WorldMark: Plug-and-Play Interface Enhances LLM Watermarking via World Knowledge

A new interface called WorldMark has been developed by researchers to enhance the effectiveness of watermarking in large language models (LLMs). This watermarking technique embeds detectable signals during decoding to trace the origin of generated text. Current methods, including logits-based, sampling-based, entropy-aware, and adaptive-strength approaches, rely on local token statistics for signal placement, which may be inadequate in open-ended text generation. WorldMark improves upon this by utilizing a World Knowledge Memory (WKM) to structure semantic and episodic knowledge within a memory graph, transforming retrieved information into token-level saliency scores and modulating watermark strength through Asymmetric Knowledge Modulation (AKM). Notably, it does not require retraining of the backbone or additional model parameters. The paper, titled "WorldMark: A Plug-and-Play World Knowledge Interface for Cross-Host Language Model Watermarking," can be found on arXiv with the identifier 2608.06416. The abstract suggests that the WorldMark system demonstrates promising results in the primary C4 evaluation, although further details are not included.