Preprint
Aug 2026
Inference-Time Attention Steering for Vision-Language-Action Driving Models
A bounded additive pre-softmax attention bias on the visual tokens of detector localized traffic actors on Alpamayo-R1's Qwen3-VL backbone is studied, suggesting the bias governs where the model looks rather than encoding a target behavior.
D. Prasad, Lars Ullrich, Knut Graichen
· 1 citation