Agentic Compositional Visual Reasoning in the Large Language Model Era

This thesis studies how artificial intelligence systems can reason more effectively about visual information in real-world environments. It focuses on approaches that break complex visual tasks into multiple steps, adapt to new situations using limited data, and remain reliable when visual tools provide imperfect information. The research includes a large review of existing studies, proposes new reasoning frameworks, and introduces evaluation benchmarks that test visual reasoning across multiple viewpoints. Together, these contributions support the development of more robust, flexible, and collaborative visual intelligence systems inspired by human reasoning.