What it is
Turns driving planning into a verifiable multiple-choice trajectory task and trains VLMs to reason before seeing options, reducing trajectory anchoring bias. Instead of exposing future trajectories during reasoning, it hides them until after the model commits to a scene-based decision. Planning becomes MCQ over explicit trajectory candidates, enabling exact verification and RL on causally grounded CoTs.
Key results
- 77.9% ACC vs 28.1% base on AD-MCQ with Qwen3-VL-8B
- 82.2% ACC vs 34.0% base on AD-MCQ with Qwen3.5-4B
- 49.5% vs 39.6% ACC on OOD nuScenes with Qwen3-VL-8B
Shows that information order matters: reasoning first and grounding later yields more faithful and transferable driving decisions.