Reproduce “Nexus: Same Pretraining Loss, Better Downstream Generalization via Common Minima” 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

Huanran Chen, Huaqing Zhang, Xiao Li, Yinpeng Dong, Ke Shen, Jun Zhu

The authors introduce Nexus, an optimizer for mixed-domain language-model pretraining that encourages gradients from different batches to align through an auxiliary inner loop. In a 3B-parameter comparison, Nexus reaches validation pretraining loss 1.602 versus AdamW’s 1.606 and GSM8k accuracy 59% versus 44%. Gains are not universal: GPQA accuracy falls, while BBH and MBPP remain unchanged.

Summary by alphaXiv.

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