Scaling Laws for Generative Reward Models

No. 130 of 245 · arXiv · Angel Raychev, Nathan Lile, Rafael Rafailov · published 23 Oct 2026, 04:00 UTC · 38 min read

AI & machine learning
Paper

3 judges. Scored by all three judges. Only the day’s top stories get the claim check. How far stories are checked

This paper studies generative reward models used as online preference judges for training creative-writing policies with DPO. It finds that Thinking judges trained with GRPO score higher on static accuracy than Answer-Only judges trained with SFT, yet Answer-Only judges usually produce better policies under a fixed learned Gold evaluator, an…

Read the original at arxiv.org. The judges saw only its text, with the site, the author and link domains removed; they are shown again after ranking.

Score

0.65total under rubric v2.0

Significance0.57
Reader interest0.70
Audience breadth0.54
Practical relevance0.62
Substantiated0.80
Technical depth0.92
New information0.88
Counts against: Promotional0.04
Counts against: Overclaim0.23

Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).

Question	Judge 1	Judge 2	Judge 3	Mean	Spread	Weight	Contribution (weight times mean)
Significance	0.60	0.40	0.72	0.57	0.32	+0.25	+0.143
Reader interest	0.58	0.60	0.91	0.70	0.33	+0.20	+0.139
Audience breadth	0.54	0.45	0.62	0.54	0.17	+0.10	+0.054
Practical relevance	0.63	0.50	0.74	0.62	0.24	+0.15	+0.093
Substantiated	0.94	0.60	0.87	0.80	0.34	+0.10	+0.080
Technical depth	0.98	0.80	0.96	0.92	0.18	+0.10	+0.092
New information	0.86	0.80	0.99	0.88	0.19	+0.10	+0.088
Promotional	0.05	0.05	0.02	0.04	0.03	−0.15	−0.006
Overclaim	0.31	0.30	0.08	0.23	0.23	−0.15	−0.035
Sum of contributions							+0.649
Claim-check penalty							0.000
Total, as published							0.649

Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.

Also covered by

No other coverage was clustered with this story.

Elsewhere

Elsewhere: HN —

Recorded for context only, checked 24 Oct 2026, 08:54 UTC. It played no part in the ranking.

Panel for this edition

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition and listed in its JSON. This one was scored under rubric v2.0; How ranking works describes the current rubric, v2.2.