A mixture-of-experts model can route the same token through a completely different expert without changing a single weight - and a new study argues that fact alone tells you almost nothing about whether the model's behavior actually shifted.
Researchers tested seven open-weight MoE checkpoints across two domains, isolating routing changes from actual content changes in a single frozen forward pass. They found the routing decision itself accounts for a narrow slice of a block's output - varying by about 1.6x - while its downstream visibility in the residual stream varies more, about 3.2x. On PubMedQA, swapping which experts handle a token moved outputs by less than half of what changing the surrounding context did, and most of that movement looked statistically indistinguishable from noise.
That matters because router activity gets used as a quick diagnostic: if a model sends a token to a different expert, researchers often assume something meaningful changed, including in self-distillation setups where a demonstration-aware teacher trains a query-only student. This paper says that assumption needs a footnote - a routing change is not proof of a behavioral change until someone actually measures the downstream exposure and tests outputs directly.
MoE interpretability was already messier than dense-model interpretability, and this adds a specific caution to the pile: don't mistake the router's diary for the model's mind.