This paper introduces a novel watermarking method for large language models (LLMs) that enhances both detectability and semantic coherence. The method employs multi-objective optimization (MOO) to dynamically adjust token-specific splitting ratios and watermark logits, ensuring that watermarks are imperceptible to humans but detectable by algorithms. The approach uses lightweight networks to generate token-specific values for these parameters, allowing for adaptive adjustments based on the context and semantics of each token. This ensures that the watermarking process does not compromise the semantic integrity of the generated text. The method is evaluated against existing watermarking techniques, demonstrating superior performance in both detectability and semantic quality. It outperforms methods like KGW, SWEET, MultiBit, SIR, and EXP-edit, particularly in maintaining semantic coherence while improving detectability. The method is also robust against common attacks, such as paraphrase and copy-paste attacks, due to its adaptive watermarking strategy. The paper highlights the importance of watermarking in ensuring the ethical use of LLMs, preventing misuse in areas such as election manipulation, fake news dissemination, and academic dishonesty. It also emphasizes the need for responsible use of watermarking algorithms to protect intellectual property rights and reduce unauthorized use of generated content. The proposed method offers a balanced approach to watermarking, ensuring both detectability and semantic integrity of LLM-generated texts.This paper introduces a novel watermarking method for large language models (LLMs) that enhances both detectability and semantic coherence. The method employs multi-objective optimization (MOO) to dynamically adjust token-specific splitting ratios and watermark logits, ensuring that watermarks are imperceptible to humans but detectable by algorithms. The approach uses lightweight networks to generate token-specific values for these parameters, allowing for adaptive adjustments based on the context and semantics of each token. This ensures that the watermarking process does not compromise the semantic integrity of the generated text. The method is evaluated against existing watermarking techniques, demonstrating superior performance in both detectability and semantic quality. It outperforms methods like KGW, SWEET, MultiBit, SIR, and EXP-edit, particularly in maintaining semantic coherence while improving detectability. The method is also robust against common attacks, such as paraphrase and copy-paste attacks, due to its adaptive watermarking strategy. The paper highlights the importance of watermarking in ensuring the ethical use of LLMs, preventing misuse in areas such as election manipulation, fake news dissemination, and academic dishonesty. It also emphasizes the need for responsible use of watermarking algorithms to protect intellectual property rights and reduce unauthorized use of generated content. The proposed method offers a balanced approach to watermarking, ensuring both detectability and semantic integrity of LLM-generated texts.