Reproduce “Multi-State Linear Attention with Generalized Second-Order Routing” 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
Hee Seung Hwang, Sanghyuk Chun, Xindi Wu, Olga Russakovsky, Zhiwei Deng
MERLIN routes segment readouts in multi-state linear attention by combining query-key alignment means and variances across compressed context segments, using a learned gate to mix separately normalized scores. On LongBench, mid-trained Gated DeltaNet-1.3B scores 21.9 versus Meanpool’s 19.3, a 13.5% relative gain. This is an average across 11 tasks up to 16K tokens, not a win on every task or evidence beyond two tested backbones.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.