Uploaded October 2025 | Updated September 2026, 6 hours ago
Scientific discovery has historically required the creative synthesis of paradigm-questioning insights, empirical observation, and mathematical
rigor—from Copernicus's hypothesis to Kepler's models to Newton's universal laws. Today, we stand at the threshold of automating this discovery
process using artificial intelligence, particularly large language models. However, current approaches often excel at recombining existing
knowledge while struggling with the creative skepticism and mathematical precision needed for genuine breakthroughs. This talk explores how
data-driven mathematical modeling serves as both a rigorous testbed and foundational component for reliable AI systems for scientific discovery.
Kazem discusses recent developments in leveraging language models for scientific model discovery, multi-modal approaches that combine symbolic
and data-driven representation learning, iterative scientific model discovery via programming with large language models, and novel benchmark
methodologies specifically designed to distinguish real discovery from memorization of existing scientific knowledge. Looking forward, he
outlines potential future directions including enhancing LLM creativity and paradigm-questioning capabilities, developing multi-modal capabilities
for direct integration of empirical data, and enhancing evaluation frameworks through collaboration with domain experts to validate authentic
scientific contributions and translate AI-generated hypotheses into scientific publications and breakthroughs.
Kazem is an AI Researcher at Capital One AI Foundations and a recent PhD graduate from Carnegie Mellon University, advised by Amir Barati Farimani. His research interests revolve around language models and AI-driven scientific discovery. His work explores leveraging large language models for
symbolic reasoning and mathematical model generation, developing AI systems that can discover and verify scientific models from data. During his PhD, he held research internships at Netflix Research and Electronic Arts, working on foundation models and large language model applications.
Scientific discovery has historically required the creative synthesis of paradigm-questioning insights, empirical observation, and mathematical
rigor—from Copernicus's hypothesis to Kepler's models to Newton's universal laws. Today, we stand at the threshold of automating this discovery
process using artificial intelligence, particularly large language models. However, current approaches often excel at recombining existing
knowledge while struggling with the creative skepticism and mathematical precision needed for genuine breakthroughs. This talk explores how
data-driven mathematical modeling serves as both a rigorous testbed and foundational component for reliable AI systems for scientific discovery.
Kazem discusses recent developments in leveraging language models for scientific model discovery, multi-modal approaches that combine symbolic
and data-driven representation learning, iterative scientific model discovery via programming with large language models, and novel benchmark
methodologies specifically designed to distinguish real discovery from memorization of existing scientific knowledge. Looking forward, he
outlines potential future directions including enhancing LLM creativity and paradigm-questioning capabilities, developing multi-modal capabilities
for direct integration of empirical data, and enhancing evaluation frameworks through collaboration with domain experts to validate authentic
scientific contributions and translate AI-generated hypotheses into scientific publications and breakthroughs.
Kazem is an AI Researcher at Capital One AI Foundations and a recent PhD graduate from Carnegie Mellon University, advised by Amir Barati Farimani. His research interests revolve around language models and AI-driven scientific discovery. His work explores leveraging large language models for
symbolic reasoning and mathematical model generation, developing AI systems that can discover and verify scientific models from data. During his PhD, he held research internships at Netflix Research and Electronic Arts, working on foundation models and large language model applications.
