Humans see it instantly — LLM doesn’t

This isn’t really about solving a maze — it’s about recognizing that no solution exists. From the exit (top-right), it’s immediately clear the region is completely disconnected. A human spots this at a glance without tracing any paths. The model, however, failed to do that. It didn’t start from the exit, didn’t check connectivity, and instead tried to construct a path — effectively assuming a solution must exist. This feels like a typical failure mode: LLMs aren’t prepared for trick cases where the correct answer is “no solution”, especially when it requires a simple global/topological check rather than step-by-step construction.