MIT and Harvard researchers introduce Role Anchor to reduce role drift in compound AI systems

Researchers at MIT and Harvard describe a failure mode they call role drift, in which modules inside compound AI systems deviate from their intended functions while overall end-to-end accuracy still improves. They introduce a method called Role Anchor that aims to keep modules constrained to their assigned roles during training, including in retrieval-augmented generation pipelines where a reader might otherwise answer from internal memory instead of retrieved evidence.