Tag
This paper studies self-consuming training in a multi-model regime, showing that human curation can backfire and degrade long-term alignment due to cross-model interactions.