检索增强与提示演进的车端大模型学习框架OA

Addressing the urgent demand for continual learning capabilities of intelligent connected ve-hicles in dynamic open environments,as well as the limitations of existing methods such as catastrophic forgetting,insufficient resource efficiency,and inadequate safety compliance,this paper proposes a contin-ual learning framework for vehicular large models that integrates retrieval-augmented generation and prompt evolution.The framework establishes a collaborative learning paradigm of"external memory-in-telligent interface-efficient update,"realizes the structured storage and efficient retrieval of driving expe-rience via a dynamic hierarchical memory bank,designs a prompt-evolving engine to generate context-a-ware adaptive prompts,and adopts a vehicle-optimized parameter-isolated fine-tuning strategy for effi-cient knowledge injection.Experimental results demonstrate that the proposed framework achieves task ac-curacy exceeding 90%across various typical driving scenarios,reduces response latency by 34.61%,de-creases memory usage by 28.99%,and lowers safety violations by 59.06%,significantly outperforming traditional continual learning and static retrieval-augmented generation methods.