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

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 the research objective and essential scientific constraints, while leaving the

This work introduces a novel benchmark for evaluating AI agents' ability to conduct scientific investigations, moving beyond simple task completion to assessing the rigor and completeness of scientific conclusions. For ML engineers, understanding such benchmarks is crucial for developing and evaluating agents capable of complex, multi-step reasoning and evidence synthesis.