Vision Language Models See What You Want but not What You See

Abstract

Knowing others' intentions and taking others' perspectives are two core components of human intelligence that are typically considered to be instantiations of theory-of-mind. Infiltrating machines with these abilities is an important step towards building human-level artificial intelligence. Recently, Li et al. built CogDevelop2K, a data-intensive cognitive experiment benchmark to assess the developmental trajectory of machine intelligence. Here, to investigate intentionality understanding and perspective-taking in Vision Language Models, we leverage the IntentBench and PerspectBench of CogDevelop2K, which contains over 300 cognitive experiments grounded in real-world scenarios and classic cognitive tasks, respectively. Surprisingly, we find VLMs achieving high performance on intentionality understanding but lower performance on perspective-taking. This challenges the common belief in cognitive science literature that perspective-taking at the corresponding modality is necessary for intentionality understanding.