Artificial Analysis Released AA-3D Benchmark for Models
The new evaluation tool uses human preference scoring to rank AI models on generating interactive 3D web scenes.
Updated on Oct. 13, 2026 in Artificial Intelligence
Live Poll
Do you trust human preference voting more than automated scoring for evaluating AI model performance?
Artificial Analysis has announced the AA-3D benchmark, a new system designed to evaluate how AI models generate functional three.js code for interactive 3D web environments. The project uses blind pairwise comparisons to rank model performance based on human preference.
Why it matters
Blind pairwise voting provides a dynamic way to assess model quality that prevents developers from optimizing AI training against fixed, static automated scoring metrics. This approach shifts evaluation toward nuanced human judgment in complex software generation tasks.
The AA-3D system ranks models using an Elo-style rating system, a method for calculating relative skill levels in zero-sum games. It evaluates model output within an offline sandbox to ensure the generated three.js code executes correctly.
The players
Artificial Analysis
An organization focused on creating benchmarks and evaluation systems to rank the performance of generative AI models.
The details
The benchmark tasks AI models with creating interactive three.js — a JavaScript library used to create and display animated 3D computer graphics in web browsers — experiences from text-based prompts. Human judges review these outputs in side-by-side, blind tests, selecting the preferred version without knowing the identity of the generating model. This mechanism avoids the limitations of existing research tools like T3Bench or Blender AI Arena, which primarily evaluate static 3D assets.
Timeline
30 September 2026: The AA-Video-T2V v2.0 benchmark launched.
12 October 2026: Artificial Analysis announced the AA-3D benchmark.
The Tech Race
The AA-3D benchmark departs from the static asset evaluation scope defined by the T3Bench project. It pushes the field toward assessing how models handle live, executable code in web-based 3D environments.
Developers and researchers can use the AA-3D rankings to identify which models currently excel at generating functional 3D web code. This resource provides a practical baseline for selecting models for web-based design and interactive interface workflows.
The takeaway
The industry is increasingly turning to human-in-the-loop benchmarks to measure the functional utility of generative code models. Watch for the initial ranking results on the Artificial Analysis platform to see which models set the standard for 3D web interaction.
Further reading
Explore deeper into how researchers evaluate generative systems in our Artificial Intelligence section.
Source note: This article includes information reported by Crypto Briefing.
Live Poll
Do you trust human preference voting more than automated scoring for evaluating AI model performance?
