Reproduce “How Linear Attention Remembers” 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
Kichang Lee, JaeYeon Park, Songkuk Kim, JeongGil Ko
The authors study how pretrained linear attention models retain and retrieve facts in a shared recurrent state. They intervene on selected fact updates and states in controlled four-choice associative recall. In GDN-340M, targeted donor-answer transfer occurs in 71.9% of cases, while 0.41% of previously correct non-target answers break. These results concern controlled recall, not general long-context reasoning or serving benefits.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.