A new study from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) demonstrates a significant advancement in AI’s ability to interpret and reason about the physical world. The research introduces a framework where large language models (LLMs) are used to generate formal, logical representations of complex, real-world scenarios, such as those depicted in the children’s book ‘Where’s Waldo?’. This logical code is then executed by a theorem prover to answer questions about the scene, effectively separating the challenge of visual understanding from the task of logical reasoning. The system, named ‘LILA’, outperformed previous methods on a suite of visual reasoning benchmarks, showing particular strength in tasks requiring compositional reasoning—understanding how objects and their relationships combine to answer a query. The work suggests a promising path toward AI that can more reliably understand and interact with its environment by leveraging the complementary strengths of neural networks for perception and symbolic systems for reasoning. Read the full article at: https://technologyreview.com/2023/10/18/1081819/mit-ai-reasoning-wheres-waldo/