Behind Aggregate Scores: What Systematically Varying Arena-Hard's LLM Judge Reveals

No. 27 of 245 · arXiv · Louise Yorke · published 23 Oct 2026, 04:00 UTC · first published 21 Oct 2026 · 43 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 workshop paper shows that LLM judges used in Arena-Hard can produce stable aggregate scores while individual prompt verdicts change substantially. Rerunning an unchanged GPT-4.1 setup altered 21.2% of prompt outcomes, and some instruction packages changed more.

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

0.71total under rubric v2.0

Significance0.64
Reader interest0.78
Audience breadth0.62
Practical relevance0.76
Substantiated0.84
Technical depth0.87
New information0.86
Counts against: Promotional0.03
Counts against: Overclaim0.24

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	0.69	0.45	0.78	0.64	0.33	+0.25	+0.160
Reader interest	0.68	0.72	0.93	0.78	0.25	+0.20	+0.155
Audience breadth	0.62	0.55	0.68	0.62	0.13	+0.10	+0.062
Practical relevance	0.77	0.70	0.82	0.76	0.12	+0.15	+0.114
Substantiated	0.99	0.75	0.78	0.84	0.24	+0.10	+0.084
Technical depth	0.98	0.70	0.92	0.87	0.28	+0.10	+0.087
New information	0.92	0.70	0.97	0.86	0.27	+0.10	+0.086
Promotional	0.04	0.05	0.01	0.03	0.04	−0.15	−0.005
Overclaim	0.29	0.30	0.12	0.24	0.18	−0.15	−0.036
Sum of contributions							+0.708
Claim-check penalty							0.000
Total, as published							0.708

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, 08:53 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.