Reasoning-augmented vision-language-action (VLA) driving models can perceive and reason about safety-critical hazards, yet sometimes fail to translate that awareness into action. Researchers characterize this as a perception-action binding gap: hazard information exists inside the model but exerts too little control over trajectory generation. Using matched simulated counterfactuals and logged driving replay, they show that in-path hazards are encoded in the visual representation, but the planned longitudinal response tracks ego-speed history and reacts only weakly to hazard proximity or the model's own reasoning trace.
The authors also demonstrate that the trajectory generator is capable of braking, but can fail due to the binding gap. As a fix, they inject a hazard-conditioned energy into the frozen flow-matching sampler of the Alpamayo-1.5 model, without retraining or modifying any weights. With exact obstacle geometry, this guidance reduces footprint intrusion from 58.1% to 14.5% and recovers safe braking whenever sufficient stopping distance remains, eliminating collisions entirely in their braking-relevant closed-loop battery.
The work separates hazard perception from hazard control in driving VLAs, showing that unsafe behavior can arise even when the model possesses both the evidence and the ability to act on it. The proposed generation-time intervention offers a simple path to bridging this gap in existing models, without the cost of retraining.