Artificial Analysis describes AA-3D, which has each model, working as an agent in an offline sandbox, turn a written brief into a web page. A headless browser and a person check every scene, then panelists compare blind pairs from the same brief. A Bradley-Terry fit converts those votes into Elo ratings.
Part of a thread · earlier post
How AA-3D works: build, check, vote, rank
➤ Each model, as an agent in an offline sandbox, turns a written brief into a web page
➤ A headless browser and a person check every scene
➤ Panelists compare blind pairs from the same brief
➤ A Bradley-Terry fit turns votes into Elo
Source: Artificial Analysis · x.comPublished · added here
