Majority Bit-Aware Watermarking For Large Language Models
Existing LLM watermarking methods trade text quality for message accuracy by limiting a 'green list' of acceptable tokens during encoding. This work introduces Majority Bit-Aware Encoding, MajorMark and MajorMark+, which relaxes this restriction. By basing the green list on the majority bit of message shards, these methods preserve strong watermark signals even with larger token sets, leading to better text quality and decoding accuracy.