Reproduce “Enabling Dynamic Computation in Looped LMs” 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
Aayush Mishra, Arnau Padrés Masdemont, Victor Conchello Vendrell, Jordi Ros Giralt, Arash Behboodi, Fabio Valerio Massoli
Ragged KV lets looped language models stop computing tokens early while allowing later tokens to reuse each token’s last available key/value cache entry at deeper reads. The authors report up to 30% lower computation and cache memory with full-depth performance in tested settings. On BBH, Ouro-1.4B reaches 67.5% pass@1 at average decode depth 3.05, versus 69.0% at fixed depth 4.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.