Evolutionary Game Theory Explains Shortcut Learning in AI

A new paper on arXiv (2605.02658v2) provides a formal definition of core and shortcut features in deep learning, using evolutionary game theory to analyze shortcut bias. The authors model data samples as players and neural tangent features as strategies, assuming core and shortcut subnetworks. They find gradient descent (GD) primarily optimizes the shortcut subnetwork, while stochastic gradient descent (SGD) primarily optimizes the core subnetwork, leading to distinct stochastically stable states. A continuous stochastic differential equation reveals the impact of data noise and optimization on shortcut bias, offering theoretical insight into why models rely on non-essential features.