Adaptive Computation

A neural network computation paradigm that allows a model to dynamically adjust the amount of computation based on input complexity, rather than applying a fixed number of computation steps to all inputs. The approach aims to improve both efficiency and reasoning capability.