DeepLoop Scales Transformer Depth via Looped Parameters
July 14, 2026
DeepLoop uses looped Transformers to increase unrolled depth by reusing a compact stack of physical blocks for multiple rounds. This method formalizes a tied-depth effect using a visit-alignment coefficient to manage residual-scaling challenges.
HOW THIS AFFECTS YOU
●
researcherYou can explore new ways to scale sequential computation without increasing stored parameter counts.