Review for NeurIPS paper: EvolveGraph: Multi-Agent Trajectory Prediction with Dynamic Relational Reasoning

Weaknesses: The ability of EvolveGraph to uncover known dynamic relations is not explored in as much detail as it could be. More specifically, the one synthetic experiment designed to evaluate this is somewhat simple, in that all relations change from "active" to "inactive" for all entities at the same moment in time, and this switch happens once. What happens when relations change at different times for different variables? What happens if the re-encoding gap is "out of sync" with the actual change in relations? How well does the model perform if relations change multiple times aperiodically?