Differentiable complexity controller for grokking via algorithmic information dynamics
Researchers propose a certified, differentiable complexity estimator that extends algorithmic information dynamics beyond the piecewise-constant Block Decomposition Method. This allows gradient-based control of complexity during training, which they apply to the phenomenon of grokking. The work is published as an arXiv preprint in machine learning.