Reproduce “Depth Memory” 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
David González-Martínez, Philipp Nazari, T. Konstantin Rusch, Shiwei Liu
Linear Depth Memory (LDM) gives transformer layers a fixed-size associative memory they can query and update, alongside the ordinary residual stream. It uses recursive least squares to fit key–value associations rather than retaining all earlier layer outputs. In matched 1.5B-parameter models trained on 100B tokens, LDM achieved validation perplexity 9.15 versus mHC’s 9.25; evidence covers tested language-modeling setups.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.