Evaluating World Models with 3D Games
sebnadeau · reddit · 2026-07-15
The author created WorldBuild Bench, using playable 3D games to evaluate how well different models understand "world models."
They argue that static questions are inadequate for measuring a model's grasp of space, time, and causality. Instead, they use three types of 3D gaming tasks:
- Arena combat games
- Physics puzzles
- Racing games
In the initial round, 8 models were tested, generating 24 browser-playable 3D games. The author emphasizes that the evaluation focuses not on traditional scores, but on dimensions closer to real-world experience: spatial consistency, temporal consistency, causal consistency, visual presentation, and completeness.
Methodologically, the project uses blind A/B testing: users play two games generated from the same brief without knowing the model names, comparing them on overall preference, game feel, world design, presentation, and completeness. The author also disclosed generation time, costs, lines of code, and underlying assets, arguing these metrics are more insightful than standard benchmark scores.
More from Models
- Kimi K3.1 Pro tops MathArena open models, closing frontier gap to about 6 months — teortaxesTex · 2026-10-26
- Ruff creator Charlie Marsh: frontier models find infinite bugs in any software — charliermarsh · 2026-10-26
- Dev testing one skill across Opus 5.5 and Sol 6.1 finds instruction following 'all over the map' — lucasmeijer · 2026-10-26
- supermemory unveils learner-1, claiming continual learning for any model — julianweisser · 2026-10-26
- Anthropic says Opus 5.5 cuts cost per task by ~40% vs Opus 5 — taufiqintech · 2026-10-26
- AI search APIs ranked: top 12 within 7 points, yet prices differ nearly 7x — FinanceYF5 · 2026-10-26