Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs

A new study on arXiv (2608.01755v1) dives into the biases found in Vision-Language-Action (VLA) models used for self-driving cars. The researchers found that when the teacher model is trained with the actual future path data, it creates what's called a 'trajectory anchoring bias.' This bias makes the model more likely to back up the given outcome instead of making decisions based on what’s happening in the scene, leading to less accurate reasoning and more false conclusions, especially in tricky situations. To tackle this, the authors suggest a method called 'Autonomous-Driving Multiple,' which aims to improve decision-making without relying on open-ended path generation. The paper is pending publication and can be viewed on arXiv.