When and How Human Curation Backfires: Preference Alignment under Multi-Model Self-Consuming Loop

In three linesTheoretical study of foundation models trained on synthetic data from other model iterations. Authors show that human curation of one model can degrade alignment of other models through cross-model interactions, unlike isolated settings where it always improves alignment.

Summary generated by Claude — human-verified