BabyAI

AI world models suffer from instruction leakage, researchers find

Researchers have identified a critical issue in compact world models that use language goals to ground spatial relations, such as "put the red block left of the blue block." They found that these models often exhibit "i…