Ask a Question

Prefer a chat interface with context about you and your work?

Cognitive Paradigms for Evaluating VLMs on Visual Reasoning Task

Cognitive Paradigms for Evaluating VLMs on Visual Reasoning Task

Evaluating the reasoning capabilities of Vision-Language Models (VLMs) in complex visual tasks provides valuable insights into their potential and limitations. In this work, we assess the performance of VLMs on the challenging Bongard Openworld Problems benchmark, which involves reasoning over natural images. We propose and evaluate three human-inspired paradigms: holistic …