Date: December 3rd at 11am

Location: Picower Seminar Room, 46-3310

Title: Modular Cognitive Architecture in Natural and Artificial Intelligence

Abstract: How do humans use language to solve complex problems that require coordination between multiple cognitive faculties? How can we engineer these behaviors into performant artificial intelligence (AI) systems? When scheduling a doctor appointment, we seamlessly integrate linguistic comprehension, memory retrieval, and constraint satisfaction. When answering “What is the sum of three, six, and four?” we parse language, perform arithmetic, and verbalize a response. Despite being nearly effortless, this functional use of language has remained difficult to model and engineer. Progress has followed two main traditions: (1) Symbolic approaches excel at systematic reasoning with correctness guarantees but struggle with flexibility and real-world complexity. (2) Neural approaches, such as large language models (LLMs), excel in learning from massive data and handling variation, but lack systematic generalization and reliable formal reasoning. Neither paradigm alone captures the full scope of human-like intelligence. This thesis investigates how we can design AI systems that combine neural flexibility with symbolic robustness, drawing from principles of human cognitive modularity. I develop neurosymbolic architectures where LLMs serve as semantic parsers extracting structured representations from language, which specialized symbolic modules then process for formal reasoning. This recasts language models from encapsulating entire reasoning systems to serving as components within larger cognitive architectures. The thesis proceeds in four parts. First, I evaluate what conceptual knowledge LLMs possess through common sense reasoning tasks, finding systematic shortcomings in areas like spatial reasoning that motivate architectural improvements beyond scaling (Chapter 2). Second, I develop a neurosymbolic architecture for logical reasoning combining an LLM parser with a symbolic theorem prover, which substantially outperforms LLM-only approaches, but reveals that parsing natural language into formal representations struggles with vagueness (Chapter 3). Third, I address vagueness through a probabilistic semantic parsing framework, where programs contain “holes” filled by distributions over completions, showing that LLMs can be well calibrated to graded human interpretations (Chapter 4). Fourth, I develop an efficient algorithm for probabilistic semantic parsing that is faster and more accurate than prior approaches (Chapter 5). Together, these contributions establish principles and tools for building AI systems that reason through modular architectures inspired by human cognition, advancing both our scientific understanding of language and thought, as well as engineering solutions for more robust artificial intelligence.