Does Human-Like Contextual Object Recognition Emerge from Language Supervision and Language-Guided Inference?

Human vision is an active, context-sensitive process that interprets objects in relation to their surroundings. While behavioral research has long shown that scene context facilitates object recognition, the underlying computational mechanisms--and the extent to which artificial vision models replicate this ability--remain unclear. Here, we addressed this gap by combining human behavioral experiments with computational modeling to investigate how structured scene context influences object recognition. Using a novel 3D simulation framework, we embedded target objects into indoor scenes, and manipulated contextual coherence between objects and scenes by using either intact scenes or their phase-scrambled versions. Humans showed a robust object recognition advantage in coherent scenes, particularly under challenging conditions such as occlusion, crowding, or non-canonical viewpoints. Conventional vision models--including convolutional neural networks (CNNs) and vision transformers (ViTs)--failed to replicate this effect. In contrast, vision-language models (VLMs), particularly those using ViT architectures and trained with language supervision (e.g., CLIP), approached human-like accuracy. This shows that semantically rich and category-structured representations are required for modelling context sensitivity. Notably, context sensitive behavior was closest to humans in VLMs when using language-guided inference at test time. This suggests that how a model accesses its representations during inference is relevant for enabling context-sensitive behavior. Together, this work offers steps towards a computational account of contextual facilitation of objects by scenes, and highlights zero-shot inference as an interesting alignment metric when benchmarking artificial and biological vision.