A new benchmark, SCICONVBENCH, evaluates LLMs on multi-turn clarification for task formulation in computational science. It assesses their ability to elicit missing information and correct internally contradictory requests. Current frontier models show room for improvement.