Phy-SRBench: A Physics Benchmark for Scientific Equation Discovery with Symbolic Regression

Scientific discovery has driven conceptual and technological advances in all areas, including non-scientific domains. Symbolic regression, mainly focusing on automating scientific discovery towards making new ones, suffers key limitations in the key components, the datasets. This paper discusses a new dataset benchmark that originates from physics, i.e., high-energy physics. These equations are domain-specific, therefore not necessarily memorized by large language models, and can be expressed in different forms thus allowing for different sets of input/output pairs.