MIT researchers found LLMs may form internal simulations of reality while improving language skills. A model achieved 92.4% accuracy on Karel puzzles without direct exposure, suggesting deeper understanding beyond mimicry.