When Sample Selection Bias Precipitates Model Collapse

In three linesRecursive training on synthetic data risks model collapse: data selection based on fragmented local references removes globally relevant tail modes. Authors theoretically prove siloed selection accelerates collapse and propose Wasserstein proxy references across silos without sharing raw data.

Summary generated by Claude — human-verified