robotics
The Driving Model Looked Harder—Then Moved Closer
A bounded inference-time attention bias shifted planned trajectories toward a highlighted traffic actor without retraining the model.
Summary
A bounded inference-time attention bias shifted planned trajectories toward a highlighted traffic actor without retraining the model.
The experiment added a pre-softmax bias to detector-localized visual tokens in 50 synthetic lane-change scenarios. Trajectory displacement increased with the bias, reaching about 17 centimeters on average and up to roughly 140 centimeters laterally. The reasoning text did not change because the mask never reached that pathway, and the steering tended to move toward the attended actor rather than encode a safe target behavior—an important limit on treating attention control as safety control.
Why it matters
A bounded inference-time attention bias shifted planned trajectories toward a highlighted traffic actor without retraining the model.
Limits and context
- The reasoning text did not change because the mask never reached that pathway, and the steering tended to move toward the attended actor rather than encode a safe target behavior—an important limit on treating attention control as safety control.
Key claims
A bounded inference-time attention bias shifted planned trajectories toward a highlighted traffic actor without retraining the model.
Qualification: The reasoning text did not change because the mask never reached that pathway, and the steering tended to move toward the attended actor rather than encode a safe target behavior—an important limit on treating attention control as safety control.
Evidence: source-2026-08-19-011
Sources
- arXiv preprint 2608.17095arXiv · primary research
Corrections
No corrections have been recorded for this story.