Evals
Measure what intelligent systems actually do.
Evals develops empirical methods for measuring AI behavior, capability, uncertainty, robustness, cooperation, and failure. Our goal is to create evaluations that reveal not only whether a model succeeds, but how and why particular behaviors emerge.
Questions and areas of inquiry
Experimental approach
We aim to define testable hypotheses, build controlled conditions, and examine behavioral evidence across repeated trials. Our research approach considers uncertainty, alternative explanations, and the limits of each measurement.
These areas describe our research direction. Specific studies and results are shared through the research archive as they become available.
Discuss Evals ResearchSignal study.
An illustration of the observation interface.
Illustrative simulation — not live research data.
An open record of inquiry.
Research papers, experiments, tools, and observations from our work.
Research archive in development.
Future research outputs appear here as they become available.