MaD Physics Benchmark Tests Agents Under Measurement Constraints

A new benchmark called Measuring and Discovering Physics (MaD Physics) evaluates agents' ability to make informative measurements and conclusions under constraints on measurement quality and quantity. Proposed in arXiv:2605.10820, the benchmark addresses a gap in existing scientific discovery benchmarks, which focus on static knowledge-based reasoning or unconstrained experimental design. MaD Physics includes three environments, each based on a distinct physical law, with altered physics to mitigate contamination from prior knowledge. The work highlights the resource-constrained nature of scientific discovery, where trade-offs between measurement quality and quantity are critical.