HATS: High-Accuracy Triple-Set Watermarking for Large Language Models

2025/12/22 by Hu, Zhiqing, Zhao, Chenxu, Lu, Jiazhong +1
#Computation and Language (cs.CL) #FOS: Computer and information sciences

paper · doi:10.48550/arxiv.2512.19378

Abstract

Misuse of LLM-generated text can be curbed by watermarking techniques that embed implicit signals into the output. We propose a watermark that partitions the vocabulary at each decoding step into three sets (Green/Yellow/Red) with fixed ratios and restricts sampling to the Green and Yellow sets. At detection time, we replay the same partitions, compute Green-enrichment and Red-depletion statistics, convert them to one-sided z-scores, and aggregate their p-values via Fisher's method to decide whether a passage is watermarked. We implement generation, detection, and testing on Llama 2 7B, and evaluate true-positive rate, false-positive rate, and text quality. Results show that the triple-partition scheme achieves high detection accuracy at fixed FPR while preserving readability.

Citations

• The Widespread Adoption of Large Language Model-Assisted Writing Across Society
• The Llama 3 Herd of Models
• Securing Large Language Models: Threats, Vulnerabilities and Responsible Practices
• LLMs with Industrial Lens: Deciphering the Challenges and Prospects -- A Survey
• Security and Privacy Challenges of Large Language Models: A Survey
• Adaptive Text Watermark for Large Language Models
• DeepSeek LLM: Scaling Open-Source Language Models with Longtermism
• The Adoption and Efficacy of Large Language Models: Evidence From Consumer Complaints in the Financial Industry
• Three Bricks to Consolidate Watermarks for Large Language Models
• Llama 2: Open Foundation and Fine-Tuned Chat Models
• On the Reliability of Watermarks for Large Language Models
• Robust Multi-bit Natural Language Watermarking through Invariant Features
• GPT-4 Technical Report
• A Watermark for Large Language Models
• Tracing Text Provenance via Context-Aware Lexical Substitution
• Pointer Sentinel Mixture Models
• Character-level Convolutional Networks for Text Classification

Related

• Three Bricks to Consolidate Watermarks for Large Language Models
• Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach
• Towards Codable Watermarking for Injecting Multi-bits Information to LLMs
• WaterMax: breaking the LLM watermark detectability-robustness-quality trade-off
• Entropy-Guided Watermarking for LLMs: A Test-Time Framework for Robust and Traceable Text Generation