GazeVLM: Active Vision via Internal Attention Control for Multimodal Reasoning
GazeVLM introduces a method for human-like visual reasoning by internalizing metacognitive oversight directly into a Vision-Language Model's attention mechanism. It autonomously generates gaze tokens to create a continuous suppression bias, allowing the model to focus on relevant details while maintaining awareness. This enables fluid switching between global and local spatial reasoning without needing external tools.