Self-Consuming Generative Models with Co-Evolving Human Preferences

The paper examines generative models that are trained through self‑consuming loops where users curate preferred samples. It shows that when training relies solely on user‑curated synthetic data, initial biases are amplified, leading the system to converge to a single dominant outcome. Introducing reference data at a sufficient rate changes the dynamics, producing a unique globally attracting equilibrium, and the authors propose an algorithm to select reference distributions and mixing weights to steer the system toward desirable equilibria while reducing data collection costs.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Machine Learning.