The paper introduces LAPDOG, a novel framework for personalized dialogue generation that leverages external knowledge to enhance persona profiles. Personalized dialogue generation aims to generate tailored responses based on persona profiles and dialogue context, but existing methods often rely on limited and static persona descriptions, which can hinder the generation of contextually rich and diverse responses. To address this, LAPDOG proposes a two-stage training process involving a retriever and a generator. The retriever retrieves relevant stories from a story corpus (e.g., ROCStory) to augment the persona profile, while the generator uses the augmented persona and dialogue history to produce personalized responses. The model is trained using a joint framework that optimizes both the retriever and generator, with the retriever being tuned towards desired metrics (e.g., BLEU, F1, ROUGE-L). Experiments on the CONVAI2 dataset show that LAPDOG significantly outperforms baselines, demonstrating the effectiveness of learnable retrieval augmentation in personalized dialogue generation. The code for LAPDOG is publicly available for further exploration.The paper introduces LAPDOG, a novel framework for personalized dialogue generation that leverages external knowledge to enhance persona profiles. Personalized dialogue generation aims to generate tailored responses based on persona profiles and dialogue context, but existing methods often rely on limited and static persona descriptions, which can hinder the generation of contextually rich and diverse responses. To address this, LAPDOG proposes a two-stage training process involving a retriever and a generator. The retriever retrieves relevant stories from a story corpus (e.g., ROCStory) to augment the persona profile, while the generator uses the augmented persona and dialogue history to produce personalized responses. The model is trained using a joint framework that optimizes both the retriever and generator, with the retriever being tuned towards desired metrics (e.g., BLEU, F1, ROUGE-L). Experiments on the CONVAI2 dataset show that LAPDOG significantly outperforms baselines, demonstrating the effectiveness of learnable retrieval augmentation in personalized dialogue generation. The code for LAPDOG is publicly available for further exploration.