Pharma LPII- Agentic Uncertainty for Robust Scientific Discovery
AI agents are increasingly able to analyze scientific data, run computational experiments, and make recommendations. This project will study how uncertainty can be modeled and used throughout that process to make agent-driven science more robust.
Faculty Advisor:
- Jason Wu
Description:
AI agents are increasingly able to analyze scientific data, run computational experiments, and make recommendations. This project will study how uncertainty can be modeled and used throughout that process to make agent-driven science more robust. Before an experiment, an agent may be uncertain about the scientist’s actual objective, such as whether a drug candidate should prioritize potency, safety, or another property, and may need to clarify that goal. During an experiment, the agent may face several reasonable analysis choices and need to determine whether it can proceed on its own or should ask the scientist for input. After the work is complete, we will study how to audit the agent’s trajectory for signs that its conclusions may be brittle or poorly supported, and how to highlight the parts of the process that deserve closer human inspection. Students will investigate a variety of methods, such as visualization, system-building, and model training to attempt to address these problems.
Prerequisites:
Python programming. Experience with machine learning, data science, human-computer interaction, or AI agents is helpful but not required. Students from Computer Science and related areas are welcome.