Robust Text Watermarking for Large Language Models via Dual Semantic Embeddings
The episode discusses a paper titled "Robust Text Watermarking for Large Language Models via Dual Semantic Embeddings." The authors propose a method that embeds a signal based on semantic alignment between context and candidate tokens, rather than surface patterns. This approach is highly robust against paraphrasing and translation, efficient to deploy in black-box APIs, and provides reliable provenance across languages.