Can an AI agent run a scientific investigation from start to finish and stop at a conclusion the evidence supports? A new arXiv preprint addresses this question by introducing PhysClosureBench, a benchmark made up of 50 expert-verified, goal-specified research tasks in physics.
The benchmark is built around the idea of scientific closure: an investigation should end only when the evidence supports a conclusion. Each task gives an agent a specific research goal, and the agent is judged on whether it can organize the work and arrive at a well-supported endpoint.
The abstract does not specify which physics subfields the tasks cover or how the authors score agent performance. Those details will matter for assessing how broadly the benchmark applies, but the core question—whether agents can close a research loop—is clearly defined.