This document summarizes a research paper on Neural Input Search (NIS), a method for optimizing vocabulary sizes and embedding dimensions for recommendation models. NIS uses reinforcement learning to search for the optimal multi-size embedding configuration for different features. It defines the problem of finding the best vocabulary sizes and embedding dimensions for each feature to maximize a model's objective function, subject to a memory budget constraint. NIS models outperformed baselines on public recommendation datasets using fewer parameters and were able to achieve better performance even when exceeding the memory budget.