Reproduce “Constant-Memory Recall: Learned Associations in a Fixed Matrix State” 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
Samuel Larson
The authors evaluate a small DeltaNet variant on synthetic episodes containing 32 newly assigned key–value links. It writes associations into a fixed-size matrix state and answers a single successor query after filler tokens are added. It achieves mean accuracy of 0.9995 across three training seeds. This result measures episode-restricted recall, not natural-language performance or recall amid new interfering associations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.