ABSTRACT With the rapid development of intelligent cockpits, personalized recommendations have become crucial for achieving high‐level cognitive intelligence in cockpit systems. The core goal is to mine the behavioural history of drivers and passengers to provide tailored, proactive interactions based on environmental conditions and user preferences. Current systems mainly rely on rules or limited offline data, focusing on specific functions or scenarios, which lack global capabilities and struggle with multi‐task collaboration, leading to inaccuracies and limited flexibility in personalized recommendations. Large language models (LLMs), with their powerful general‐purpose understanding capabilities, have demonstrated significant advantages in reasoning about complex user intentions and enhancing interaction recommendation performance. However, LLMs have not been applied to cockpit personalized interaction recommendations. To bridge this gap and effectively balance the complexity of cockpit systems under highly diverse multi‐task and personalized requirements, this paper proposes an innovative two‐stage recommendation framework, PC‐LLMRec, specifically designed for customized recommendations in intelligent cockpits. The framework employs full‐parameter baseline model optimization and personalized adapter construction to achieve general recommendations in the cloud and personalized adjustments on the vehicle end. This allows precise capture and interpretation of both common behavioural patterns and individualized user needs, enabling cross‐scenario, multi‐task proactive recommendation. To enhance the adaptability of PC‐LLMRec, this paper also constructs an instruction‐following dataset tailored to proactive cockpit interaction recommendations. This dataset includes extensive user interaction context and real recommendation labels, ensuring effective fine‐tuning between global recommendations and personalized services. Extensive experimental results demonstrate that PC‐LLMRec excels in accuracy and adaptability across various recommendation scenarios, outperforming existing context‐learning‐based methods, retrieval‐augmented prompt strategies, and other state‐of‐the‐art models.
Dong et al. (Thu,) studied this question.