Stop, Collaborate, Don't Collapse

Non-contrastive self-supervised methods like BYOL and SimSiam train two branches to predict each other's representations from different augmented views (a pretext task), but drop the negative-pair term used in contrastive losses like SimCLR's InfoNCE. Without negatives to push dissimilar samples apart, what specific mechanism do these methods rely on to avoid representation collapse (all outputs converging to a constant vector)?

A momentum encoder alone, with no architectural asymmetry between branches

An asymmetric predictor network on one branch combined with a stop-gradient on the target branch

A much larger batch size than contrastive methods use, which implicitly supplies negative pairs

L2 weight decay applied only to the final projection layer