MazeBench: Open-Source 3D Benchmark for Long-Horizon Agents

MazeBench is a fully open-source, browser-based 3D puzzle world for evaluating long-horizon planning and visuospatial reasoning in agents; without Python tools, agents solve only about 1% of levels, and any model with any harness can compete.

2026-08-21 ~ 2026-08-21 · 3 related posts

• MazeBench benchmark tests long-horizon agents: 0% score without Python tools — JFPuget · 2026-08-21
• MazeBench is fully open source: any model with any harness can enter — xeophon · 2026-08-21
• MazeBench: build persistent 3D puzzle worlds in a browser, then eval models in the same engine — xeophon · 2026-08-21