自动驾驶系统推理的调查：开放挑战与新兴范式

The development of high-level autonomous driving (AD) is transitioning from perception-centric limitations to a more fundamental bottleneck: a lack of robust and generalizable reasoning. Current AD systems manage structured environments but often struggle in long-tail scenarios and complex social interactions that require human-like judgment. The emergence of large language models (LLMs) and multimodal models (MLLMs) presents a transformative opportunity to integrate a powerful cognitive engine into AD systems, moving beyond mere pattern matching towards genuine comprehension. However, there is a critical lack of a systematic framework to guide this integration. A comprehensive review of this emerging field is provided to bridge this gap.