Reproduce “SLEM: Spectrum-Aware Latent Expert Merging” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Dohyeon Kim, Bedionita Soro, Bruno Andreis, Sung Ju Hwang

SLEM reduces the stored expert set in mixture-of-experts models by learning a shared latent representation of expert weights, selecting similar experts, and interpolating their subnetworks in that space. It uses pretrained weight spectra to set contributions without calibration examples, but requires training an Expert VAE. Across three evaluated models, it leads generation-benchmark averages at both tested reduction rates, though other categories are mixed.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.