← News

Agents & MCPSources: arXiv · 19 Nov 2026

3D GameDev Arena benchmark tests whether coding agents can build full 3D games from scratch

A new arXiv paper introduces 108 long-horizon game-development tasks where automated agents plan, code, and playtest browser games, with human preference judged in a public arena.

In plain words

Researchers created a benchmark that asks AI coding agents to build complete 3D browser games from a written description and design images, with no human help. Each game goes through an automated cycle of planning, implementation, testing, and repair. The benchmark then checks both whether the code works and whether people actually enjoy playing the result.

What was announced

The paper's authors position 3D GameDev Arena as a first benchmark for extreme long-horizon game development, spanning 108 from-scratch tasks across 14 genres. Existing benchmarks, they argue, rarely assess player enjoyment, even though quality in games rests on it, and large 3D games tightly couple mechanics, scenes, assets, and interaction logic. In the pipeline, an agent receives only a text specification and UI/UX reference images, then produces a self-contained browser game through a fully automated plan–implement–test–repair loop. Runs average roughly 1,000 LLM calls, and the longest run used 14,407 calls, 2.13 billion tokens, and 608 assets.

Evaluation combines three layers: objective code-level acceptance against the specification, agent-based QA and playtesting of 3D presentation and playability, and blind pairwise human preference using a Bradley-Terry model in a public online arena. The authors report preliminary findings that long-horizon 3D game development remains far from solved, and that code-level correctness alone does not dictate which game players prefer. The abstract does not name the agents tested or report per-model rankings, so those details are not established here.

What it does not show

  • The evidence is the paper's abstract only; full methodology, tested agents, and results are not covered.
  • Findings are described as preliminary by the authors and have not been independently confirmed.
  • The abstract does not report per-agent scores or rankings from the arena.

For security and IT teams

Scope
The benchmark covers from-scratch 3D browser game development from text specifications and reference images, not general repository coding tasks.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.