Reproduce “Scaling Zero-Order Pretraining through Model Sharding” 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

Francois Chaubard, Mykel J. Kochenderfer, Chris Ré

The paper presents SOMA, which trains independently clustered experts with local losses instead of one monolithic zero-order model. It combines experts at inference after routing contexts to their shards. Near 8.44 million parameters and 150 aggregate GPU-hours, two-expert SOMA reaches 1.76 test nats/byte, versus 2.00–2.11 for monolithic SPSA and 2.21 for EGGROLL. However, BPTT performs better at every tested budget.

Summary by alphaXiv.

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