Skip to content

Research artifacts and CPU analysis — arXiv:2610.30765

Choose a tag to compare

@RaychevAngel RaychevAngel released this 23 Oct 08:08

Scaling Laws for Generative Reward Models

Angel Raychev · Nathan Lile · Rafael Rafailov

arXiv:2610.30765 · Project page

This release contains the manuscript, numerical research artifacts and
reproducible CPU analysis. It verifies retained tables, regenerates 22 vector
figures, reproduces descriptive size fits and recounts retained comparison
receipts. Linux CI on Python 3.10, 3.11 and 3.12, plus a fresh anonymous clone, have passed
the published commands. No GPU or model credentials are required.

Scope update, 23 October 2026: the custom training implementation and
historical launchers are not included in this public package. An earlier
release included them; the paper's original availability paragraph reflects
that earlier scope. No model weights or raw story datasets are redistributed.
See the repository's dated release-scope note.

The attached manuscript and TeX archive retain the scientific content already
published on the project page. This release makes no change to arXiv.

uv sync --frozen
make verify test figures summary size-fits receipts