As large-scale language models (LLMs) are increasingly used in diverse real-world applications, personalization of LLMs is becoming increasingly important. While various LLM personalization approaches, including prompt-based and learning-based methods, have been actively studied, the development of effective decode-time algorithms has been overlooked despite their potential. In this paper, we propose Contrasting Personal Preference (CoPe), a novel decode-time approach applied after parameter-efficient fine-tuning (PEFT) on user-specific data. The core idea is to maximize each user's implicit reward signal, specifically leveraging reward-guided decoding for personalization. We evaluate CoPe across five open-ended personalized text generation tasks. Experimental results show that CoPe achieves robust personalization performance on ROUGE-L, with an average improvement of 10.57%, without relying on an external reward model or additional training procedures.
