This paper proposes optimization methods for personalizing large language models (LLMs) through retrieval augmentation. The authors introduce two optimization algorithms for improving retrieval models that deliver a limited number of personal documents to LLMs for personalized generation. The first method uses reinforcement learning, where the reward function is defined using any arbitrary metric for personalized generation. The second method employs knowledge distillation from the downstream LLM to the retrieval model. Additionally, the paper introduces a pre- and post-generation retriever selection model that decides which retriever to use for each LLM input. The methods are evaluated on the LaMP benchmark, which consists of seven diverse personalization tasks, including three personalized text classification tasks and four personalized text generation tasks. The results show statistically significant improvements in six out of seven datasets. The best-performing method achieves an average of 5.5% state-of-the-art improvements across all LaMP datasets. The paper also discusses the importance of retrieval selection in personalizing LLMs, as different retrieval models may be optimal for different tasks. The authors propose two retriever selection models, RSPG-Pre and RSPG-Post, which select the most appropriate retriever for each input. The results show that RSPG-Post performs best on six out of seven datasets. The paper concludes that the proposed methods significantly improve the performance of LLM personalization through retrieval augmentation.This paper proposes optimization methods for personalizing large language models (LLMs) through retrieval augmentation. The authors introduce two optimization algorithms for improving retrieval models that deliver a limited number of personal documents to LLMs for personalized generation. The first method uses reinforcement learning, where the reward function is defined using any arbitrary metric for personalized generation. The second method employs knowledge distillation from the downstream LLM to the retrieval model. Additionally, the paper introduces a pre- and post-generation retriever selection model that decides which retriever to use for each LLM input. The methods are evaluated on the LaMP benchmark, which consists of seven diverse personalization tasks, including three personalized text classification tasks and four personalized text generation tasks. The results show statistically significant improvements in six out of seven datasets. The best-performing method achieves an average of 5.5% state-of-the-art improvements across all LaMP datasets. The paper also discusses the importance of retrieval selection in personalizing LLMs, as different retrieval models may be optimal for different tasks. The authors propose two retriever selection models, RSPG-Pre and RSPG-Post, which select the most appropriate retriever for each input. The results show that RSPG-Post performs best on six out of seven datasets. The paper concludes that the proposed methods significantly improve the performance of LLM personalization through retrieval augmentation.