NEBench R1 evaluates LLM reasoning in nuclear engineering
November 18, 2026
NEBench R1 is a reasoning-oriented benchmark featuring 7,546 questions from U.S. NRC reactor operator exams covering 18 nuclear facilities. Tested LLMs achieved accuracy between 34.0% and 73.9%, with an LLM-based judge used to evaluate the quality of their reasoning steps.
HOW THIS AFFECTS YOU
●
researcherYou can use this to test if LLMs are actually reasoning through nuclear engineering problems or just guessing.
●
policyThis provides a way to vet models before using them in highly regulated, safety-critical sectors like nuclear energy.