Roughly, when you train a model to make its predictions align to its own predictions in some way, you create a scenario where the simplest "correct" solution is to output a single value under diverse inputs, aka representation collapse. This guarantees that your predicted representations agree, which is technically what you want it to do but it's degenerate.

EMA helps because it changes more slowly than the learning network which prevents rapid collapse by forcing the predictions to align to what a historical average would predict. This is a harder and more informative task because the model can't trivially output one value and have it match the EMA target so the model learns more useful representations.

EMA has a long history in deep learning (many GANs use it, TD-learning like DQN, many JEPA papers, etc.) so authors often omit defense of it due to over-familiarity or sometimes cargo culting. :)