T-maze

The T-Maze is a synthetic benchmark used to evaluate the ability of models to handle long-horizon decision-making tasks in partially observable environments, featuring corridors that can extend up to one million steps.

Papers using T-maze (2)

ELMUR: External Layer Memory With Update/rewrite For Long-horizon RL Problems2025Why Evolve When You Can Adapt? Post-evolution Adaptation Of Genetic Memory For On-the-fly Control2025