Reproduce “RouteRelay: Event-Triggered Cross-Layer Route Reuse for Efficient Dynamic Sparse Attention” 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

Bin Li, Sisi Liu, Chenyang Hu, Chaoyang Zhang, Wei Li, Hui Song

RouteRelay maintains sparse-attention route indices across Transformer layers, checking selected chunks against near-miss and random sentinel candidates before selectively rebuilding unstable rows. Each layer still computes attention with its own queries, keys, and values. In synthetic cross-layer drift of 0.03, it achieved more than 99.99% top-k route recall while rerouting 55.4% of rows, while the measured CPU pipeline was slower than full routing.

Summary by alphaXiv.

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