Uncovering Mesa-Optimization Transformers in Deep Learning - arxiv.org