Subspace ablation reveals functional modularity in vision, language, and RNN models despite overlapping neurons

The paper defines three nested modularity levels and supplies a single optimization framework to recover them. Unit-level requires disjoint neuron sets. Orthogonal asks for a rotation yielding independent subspaces. Linear asks for a bounded map isolating task-relevant computations. Causal validation uses selective ablation on the recovered subspaces rather than single units.

Vision models show content subspaces align with higher visual cortex and style subspaces with lower visual cortex. Language models show semantic subspaces outperform syntactic ones in predicting language cortex activity. Multi-task RNNs previously reported as non-modular yield sharp task double dissociations once linear decompositions are applied. These patterns hold across the reported model classes.

Operationally the result supplies a practical route for targeted intervention: subspace ablation can suppress or preserve specific computations even when no neuron is specialized. This directly affects mechanistic interpretability pipelines that currently rely on neuron-level localization and extends to biological data alignment studies that treat single-unit recordings as ground truth.

Next steps include testing whether the same linear decompositions remain stable under continued training and whether they transfer across model scales without re-optimization.