Size (or length) generalization is a key challenge in designing neural modules to perform algorithmic tasks. Specifically, when can a neural model with bounded complexity generalize to problem instances of arbitrary size? In this talk, I will focus on approaches to achieve size generalization by "aligning" the neural models with certain algorithmic structures, so as to facilitate a neural model learning "procedures" instead of merely fitting data. I will first present a theoretical result to show the benefit of algorithmic alignment in extrapolating for the graph shortest path distance estimation. We will then present examples of designing practical and efficient neural models for various geometric optimization problems via algorithmic alignments.