Humans and Mice Navigate Mazes Alike. Can AI Beat Them?

Paper | Poster Page

Mice are natural navigators, capable of few-shot learning in complex environments. Here, we explore how their performance in maze navigation can serve as a benchmark for comparing learning across species and artificial agents. We tested human participants on a virtual binary maze game adapted from a prior mouse study and found not only similar performance, but also striking parallels in learning dynamics. Like mice, humans rapidly optimized reward acquisition, exhibited sudden insights about the maze structure, and showed knowledge of the path home from their very first maze incursion. We then used this embodied navigation task to compare AI agents with both species. We showed that two canonical agents — a Deep Q-Learning (DQN) and a Large Language Model (LLM) — were outperformed by the biological learners. These results highlight the potential of naturalistic learning tasks for cross-species comparisons, and expose challenges and opportunities for advancing AI.