Reasoning agents are the next frontier of AI. My research investigates how we can develop AI reasoning agents for the benefit of society, focusing both on designing novel AI reasoning methodologies, and adapting them for applications such as health, education, and global fairness. My group’s research draws on methods in natural language processing, deep learning, machine learning, and artificial intelligence to investigate these problems.
Topics that I focus on include:
LLM Representations of Knowledge. Figuring out how to go from an LLM to a reasoning agent requires understanding what LLMs know, how they represent that knowledge, and how they compose that information internally. My research investigates how LLM subnetworks (and other LLM representations) encode discrete forms of knowledge, how those representations can be modified, and how closely they align with measurements of the human brain [1,2,3,4,5,6]
Reasoning Algorithms. LLMs fail dramatically and unexpectedly when presented with seemingly simple reasoning problems that humans effortlessly solve. We draw on methods from diverse research areas (e.g., symbolic systems, neuroscience, cognitive science, psychology) to devise new methods and frameworks for LLM reasoning. [1,2,3,4,5,6,7]
Large-scale AI Development. LLMs behave differently at small scale compared to large scale. Our work bridges the research gap between these two settings by developing open-source, open-weight, and open-data foundation models. We specifically focus on developing multilingual models trained on compliant data to enable use in diverse regulatory and cultural settings [1,2,3,4]
AI Democratization. Much like previous generations of AI advancement, AI reasoners will be experienced differently by different groups based on their current digital and AI maturity. Across practice areas in health, education, and global fairness, my works works with end users to develop models, theories, and evaluations that enable responsible development of LLM-based AI. [1,2,3,4]
