Reproduce “M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling” 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

Mayank Mishra, Shawn Tan, Ion Stoica, Joseph E. Gonzalez, Tri Dao

M²RNN combines nonlinear recurrent updates with a matrix-valued state: keys and values write outer products to memory, queries read from it, and an input-dependent gate blends states. Its recurrence remains sequential. On a synthetic permutation task, it maintains at least 99.5% accuracy through length 512; at 7B parameters, its hybrid reaches WikiText perplexity 13.00, versus 13.51 for Gated DeltaNet.

Summary by alphaXiv.

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