[2Win5] Poster session 2

[2Win5-37] A Watermark for Large Language Models in Long Text Generation

〇Koichi Nagatsuka¹, Yasuhiro Sogawa¹ (1.Hitachi, Ltd.)

Keywords:digital watermark, large language model, long-form text generation

In recent years, as large language models (LLMs) have advanced, there is growing concern about the misuse of LLMs for disseminating misinformation. A widely recognized method for distinguishing text generated by LLMs is digital watermarking, which adjusts decoding probabilities using a secret key. However, this approach has the drawback of degrading text quality, particularly for shorter texts. On the other hand, long-context models, which have recently garnered attention, are capable of handling lengthy texts. The extent to which digital watermarking affects text quality in long-text generation has not been thoroughly investigated. This study investigates the trade-off between watermark detection accuracy and text quality in long-text generation by leveraging publicly available long-context models and evaluating their performance. Experimental results demonstrate that it is possible to improve detection accuracy while mitigating adverse effects on text quality in long-text generation.