Work

Expert Collapse Is Not Inherited from Weights:\\Depth-Localized Re-Collapse across Three Mixture-of-Experts Architectures,\\and the Dimensional Artifact of the Weight--Activation Gap

Year: 2026
Type: preprint
Abstract: Kim (2026) reports a ``weight--activation disconnect'' in Mixture-of-Experts models: regularizing expert weights toward orthogonality leaves activation-space overlap unchanged, and activation MSO sits roughly $1{,}000\times$ above weight MSO. The number is quoted as evidence that weight-space geometry carries no information about functional specialization. We show that this ratio is an artifact of dimension. MSO is a mean of squared cosines, so for random vectors its expectation is $1/d$---and the two spaces being compared have $d\approx10^{6}$--$10^{8}$ (expert weight matrices) versus $d\approx10^{2}$--$10^{3}$ (residual-stream activations). Kim's $1{,}050\times$ ratio is the same order as what two unrelated spaces of those dimensionalities produce ($2{,}048\times$ for his model). Normalized by the random floor, his model's ratio is $0.51\times$, and the activation space is in fact \emph{ less } aligned than the weight space. Re-measuring five models spanning $130$M to $47$B parameters with a single metric, we find that the disconnect does not generalize. Floor-normalized, the ratio of activation to weight overlap spans six orders of magnitude across the five models, so it is a property of each model rather than of MoE. And three of them sit at the weight-space independence floor ($R_w=1.1$--$1.2$), so their weight geometry carries no ranking at all, while their activation overlap spans a factor of $30$: weight space stops discriminating before activation space does. Our main result is depth-resolved, and it replicates across architectures. In all three $1$B+ models we measure---Mixtral ($8$ experts, top-2), OLMoE ($64$ experts, top-8) and DeepSeek-MoE ($64{+}2$, top-6)---expert outputs are statistically independent through the middle of the network and then re-align sharply in the last one to three layers: swings of $16.7\times$, $336\times$ and $65\times$ against a flat weight profile. The magnitude of the final collapse varies by a factor of $22$ across the three; its location does not. Over the same layers the weight geometry is either flat (Mixtral, $1.09\times$ after consensus removal) or pinned at the floor (OLMoE, DeepSeek, $R_w\approx1.2$). Collapse is generated in the forward pass near the output, not inherited from the weights of the layers in which it appears. Finally, we show that the natural way to test this---subtracting the cross-expert consensus---has its own null distribution, $1/(E-1)^2$, which is $5\times10^{4}$ times larger than the independence floor when $E=64$. Without it, centering manufactures overlap that is not there.
Author: 庆君 张
Language: English
Open Access status: green
OpenAlex ID: W7235735388
Cites: 10
Cited by:
Related to: 10