A reward-model paper that's been quietly revised for over a year just got its fourth update.
R3, short for Robust Rubric-Agnostic Reward Models, first went up on arXiv in May 2025. This week's revision is the fourth version of that same paper, not a new release. The pitch hasn't changed: a reward-modeling framework that scores language-model outputs against any rubric instead of the single narrow objective most reward models are trained on, and that explains its scores with reasoning instead of returning an unexplained number. The code, training data, and models remain posted as open source on GitHub.
Reward models are the graders behind reinforcement learning from human feedback, and most are built for one job, then treated as a black box when their score does not match a human's judgment. A rubric-agnostic model that shows its reasoning could save teams from retraining a separate scorer every time they change what they are optimizing for. That is a real convenience, not a breakthrough.
This is less a launch than a paper still finding its final form well over a year after it first appeared, worth a note, not a headline.