Reproduce “Train Separately, Merge Together: Modular Post-Training with Mixture-of-Experts” 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
Jacob Morrison, Sanjay Adhikesaven, Akshita Bhagia, Matei Zaharia, Noah A. Smith, Sewon Min
BAR develops capability-specific model components separately and combines them into one model, allowing domain updates without retraining the system. It preserves an anchor expert while training domain experts, then combines them with a learned router. In experiments on a 7B model, BAR scores 49.1, versus 47.8 for post-training-only retraining and 50.5 for full retraining; it does not win every task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.