Human Baseline on Mazebench: Top 10 Score Average 70%
patience_cave · x · 2026-09-02
In contrast to Fable's low score, top human performers on the Mazebench are crushing it with an average score of 70% across 30-50 hours of playtime. This highlights the gap where humans can easily play these games without tools, while AIs require billions of tokens.
More from Research
- ECCV 2026 hosts CDEL workshop on curated data for efficient learning — y_m_asano · 2026-09-02
- 8-Year Study Reveals Implicit Symbolic Structure in LLMs — shangbinfeng · 2026-09-02
- Mythos 5.1 Prefers High-Stakes Tasks and Outcome Agency — Sauers_ · 2026-09-02
- Authors Are Better Judges of Their Work Than Peer Reviewers — weijie444 · 2026-09-02
- R3: Training Robots to Reason in Natural Language via RL — tw_killian · 2026-09-02
- Measuring model capability gaps through reasoning traces — himanshustwts · 2026-09-02