Multimodal LLM Fails at Map Reading, Hallucinates Place Names

GaryMarcus · x · 2026-08-23

Gary Marcus shares a failure case of a multimodal LLM processing maps: the model incorrectly identified the location "Venice" as "Venecky" and generated the sentence "We live in Venecky." This incident reinforces his previous predictions about the persistent problems multimodal LLMs face with maps.

Original post →

More from Fun

Fun channel →