Reproduce “Joint or Per-Option? Compact Encoders for Closed-Set Decisions from 17 Million to 1 Billion Parameters” 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
Tardelli Ronan Coelho Stekel
The authors introduce Crivo, which returns probabilities over allowed choices without generating text. In a three-seed, 32M-parameter study, per-option scoring beats an approximate compute-matched joint reader by 0.028 chance-corrected skill on held-out pub15, but not from raw initialization. On all 209 System1-Arena tasks, Crivo-68M scores 0.466 versus 0.416 for a 435M zero-shot classifier; its unseen-task advantage is not Holm-significant.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.