From Research Goals to Scientific Closure: Evaluating Agents Through Evidence-Supported Physics Research

Researchers introduced PhysClosureBench, a benchmark of 50 expert-verified, goal-specified research tasks spanning strong-field, attosecond and neighboring physics areas, to test whether AI agents can organize scientific investigations that reach well-supported conclusions. Each task specifies a research goal, and the benchmark evaluates agents on evidence-supported closure rather than open-ended exploration.

Can AI agents organize scientific investigations that lead to well-supported conclusions? To evaluate this capability, we introduce PhysClosureBench, a benchmark of 50 expert-verified, goal-specified research tasks spanning strong-field, attosecond, and neighboring physics areas. Each task specifies