Visual Interestingness Decoded from Multimodal AI Using Neuroscience Methods

A team of researchers from an undisclosed institution investigated the multimodal vision-language model Qwen3-VL-8B to assess its encoding of human visual interest principles. They utilized a pre-established Common Interestingness (CI) score, derived from extensive human engagement data on Flickr, to analyze internal representations within the model's vision and language components using neuroscience-inspired techniques. Their findings indicated that CI information can be linearly decoded from the model's final layers, implying that transformer models might reflect certain elements of human attention. This research seeks to enhance the understanding of cognition and promote responsible AI applications in communication and marketing. The study was published on arXiv with the identifier 2605.08188.