Intrinsic Task Symmetry Drives Generalization in Algorithmic Tasks

Researchers propose that intrinsic task symmetries drive 'grokking' - the sudden transition from memorization to generalization in neural networks. The study identifies a three-stage training process and introduces diagnostic tools to predict and accelerate the onset of generalization in algorithmic reasoning tasks.