interpretabilityfactbullishActual routing decisions in MoE models explain token representations better than matched alternative routes, even though expert subspaces overlapMachine Learning02 Aug 2026http://arxiv.org/abs/2607.28308v1