Besides, we also introduce *stochastic* RINS where we select the number of recursion rounds from a binomial distribution. This *improves* performance in SigLIP (despite also *saving* training flops). But in LM, there is a tradeoff between flexibility and maximum performance gain.
