Our paper "A mechanistic analysis of a transformer trained on a symbolic multi-step reasoning task" is accepted at ACL 2024
arXiv: https://arxiv.org/abs/2402.11917