DEFT-RLVR tackles future-trajectory leakage in autonomous driving VLM reasoning

Introduction

Vision-Language-Action models are becoming an increasingly important direction for autonomous driving. The goal is not only to predict an action, but also to make the model explain why a maneuver is appropriate: slowing down for a pedestrian, staying in lane because of nearby traffic, or selecting a safer path through a complex scene. Chain-of-thought supervision has therefore become a common tool for improving the reasoning component of driving VLMs.

This paper highlights a subtle but important weakness in that pipeline. In many annotation setups, the teacher model that produces the reasoning trace is shown the logged ground-truth future trajectory. Once the future outcome is visible, the model may no longer infer a decision from the scene. Instead, it can rationalize the trajectory it has already seen. The authors call this trajectory anchoring bias, and report that it leads to less causally faithful reasoning and more severe hallucinations, especially in challenging causal scenarios.

Key ideas

• The issue is premature exposure, not reasoning supervision itself. If the ground-truth trajectory is available before the model explains a decision, the explanation can become a polished after-the-fact story.
• Removing the trajectory creates another problem. Open-ended trajectory generation mixes high-level intent, geometric path synthesis, and low-level driving dynamics, making the learning target harder to verify.
• AD-MCQ turns planning into a multiple-choice task. Instead of generating a full future path from scratch, the model selects among explicit trajectory candidates.
• DEFT-RLVR changes the role of future trajectories. Future trajectories are no longer decision-time anchors; they are deferred and used after the decision as verification targets in a reinforcement learning with verifiable rewards setup.
• The approach stays within VLM-only inference. The difficulty of the task can be adjusted through candidate construction, while experiments indicate improved autonomous-driving reasoning without sacrificing general visual capabilities.

Why it matters

The main contribution is a reframing of what “explainable driving reasoning” should mean. For safety-critical systems, a convincing textual rationale is not enough. If the rationale is produced after the answer is revealed, it may have limited value for auditing, debugging, or validating an autonomous system. The model must be encouraged to ground its reasoning in scene evidence before the future trajectory is known.

AD-MCQ is also a pragmatic design choice. Multiple-choice trajectory selection makes planning decisions easier to verify than open-ended trajectory generation, while still forcing the VLM to compare plausible alternatives. It also gives researchers control over task difficulty by changing how candidates and distractors are constructed.

The approach does introduce a new dependency: the quality and diversity of candidate trajectories matter. If the candidate set is weak, the model may learn shortcuts tied to the options rather than robust driving judgment. Even so, DEFT-RLVR surfaces a critical data-design issue in autonomous-driving VLMs: models should not be allowed to see the future first and then pretend they reasoned their way there.

Source: Hugging Face Daily Papers