Reproduce “Understanding Mixture-of-Experts Language Models Beyond Expert Boundaries” 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
Zhenhong Zhou, Mao Zheng, Mingyang Song, Qiankun Li, Kun Wang, Junfeng Fang
The authors compare neuron-level contributions across text domains in six mixture-of-experts models. RAMP estimates each routed neuron's contribution from its activation, routing weight, and output-projection norm. At 6.25% selection, same-domain overlap is 57.77–74.33% across models, versus 15.23–23.68% across domains. This supports domain-associated organization within shared experts, not causal or human-readable neuron meanings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.