Researchers reveal how vision-language models infer scene context from isolated objects, uncovering gaps between accuracy levels that could affect AI robustness

Study systematically analyzes how vision-language models (VLMs) infer scene information from single objects presented on masked backgrounds

VLMs demonstrate above-chance performance in predicting both fine-grained scene categories and broad context (indoor vs. outdoor) from single objects alone

Object properties that influence human scene perception similarly modulate VLM performance, suggesting shared underlying mechanisms

Researchers found that accurate inference at one context level (e.g., object identity) does not guarantee accuracy at other levels (e.g., superordinate context), revealing partial dissociability in model predictions