Why it is a question
Assran et al. (LIT-231, read in full in NOTE-850) show that MSN’s regularizer is a KL penalty pulling soft cluster sizes toward uniform and that SwAV enforces equal sizes outright. With mini-batches of two to eight ImageNet classes, SimCLR, MSN and VICReg lose 11 to 18 points of ImageNet linear accuracy while MAE and data2vec do not (THEORY-306). For VICReg the prior is derived only for a simplified loss with γ ≫ α, with an informal last step and a cited property of K-means. For SimCLR nothing is derived. The imbalance is extreme and artificial, each cell is a single run, and the methods compared differ in architecture. Until 2026-10-19 this was THEORY-306’s promote_when. THEORY-306’s body disclaimed it, so the clause moved here when the title was narrowed.
What would answer it
For the prior: a derivation of the cluster-size preference for the losses as used (VICReg with its hinge variance term, SimCLR at finite temperature), or a measurement of the cluster-size distribution that trained VICReg and SimCLR models induce on data with known unequal classes. For the effect: the imbalance experiment repeated with batches sampled uniformly from a naturally long-tailed dataset, at one architecture across methods, with seeds. It would answer no for a method if its learned partitions track unequal class sizes as well as an instance-based method’s do.
What would not answer it: more results with two-class batches, or gains from a prior matched with labels.