TL;DR
This paper identifies "trajectory anchoring bias" in autonomous driving VLMs, where models trained on ground-truth future trajectories merely rationalize known outcomes post-hoc rather than inferring decisions from scene evidence. To fix this, the authors introduce AD-MCQ, a multiple-choice trajectory benchmark, and DEFT-RLVR, a training framework that hides candidate trajectories until after the model commits to a scene-based decision. This approach significantly improves driving reasoning and decision-making while preventing the catastrophic degradation of general visual capabilities typically caused by standard fine-tuning.
What Problem Does This Paper Address?
Vision-Language-Action (VLA) models are increasingly used in autonomous driving (AD) to handle high-level reasoning via Chain-of-Thought (CoT) prompting. However, standard data annotation pipelines show the teacher model the ground-truth (GT) future trajectory before it writes the reasoning trace.
The authors discover that this creates a severe trajectory anchoring bias: instead of driving based on scene evidence, the model works backward to justify a pre-revealed outcome. This leads to post-hoc rationalizations and severe hallucinations—such as inventing traffic signs that don't exist just to explain why it decided to turn. Hiding the trajectory completely and forcing open-ended coordinate generation is also problematic, as it entangles high-level reasoning with complex, low-level geometric synthesis.
Methods
To solve this, the authors propose a two-part framework:
- AD-MCQ (Autonomous-Driving Multiple-Choice Question): Instead of generating raw coordinates, the planning task is recast as selecting from a set of discrete, scene-specific candidate trajectories generated via a K-means codebook ($K=8192$). This turns trajectory planning into an exactly verifiable decision.
- DEFT-RLVR (Deferred Exposure of Future Trajectories for RLVR): To prevent the model from using the multiple-choice options as a new shortcut, exposure to the candidates is deferred using a two-turn interaction:
- Turn 1 (Causal Decision Reasoning): The model analyzes the scene context alone (visuals, ego history, and navigation commands) and commits to a high-level decision (e.g., "decelerate to a stop").
- Turn 2 (Executable-Trajectory Grounding): The candidate trajectories are revealed, and the model must select the option that best matches its Turn 1 commitment.
The entire sequence is optimized jointly using Group Relative Policy Optimization (GRPO). Crucially, they use Structured Rubric Rewards: an offline vision-language model generates a scene-specific rubric, and an online text-only judge evaluates the Turn 1 reasoning trace only if the final trajectory selection is correct. This keeps the evaluation efficient and focused on text quality without visual token dilution.
Datasets
The authors leverage data from the Waymo Open E2E dataset and an internal driving corpus, pooling over 514,000 scene-trajectory pairs. They introduce:
- AD-MCQ-500: A curated evaluation test split of 500 causally demanding, long-tail scenes (hard braking, sharp turns, stopping from motion) with structured hard distractors.
- nuScenes: Used as an out-of-distribution (OOD) external dataset to evaluate cross-domain transfer capability.
Key Insights and Ideas
- Information Order Matters: Revealing the answer before the rationale ruins the causal faithfulness of CoT. By simply changing when the model sees the future options, it is forced to actually "look" at the road and reason from evidence.
- MCQ as a Verification Interface: Multiple-choice formats are often criticized in LLM evaluation, but here, casting trajectory planning as an MCQ elegantly decouples high-level causal reasoning from low-level geometric coordinate generation.
- Avoiding the Specialization Trap: Directly fine-tuning VLMs on trajectory tokens (SFT) causes a catastrophic drop in general visual capabilities because it warps the pretrained token distribution. DEFT-RLVR actually preserves or slightly improves general visual perception because RL reinforces existing reasoning pathways rather than forcing the internalization of a new coordinate vocabulary.
Key Results
- Driving Performance: On Qwen3-VL-8B-Instruct, DEFT-RLVR boosted MCQ accuracy from 28.1% (base) to 77.9%, outperforming the candidate-visible baseline (JEFT + RLVR) by 16.8 percentage points.
- Causal Faithfulness: Human and automated metrics showed substantial gains in Causal Faithfulness Scores (CFS) and High-Level-Decision Consistency (HLD), proving the reasoning was genuinely grounded in scene evidence.
- Generalization & Retention: DEFT-RLVR maintained or slightly improved scores across 12 general vision-language benchmarks (averaging ~56.09% vs 54.81% base), whereas standard imitation learning (SFT) caused general capabilities to plummet toward zero. It also successfully generalized to the out-of-distribution nuScenes dataset.
Limitations and Future Work
- Discrete Codebook Constraints: While an 8,192-token codebook provides a low reconstruction error (~0.29m Average Displacement Error), it still quantizes continuous driving paths, which might limit fine-grained control in highly complex or dynamic environments.
- Dependence on a Strong Judge: The framework relies on an offline/online teacher model (like Qwen3.6-35B) to generate rubrics and grade reasoning traces, introducing a dependency on external model capabilities during training.
- Future Directions: The authors suggest exploring how to scale candidate construction to dynamically changing environments and extending this verifiable reasoning paradigm to closed-loop simulation testing.
Summary updated