Local Perception Improves Visual Reasoning Generalization

The human visual system processes information through a series of local, foveated glimpses, a stark contrast to most modern computer vision models that process entire images globally in a single pass. This paper explores whether local, sequential vision models offer fundamental computational advantages, particularly concerning visual state tracking and length generalization. Inspired by recent studies on length generalization in language models, the researchers trained vision models on simple tasks requiring the aggregation of local information across an image. Their experiments revealed that, much like language models, global vision models can develop shortcuts, leading to a failure in generalizing across varying task lengths or complexities. Crucially, the study demonstrates that recurrent vision policies employing strictly local perception can effectively overcome these generalization failures. This suggests that local attention might be a vital, yet often overlooked, requirement for achieving robust compositional generalization in artificial vision systems. The findings highlight a potential path toward more human-like and adaptable AI vision.