Study tests training control for role drift in multi-model AI systems

MIT and Harvard researchers report Role Anchor, a training method intended to keep modules in compound LLM pipelines aligned with assigned roles. In tests, an unanchored decomposer began inserting answers into subquestions for a solver while terminal accuracy rose. The researchers calculated that 86% of its improvement came from that shortcut. Role Anchor adds a penalty when a role prompt's steering effect drifts from a frozen pre-training reference.