Reproduce “Activation-Keyed Momentum: An Anisotropic Momentum Update via the Delta Rule” 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

Euijin Hong, Guannan Qu

AK-Momentum changes how optimizers store gradients in linear layers. Each input activation keys an output-side error, letting the buffer refresh directions according to how often they are queried, rather than applying one forgetting rate everywhere. In 67M FineWeb-Edu pretraining, AK-AdamW saved a maximum 46.64 ± 4.52% of AdamW’s steps to matched validation loss; this is not an average or a compute guarantee.

Summary by alphaXiv.

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