Multi-head Transformers Provably Learn Symbolic Multi-step Reasoning via Gradient Descent

Devs

Multi-head Transformers Provably Learn Symbolic Multi-step Reasoning via Gradient Descent | Read Paper on Bytez