Looped and recurrent Transformers challenge depth limits

SMELT, loop-native attention residuals, SanSi, and Multi-Stream Attention Residuals suggest recurrent computation can improve scaling efficiency, extrapolate depth, and reduce iso-loss compute. Newly reviewed LatentLoop adds asymmetric compute allocation across loop visits, reporting lower NLL, estimated compute savings, and benchmark gains across 15 configurations. Broad-task generality, frontier-scale training, and real-hardware efficiency remain unresolved.