ZendoWorld Challenges AI in Active Visual Concept Induction

Building truly intelligent systems requires agents that can not only interpret complex visual information but also formulate hypotheses about underlying rules and then strategically design experiments to validate those hypotheses. To rigorously evaluate these capabilities, a new interactive environment called ZendoWorld has been introduced. ZendoWorld presents AI agents with a task where they must infer a logical rule governing visual scenes. Agents then propose new scenes to the environment to gather feedback and refine their understanding. Various AI approaches, including VLM reasoning, Bayesian methods, and neuro-symbolic techniques, were tested within this framework. The findings indicate that simply predicting labels accurately for observed examples does not mean an AI has grasped the underlying rule. A major weakness identified is that VLM-based agents struggle to propose informative experiments, failing to actively reduce uncertainty in their hypotheses. Human performance on the same task highlights a substantial gap in inductive reasoning, particularly for more complex rules, pointing to clear areas for future AI improvement.