastive Framework for Unsupervised Domain Adaptive Graph Classificationcited by 46
  5. 2023Origin Tracing and Detecting of LLMscited by 39
  6. 2023GPT-who: An Information Density-based Machine-Generated Text Detectorcited by 38
  7. 2021The Threat of Offensive AI to Organizationscited by 37
  8. 2023Smaller Language Models are Better Black-box Machine-Generated Text Detectorscited by 35
  9. 2023Efficient Detection of LLM-generated Texts with a Bayesian Surrogate Modelcited by 18
  10. 2023Evaluating Large Language Model Creativity from a Literary Perspectivecited by 7

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2024: 7; 2025: 13; 2026: 11.

31 papers.

Sort:

By area:

  1. 2024Watermarking Makes Language Models Radioactivecited by 52
  2. 2024SoK: Watermarking for AI-Generated Contentcited by 43
  3. 2024Watermarking Techniques for Large Language Models: A Surveycited by 20
  4. 2024Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approachcited by 17
  5. 2025Dataset Protection via Watermarked Canaries in Retrieval-Augmented LLMscited by 11
  6. 2025In-Context Watermarks for Large Language Modelscited by 9
  7. 2025Position: LLM Watermarking Should Align Stakeholders' Incentives for Practical Adoptioncited by 7
  8. 2026On Protecting Agentic Systems' Intellectual Property via Watermarkingcited by 7
  9. 2025On the Empirical Power of Goodness-of-Fit Tests in Watermark Detectioncited by 5
  10. 2026PASA: A Principled Embedding-Space Watermarking Approach for LLM-Generated Text under Semantic-Invariant Attackscited by 4
  11. 2025Multi-use LLM Watermarking and the False Detection Problemcited by 3
  12. 2025Addressing Tokenization Inconsistency in Steganography and Watermarking Based on Large Language Modelscited by 3
19 more
  1. 2024Lost in Overlap: Exploring Logit-based Watermark Collision in LLMscited by 2
  2. 2025SoK: Are Watermarks in LLMs Ready for Deployment?cited by 2
  3. 2025SimKey: A Semantically Aware Key Module for Watermarking Language Modelscited by 2
  4. 2025HeavyWater and SimplexWater: Distortion-Free LLM Watermarks for Low-Entropy Next-Token Predictionscited by 1
  5. 2026MirrorMark: Generalizable Mirrored Sampling for Multi-bit LLM Watermarkingcited by 1
  6. 2026Improving the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Modelscited by 1
  7. 2024Watermark Smoothing Attacks against Language Modelscited by 0
  8. 2024Watermarking Generative Categorical Datacited by 0
  9. 2025Disappearing Ink: Obfuscation Breaks N-gram Code Watermarks in Theory and Practicecited by 0
  10. 2025Adaptive Testing for Segmenting Watermarked Texts From Language Modelscited by 0
  11. 2025UniMark: Artificial Intelligence Generated Content Identification Toolkitcited by 0
  12. 2025CODE ACROSTIC: Robust Watermarking for Code Generationcited by 0
  13. 2026WorldCup Sampling for Multi-bit LLM Watermarkingcited by 0
  14. 2026On Google's SynthID-Text LLM Watermarking System: Theoretical Analysis and Empirical Validationcited by 0
  15. 2026Rethinking LLM Watermark Detection in Black-Box Settings: A Non-Intrusive Third-Party Frameworkcited by 0
  16. 2026Trustworthy AI: Ensuring Reliability and Accountability from Models to Agentscited by 0
  17. 2026Linguistics-Aware Non-Distortionary LLM Watermarkingcited by 0
  18. 2026Signature filtering: a lightweight enhancement for statistical watermark detection in large language modelscited by 0
  19. 2026Break Step: Recursive Training Resonates with Replayed Sampling Noisecited by 0