English

Sparse Crosscoders for diffing MoEs and Dense models

Machine Learning 2026-03-09 v1

Abstract

Mixture of Experts (MoE) achieve parameter-efficient scaling through sparse expert routing, yet their internal representations remain poorly understood compared to dense models. We present a systematic comparison of MoE and dense model internals using crosscoders, a variant of sparse autoencoders, that jointly models multiple activation spaces. We train 5-layer dense and MoEs (equal active parameters) on 1B tokens across code, scientific text, and english stories. Using BatchTopK crosscoders with explicitly designated shared features, we achieve 87%\sim 87\% fractional variance explained and uncover concrete differences in feature organization. The MoE learns significantly fewer unique features compared to the dense model. MoE-specific features also exhibit higher activation density than shared features, whereas dense-specific features show lower density. Our analysis reveals that MoEs develop more specialized, focused representations while dense models distribute information across broader, more general-purpose features.

Keywords

Cite

@article{arxiv.2603.05805,
  title  = {Sparse Crosscoders for diffing MoEs and Dense models},
  author = {Marmik Chaudhari and Nishkal Hundia and Idhant Gulati},
  journal= {arXiv preprint arXiv:2603.05805},
  year   = {2026}
}

Comments

5 pages, 3 figures

R2 v1 2026-07-01T11:05:58.753Z