BREAKING
Arena.ai Adds Factuality Scoring
How Factuality Scoring Works
1
Sample battles
↓
2
Extract claims
↓
3
Search-verify
↓
4
Blend with votes
0
M+
claims labeled
0
%
true claims (Text)
0
%
true claims (Search)
Text Arena Rank Shifts
GPT-5.5 +13
13
Xiaomi +3
3
Meta -3
3
Muse Spark -13
13
Two Signals, Combined
Human Preference
75%
●
Rewards helpful style
●
Misses false claims
Factuality
25%
●
Verifies atomic claims
●
Adjustable 0-100%
Liked Isn't Always Correct
AI NEWS BLITZ
Arena.ai just added a factuality signal that reshuffled its model rankings.