This paper proposes a meta-learning framework to address the problem that generative, explainable, and flexible recommender systems based on large-scale language models (LLMs) are not suitable for cold-start user situations (i.e., situations where there is little or no interaction history). Existing supervised learning fine-tuning and collaborative filtering methods are effective when there is a large amount of user-item data, but they suffer from high maintenance and update costs. In this paper, we propose a meta-learning framework that treats each user as a task and learns soft prompt embeddings using Reptile and MAML optimization. The learned vectors are added to the input tokens as differentiable control variables representing user behavior priors. We meta-optimize the prompts by episode sampling, inner-loop adaptation, and outer-loop generalization. We demonstrate that the proposed model outperforms existing methods in terms of NDCG@10, HR@10, and MRR metrics on MovieLens-1M, Amazon Reviews, and Recbole datasets, and achieves real-time processing speeds of less than 300ms on a consumer GPU. Additionally, this method also supports personalization for users without a history, and its adaptation speed of 275ms suggests that it can be applied to real-time risk profiling in financial systems, contributing to improving the stability of payment networks.
