A preprint submitted on 13 October 2026 introduces SCOPE, a framework in which multiple scientist agents independently pursue conjectures, along with ClosureBench, which the paper describes as a set of 60 closure-discovery tasks across five scientific domains.[1] The authors argue that most symbolic regression methods optimize candidate equations, but a better fit does not establish whether the scientific conjecture behind the equation should be kept, revised or rejected.[1] The authors wrote that a single conjecture may have several mathematical forms, and that similarly accurate equations may imply different explanations.[1]
In SCOPE, multiple scientist agents independently pursue conjectures through equation modeling, question formulation and active experimentation, then exchange evidence to revise their conjectures and research directions.[1] The findings are consolidated into a scientific report that links models, conclusions and evidence.[1] ClosureBench embeds an unknown algebraic relation in a known PDE system, and agents control experimental conditions in OpenFOAM to acquire new data for equation discovery.[1] Beyond equation accuracy, the authors report that the benchmark proposes three metrics for the discovery process: evidence faithfulness, mechanism insight and experimental effectiveness.[1]
The authors report that SCOPE achieves higher equation accuracy than diverse baselines on both ClosureBench and a real-world scientific discovery case, and that it produces more accurate, reliable and effective discoveries.[1] The authors wrote that existing equation-discovery benchmarks primarily assess recovered expressions, leaving the evidence and scientific conclusions underexamined.[1]