Head-to-head empirical comparison with D²MoE and other dynamic expert-precision methods

Determine the head-to-head empirical performance of DPA-MoE relative to published dynamic and mixed expert-precision methods, particularly D²MoE, whose comparison could not be reproduced because no public implementation was available.

Background

DPA-MoE is evaluated primarily against Uniform-kk allocation, with ablations isolating the contributions of router gate scores and offline residual profiles. The paper identifies D²MoE as the closest published method because it also uses nested or prefix-like quantization for expert precision control.

A direct empirical comparison remains unresolved because the authors could not faithfully reproduce the relevant published methods, and D²MoE did not provide a public implementation. Establishing such a comparison would clarify whether DPA-MoE's reported gains extend beyond the uniform-allocation baseline and how its training-free, token-aware allocation compares with methods requiring learned precision policies.

References

This stands in for a head-to-head comparison against published methods, which we could not reproduce faithfully: the closest, D\textsuperscript{2}MoE, has no public implementation.

— DPA-MoE: A Runtime Precision Control Framework for MoE Inference  (2610.21378 - Kim et al., 15 Oct 2026) in Section 5.5, “Which Signal Drives the Gain?”; see also Appendix, Section “Comparison with Dynamic Expert-Precision Methods”