Glance-or-Gaze: AI Framework for Adaptive Visual Search

Researchers have introduced Glance-or-Gaze (GoG), an autonomous system that transitions extensive multimodal models from merely observing to engaging in active visual planning. GoG features a Selective Gaze mechanism that intelligently decides whether to focus on the broader context or to concentrate on high-value areas, effectively filtering out irrelevant data prior to retrieval. To improve performance on intricate visual queries, a dual-stage training method known as Reflective GoG Behavior Alignment has been developed. This work tackles the challenges posed by the limitations of static parametric knowledge in large multimodal models and the indiscriminate retrieval of entire images in search-augmented methodologies. The paper can be found on arXiv with ID 2601.13942.