EAIRA

Argonne-led preprint proposing four classes of evaluation for AI models used as scientific research assistants: multiple-choice questions, open responses, laboratory-style experiments and field-style experiments.