Title: WaterSearch: Investigating Seed Pooling to Optimize the Quality-Detectability Balance in LLM Watermarking

Abstract: Watermarking serves as an essential protective mechanism for text produced by Large Language Models (LLMs). By embedding distinct signals within model outputs, these techniques facilitate accurate attribution and bolster the security of AI-generated content. While current strategies effectively manipulate token generation probabilities to embed these signals, they are fundamentally constrained by a conflict between detectability and text quality. The randomness and signal intensity necessary for reliable watermarking often compromise the utility of the text for subsequent tasks. To address this, we introduce a novel embedding approach that utilizes controlled seed pools to enable diverse, parallel generation of watermarked content. Leveraging this approach, we present WaterSearch, a flexible, sentence-level framework based on search algorithms that can be integrated with various existing watermarking methods. WaterSearch improves text quality through the joint optimization of two primary factors: watermark signal properties and distribution fidelity. Additionally, the framework incorporates a sentence-level detection mechanism designed to withstand various attacks. Our evaluation spans ten diverse tasks using three prominent LLMs. Extensive testing reveals that WaterSearch delivers an average performance boost of 51.01% compared to state-of-the-art baselines, maintaining a watermark detectability strength of 95%. In more difficult contexts, such as generating short texts or outputs with low entropy, the method achieves performance increases of 47.78% and 36.47%, respectively. Furthermore, WaterSearch demonstrates strong resilience against attacks involving synonym substitution, paraphrasing, and insertion, preserving high detectability rates. The source code for this project is accessible at https://github.com/Yukang-Lin/WaterSearch.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC