Subspace Modularity Found in Vision and Language Models

Research identifies functional modularity at the orthogonal and linear subspace levels in neural networks, even when unit-level modularity is absent. This framework uses selective ablation to prove that task-relevant computations occupy independent subspaces.