Giacomo Camposampiero

We introduce a novel, scalable framework to evaluate compositional generalization, leverage it to evaluate more than 5k models, and propose a family of neural models pushing the Pareto frontier on this task.