ResBM: a new transformer-based architecture for low-bandwidth pipeline-parallel training, achieving 128× activation compression [R]

Macrocosmos has released a paper on ResBM (Residual Bottleneck Models), a new transformer-based architecture designed for low-bandwidth pipeline-parallel training. https://arxiv.org/abs/2604.11947 ResBM introduces a residual encoder-decoder bottleneck across pipeline boundaries, with the goal of reducing inter-stage communication while preserving an explicit low-rank identity path. The paper reports SOTA 128× activation compression without significant loss in convergence relative to uncompress