Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs
AI Summary
This paper argues that exposing logged ground-truth future trajectories to teacher models during chain-of-thought annotation creates trajectory anchoring bias: models rationalize observed outcomes instead of inferring decisions from scene evidence, leading to less causally faithful reasoning and more severe hallucinations in difficult scenes. It introduces Autonomous-Driving Multiple-Choice Questions (AD-MCQ), which frames planning as selection among explicit candidate trajectories, and Deferred Exposure of Future Trajectories for RLVR (DEFT-RLVR), which withholds future trajectories until after the decision. The abstract reports improved autonomous-driving reasoning while preserving or improving general visual capabilities.
Why it's worth reading
As chain-of-thought supervision spreads in autonomous driving, this work identifies future-trajectory leakage as a hidden annotation shortcut and proposes a concrete verification-oriented training framework.
Deep Read
What Happened
[Original facts] The paper studies chain-of-thought supervision for autonomous-driving vision-language-action models. It argues that common annotation pipelines expose logged ground-truth future trajectories to the teacher model before reasoning is produced. It proposes AD-MCQ and DEFT-RLVR for candidate-trajectory selection and post-decision verification.
Core Technology
[Original facts] AD-MCQ formulates planning as selecting among explicit trajectory candidates, avoiding the need to jointly generate a high-level decision, precise geometry, and low-level dynamics. DEFT-RLVR delays access to future trajectories, turning them from pre-decision anchors into post-decision verification targets.
Key Evidence & Numbers
[Original facts] The abstract reports trajectory anchoring bias, less causally faithful CoTs, and more severe hallucinations when the ground-truth trajectory is exposed, especially in causally challenging scenes. The supplied abstract gives no dataset details, sample counts, metric values, or percentage improvements, so the magnitude of the gains cannot be quantified here.
Why It Matters
[Analysis] If the outcome is available while a reasoning trace is created, a fluent CoT may explain the result after the fact rather than represent evidence-based decision-making. Delayed exposure creates a cleaner interface for testing whether the model decides first and verifies afterward.
Practical Impact
[Analysis] AD-MCQ could support scalable planning benchmarks with controllable difficulty through candidate construction. DEFT-RLVR may fit reinforcement-learning and verifiable-reward pipelines, but deployment still requires closed-loop simulation, control constraints, and safety evaluation.
Limitations & Uncertainty
[Original facts] The supplied material does not specify how candidate trajectories are generated, how rewards are defined, which base VLMs or autonomous-driving datasets are used, or whether closed-loop driving results are reported. [Unverified inference] If the candidate set omits a safe action, the model may only be correct within a constrained option space; the full paper is needed to assess this risk.
Original Sources
- arXiv:2608.01755
- Source label: hf-papers
- Provided publication timestamp: 2026-08-04