Session 1

With each passing day, the vision of language agents capable of driving scientific discovery feels ever closer to reality. LLM-based systems that can follow complex instructions, use external tools, and take actions to complete sophisticated scientific tasks are beginning to resemble AI scientists.

Yet, despite this accelerating progress, today’s LLMs continue to struggle with basic reasoning and generalization failures (e.g., the Reversal Curse). In this talk, I will explore both sides of this paradox and reflect on a central question: What would it take to build truly transformative language agents for scientific discovery? I will advocate for two key lines of effort: rigorous benchmarking and fundamental understanding. First, we will discuss rigorously benchmarking agents in realistic scientific tasks, including chemistry and scientific coding, highlighting both their capabilities and limitations. Then, we will turn inward to examine the architectural foundations of these systems, exploring the limits of the Transformer in multi-step reasoning and the fundamental causes of failures like the Reversal Curse.

By contrasting rapid progress with foundational constraints, this talk aims to surface some of the key ingredients and missing pieces on the path toward truly transformative language agents for scientific discovery.

Huan Sun is an endowed College of Engineering Innovation Scholar and an associate professor in the Department of Computer Science and Engineering at The Ohio State University. Her research focuses on natural language processing and artificial intelligence, with particular interest in large language models, agents, and their safety risks. Her recent notable work includes Mind2Web, SeeAct, MMMU, ScienceAgentBench, Grokked Transformers, AmpleGCG, and EIA. She has received many awards, including Best Paper Finalist at CVPR 2024, two Honorable Mentions for Best Papers at ACL 2023, ACM SIGMOD Research Highlight Award 2022, Best Paper Award at BIBM 2021, NSF CAREER Award, and the SIGKDD Dissertation Runner-Up Award. Her team got third place in the inaugural Alexa Prize TaskBot Challenge in 2022, as the only award-winning team from the US.

Active design of therapeutic molecules requires the joint optimization of multiple, potentially competing properties. Multi-objective Bayesian optimization (MOBO) offers a sample-efficient framework for identifying Pareto-optimal drug candidates. At the heart of MOBO is the acquisition function, which determines the next candidate to evaluate by navigating the best compromises among the objectives. In this talk, I show a natural connection between the Pareto front and the extreme quantile of the joint cumulative distribution function (CDF). This link motivates the proposed Pareto-compliant CDF indicator and the associated acquisition function, BOtied. BOtied inherits invariance properties of the CDF well suited for the functional landscape of molecules. Moreover, an efficient implementation with copulas allows it to scale to many objectives. Outperforming state-of-the-art MOBO acquisition functions on a variety of synthetic and real-world problems, BOtied promises to drive model-based decisions for drug discovery.

Ji Won is a Principal Machine Learning Scientist at Prescient Design, Genentech. Her current research probes hierarchical, sparsity-inducing structures in high-dimensional data that can inform inference and adaptive decision-making. She focuses on developing algorithms in Bayesian optimization, calibration, and MCMC sampling inspired by challenges in molecular design. She received her Ph.D. in Physics from Stanford University, where she worked on hierarchical Bayesian methods for cosmology. During her Ph.D., she interned at NASA Ames and the Center for Computational Astrophysics at the Flatiron Institute. She holds a B.S. in Mathematics and a B.S. in Physics from Duke University.

This talk will explore how generative AI agents can enable drug discovery and development. I’ll introduce the Virtual Lab—a collaborative team of AI scientist agents conducting in silico research meetings to tackle open-ended R&D projects. The Virtual Lab designed new nanobody binders to recent Covid variants that we experimentally validated. Then I will discuss some interesting opportunities in designing and optimizing multi-agent interactions.

James Zou is an associate professor of Biomedical Data Science, CS and EE at Stanford University. He works on advancing the foundations of ML and in-depth scientific and clinical applications. Many of his innovations are widely used in tech and biotech industries. He has received a Sloan Fellowship, the Overton Prize, an NSF CAREER Award, two Chan-Zuckerberg Investigator Awards, a Top Ten Clinical Achievement Award, several best paper awards, and faculty awards from Google, Amazon, Adobe and Apple. His research has also been profiled in popular press including the NY Times, WSJ, and WIRED.