Atomic Canyon releases nuclear AI benchmark developed with Idaho National Laboratory

Atomic Canyon announced the release of NuclearBench, an open-source benchmark developed with Idaho National Laboratory to test how language models answer nuclear-energy questions. The benchmark uses 280 questions and evaluates both answer accuracy and faithfulness to the source material supplied to a model. Organizations can use the public resource to compare models for nuclear power operations. It extends Atomic Canyon's earlier FermiBench work, which tests whether models retrieve the right information from nuclear records, into evaluating generated answers. NuclearBench is designed to be updated as new models become available.