Attention Does Not Equal Relevance in Autonomous Driving
November 19, 2026
Analysis of 80,000 synthetic driving scenes using the PlanT 2.0 Transformer model shows that attention weights do not accurately capture the causal impact of objects on driving decisions. The study compares attention-based estimation against counterfactual analysis to test how much input elements actually drive model outputs.
HOW THIS AFFECTS YOU
●
researcherYou should be careful about using attention maps as a proxy for explainability in safety-critical systems.