Generalizable Reasoning through Compositional Energy Minimization

TL;DR

Generalization is a key challenge in machine learning, specifically in reasoning tasks, where models are expected to solve problems more complex than those encountered during training…

Opening excerpt from the authors’ abstract. source