FearCaut-Qwen: Affective Steering in a Vision-Language Model Shifts the Decision Criterion for Hazard Assessment
The study investigates why Vision-Language Models (VLMs) are overly cautious when deciding if a scene is hazardous. By using mechanistic interpretability, researchers localized an affective circuit related to fear and used activation steering to manipulate it. This manipulation successfully shifted the model's decision criterion, proving that the model's willingness to flag a hazard is a distinct system property from its ability to correctly identify unsafe objects.