Looping Transformer Layers Cuts Model Size While Boosting Math and Coding

Original: ARLoop: Fully-Open Pre-trained Looped Transformers with Attention Residuals

Source: arxiv.org ↗