Update 2024-08-29-style-control.md

_posts/2024-08-29-style-control.md

@@ -18,6 +18,13 @@ authors:
     url: "https://infwinston.github.io/"
 ---
 
+### Updated Notes (2025/05/15):
+- Starting from May 15, 2025, Style Control becomes the default view on [lmarena.ai/leaderboard](https://lmarena.ai/leaderboard).
+- For consistency, we change the anchor model to be the current #1 model on the non-style-control leaderboard (specifically, `gemini-2.5-pro-preview-05-06` to 1446). This does not change the interpretation of the scores because the Bradley-Terry coefficients are translation invariant. We perform this translation to preserve consistency between the style-control and non-style-control leaderboards.
+- Moving forward, the non-style-control leaderboard will be referred to as the "raw leaderboard", and the default will be the style-control leaderboard. We will no longer explicitly mention style control.
+
+---
+
 Why is GPT-4o-mini so good? Why does Claude rank so low, when anecdotal experience suggests otherwise?
 
 We have answers for you. We controlled for the effect of length and markdown, and indeed, _the ranking changed_. This is just a first step towards our larger goal of disentangling **substance** and **style** in Chatbot Arena leaderboard.