TLDR Tech / Practical
Advanced AI Models Struggle With Basic Maze Navigation
An assessment of GPT-6 Astra and Claude Opus 5.5 involved presenting them with 50 interactive 2D mazes to gauge their long-horizon action and causal reasoning abilities. Although these state-of-the-art models navigated mazes without switches more successfully than earlier versions and showed fewer limitations regarding task length, all five models evaluated were unable to complete mazes that included switch-gates. The study observed that Astra and Opus 5.5 encountered difficulties in unique ways, for example, by retracing steps over already visited areas or repeatedly performing unproductive pickup operations.
Read the full story