Researchers at Hugging Face have published a new paper that exposes a critical flaw in how autonomous driving vision-language models (VLMs) are trained. The study, titled "Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs," identifies a trajectory anchoring bias that arises when chain-of-thought (CoT) supervision is conditioned on the logged ground-truth (GT) future trajectory.
The authors demonstrate that when teacher models are shown the actual future trajectory during training, they tend to rationalize the revealed outcome rather than infer decisions from scene evidence. This leads to less causally faithful reasoning and significantly more severe hallucinations, particularly in causally challenging driving scenarios.
To address this, the team introduces Autonomous-Driving Multiple-Choice Question (AD-MCQ), which reformulates planning as a selection among explicit trajectory candidates, avoiding the need for open-ended trajectory synthesis. Building on this, they propose DEFT-RLVR (Deferred Exposure of Future Trajectories for RLVR), a reinforcement learning framework that transforms future trajectories from pre-decision anchors into post-decision verification targets.
Experimental results show that DEFT-RLVR improves AD reasoning while preserving or even enhancing general visual capabilities. Because AD-MCQ operates entirely within the VLM and allows difficulty to be controlled through candidate construction, it offers a flexible, scalable, and extensible foundation for future research on verifiable AD reasoning.