A training-free MoE compression kept about 90 percent of reasoning performance after cutting 60 percent of parameters, against 73 percent for the prior method
AspireAI community · arXiv (Cary, Jin, Huang, Ye, Li, Hua, Liu & Gao)
Mixture-of-Experts models cut compute through sparse activation, but the stored expert parameters remain a memory bottleneck. GS-MoE attacks that with no training at all, compressing along two axes at once: a shared-subspace rank that captures structure common across experts while retaining expert-specific projections, and Nystrom-based neuron compression that reduces the intermediate width inside each expert.
Expert grouping is driven by per-expert reconstruction-error reductions derived from the reconstruction objective, and budgets are allocated across layers with a gradient-free layer-importance score, then distributed as non-uniform ranks and widths by eigenvalues and leverage. On DeepSeek-MoE-16B-Base, at a 60% reduction in total model parameters, GS-MoE retains approximately 90% of the uncompressed model's average reasoning performance, compared with 73% for D2-MoE, while achieving 20% lower out-of-domain perplexity.
The limitation is scope: the headline comparison is a single base model at a single compression rate, and the paper's other results are relative to named baselines rather than absolute. The authors also report the complementary benefit is strongest at high compression rates on fine-grained MoE models, so the saving depends on how much redundancy a given model actually carries.
https://arxiv.org/abs/2610.32009
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.