Addressing divergent representations from causal interventions on neural networks