Status: Manuscript under review at IEEE Transactions on Intelligent Transportation Systems, 2026.
Autonomous vehicles operating in mixed traffic networks frequently encounter unpredictable, interactive scenarios characterized by complex human-machine interactions and partial observability. While traditional motion planning approaches have made significant strides, this paper explores the emerging paradigm of reasoning-based autonomous driving and its impact on broader transportation operations. By integrating vision-language-action models, this paradigm shifts from purely reactive behaviors to explicit deliberation through intermediate representations, such as rationales and structured plans. To organize this rapidly evolving field, we adopt a unified two-loop offline and online framework. The offline loop focuses on acquiring generalized driving priors, including foundational policies, world models, and evaluators, through supervised learning and reinforcement learning techniques. The online loop then refines these learned priors at decision time through mechanisms such as test-time adaptation, sampling-based predictive control, and lookahead tree search. We synthesize how reasoning tokens and world models bridge these two loops to enable superior decision-making. Finally, we outline future research directions and open questions for achieving safer, more robust autonomy in mixed traffic.