Reproduce “From Large Language Models to Efficient System One Models via Adaptive Computation” 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

Minh-Anh Nguyen

NESTONE trains one decision model with shared, progressively deeper exits and a gate that continues computation when an early choice seems unreliable. On seven decision benchmarks, its 12-layer exit scores 64.4% versus 61.5% for a separately trained model. At δ=2, gating averages 20.2 layers, with a 1.68-point macro-average drop; near-parity applies only to the four largest benchmarks.

Summary by alphaXiv.

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