Researchers propose Hierarchical Vocabulary Routing (HeRo), a watermarking framework that allows large language models to embed metadata with fine-grained access control. Unlike existing multi-bit schemes that require revealing the entire message for verification, HeRo recursively partitions the vocabulary so verifiers can decode only their authorized portions.
- The method distributes watermark information across hierarchical layers to support selective disclosure.
- It preserves the unbiasedness of the underlying sampling process to maintain text quality.
- Experiments show high detection accuracy and low latency while preventing unnecessary information exposure.
This approach addresses privacy concerns by allowing different verifiers to access only specific parts of the embedded payload.