interpretabilityfactneutralComplex behaviors like refusal to answer harmful queries live in multi-dimensional subspaces rather than single linear directionsMachine Learning27 Jul 2026http://arxiv.org/abs/2607.02396v1