MazeBench Open-Sourced: 3D Open-World Benchmark for Long-Horizon Planning

MazeBench has been officially open-sourced. It is a 3D open-world puzzle benchmark built on a JavaScript engine, designed to evaluate agents' long-horizon planning, visual-spatial reasoning, and continual learning. The authors state that current best agents can only pass the initial few levels, highlighting severe limitations in complex spatial tasks.

Confirmed

• MazeBench is an open-source local browser application that provides an interactive play interface, leaderboard, and save functionality, evaluating coding agents within the same JavaScript engine.
• The benchmark environment contains over 200 rooms and puzzles, with over 100 rooms co-designed by the author's family.
• Evaluation emphasizes out-of-distribution (OOD) difficulty and continual learning: some levels require the agent to control the camera, and there are reusable techniques across levels, requiring cross-level learning.
• To prevent agents from infinite loops wasting compute, the benchmark introduces a novelty score mechanism: agents can run approximately O(M) tokens, but are terminated if the system detects they are looping.
• The project is sponsored by Prime Intellect and guided by mentors.

Why it matters

• Traditional benchmarks are mostly static QA sets, while MazeBench provides a dynamic, highly challenging 3D interactive environment that directly exposes the severe shortcomings of current top agents in complex spatial navigation and long-horizon planning, pointing to new breakthrough directions for future agent research.

2026-07-28 ~ 2026-07-29 · 14 related posts

• [source] MazeBench is a 3D benchmark that leaves today’s agents stuck in the opening rooms — patience_cave · 2026-07-28
• [source] MazeBench opens a 3D puzzle benchmark for coding agents with 200+ rooms — patience_cave · 2026-07-28
• MazeBench launches as an open-source 3D puzzle benchmark for agent evaluation — patience_cave · 2026-07-28
• MazeBench adds novelty-based loop stopping to keep long runs from wasting compute — xeophon · 2026-07-28
• MazeBench turns maze solving into a continual-learning benchmark with camera control — xeophon · 2026-07-28
• MazeBench launches as an open-source 3D spatial reasoning benchmark — patience_cave · 2026-07-28
• [source] Mazebench goes open source with 100+ designed rooms and a blog release — patience_cave · 2026-07-28

7 near-duplicate retellings: teortaxesTex · JasonBotterill · adonis_singh · basedjensen · giffmana · majidmanzarpour · i_dg23