Reproduce “Latent Lattice Language Models” 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

Shivam Duggal, Zongze Wu, William T. Freeman, Antonio Torralba, Richard Zhang, Eli Shechtman

The authors introduce the Lattice, adding persistent per-layer states that recur across tokens alongside the Transformer’s vertical residual stream. These states condition attention and the MLP without adding token positions or KV-cache entries. After 20B FineWeb-Edu tokens, its 405.6M-parameter model reaches perplexity 14.40 versus 15.41 for a 406.7M baseline at approximately matched forward compute: 0.72 versus 0.71 GFLOP/token.

Summary by alphaXiv.

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