Learning Trajectories Leave Decodable Traces in Neural Network Weight Space and Internal Representations
arXiv:2611.24782v1 Announce Type: cross Abstract: Neural networks trained on the same task develop related weight-space structures, yet remain distinguishable after learning.