Low-Dimensional Execution Manifolds in Transformer Learning Dynamics: Evidence from Modular Arithmetic Tasks

arXiv:2602.10496v2 Announce Type: replace Abstract: We investigate the geometric structure of learning dynamics in overparameterized transformer models through carefully controlled modular arithmetic tasks. Our primary finding is that despite operating in high-dimensional para...

🔗 Read more: https://arxiv.org/abs/2602.10496

#News #Neuro #Software #Policy #AI #Space #Academic