ESM-bench: Can AI Agents Understand Earth System Model Physics and Code?
Date:
Oral presentation on ESM-bench at the 4th Annual Good Systems Smart Cities and AI Innovations Symposium at the University of Texas at Austin.
Talk Focus
ESM-bench evaluates whether AI agents can understand Earth System Model physics and code, rather than producing plausible but ungrounded answers about scientific software.
Key Points
- Benchmarking AI agents on Earth System Model physics and code understanding
- Evaluating model behavior with a structured task set and scoring rubric
- Connecting geoscience, AI evaluation, and trustworthy AI systems for scientific modeling
Related Work
This talk builds on the ESM-bench preprint and project: