Discovering and Engineering the Computation Underlying Large Intelligent Agents

Explore a thought-provoking colloquium talk by Pratyusha Sharma from MIT, who investigates how deep networks learn and represent latent compositional structure in language and intelligent behavior. The presentation offers three perspectives: experimental methods to characterize learned solutions in Large Language Models and improve their generalization without gradients, deciphering the structure of sperm whale communication systems to reveal a unique combinatorial language, and applying these insights to develop embodied agents with hierarchical and compositional "language of thought" capabilities for enhanced reasoning and planning. Sharma, a Ph.D. student at MIT's Computer Science and Artificial Intelligence Lab advised by Antonio Torralba and Jacob Andreas, has published in prestigious journals including Nature Communications and presented at TED AI. Her interdisciplinary work spans machine learning, natural language processing, robotics, and marine biology, and has been featured in the New York Times, National Geographic, and BBC.