ThinkDrive: Adaptive Dual-Process Reasoning for Autonomous Driving via Uncertainty-Triggered Causal Deliberation

End-to-end autonomous driving remains fragile in long-tail scenarios, while incorporating vision-language models (VLMs) introduces substantial deliberation latency that cannot interfere with the real-time planning loop. We present ThinkDrive, a dual-process driving framework designed under explicit real-time queuing constraints. The framework contains four coordinated components. First, a Scene Complexity Estimator regulates System-2 activation through a trigger cool-down mechanism, allowing at most one asynchronous request every ⌈L2/Δt⌉ frames and thereby preventing queue saturation under a System-2 latency of L2=565 ms. Second, a multi-modal System-1 planner generates K1=5 candidate trajectories within 44 ms and is trained with winner-takes-all imitation learning together with explicit score supervision. Third, a two-stage Causal-CoT module uses the VLM to identify risk agents and predict a preferred spatial goal GVLM, after which a single batched scm_rollout selects the safest candidate and extracts its endpoint as a world-coordinate goal anchor gS2. Fourth, a Goal-Anchor Replanning module transforms gS2 into the current ego frame and selects the candidate whose waypoint at the remaining time horizon is closest to the transformed goal. This design avoids coordinate-space mixing, mitigates bias caused by mismatched temporal horizons, and prevents semantic instability across replanning cycles. On nuPlan test14-hard, ThinkDrive with InternVL2-8B and a 6.8% trigger rate achieves 74.9 PDMs, outperforming AdaThinkDrive at 73.1 while maintaining a nominal latency of 44 ms.