Paper

Transformers Provably Learn Graph Search: Training Dynamics and the Exponential Power of Depth · ByteTrail