Synthetic Dataedit

Qiao's research on synthetic data examines generated data, recursive training, and model collapse. Related questions include recursive synthetic-data training, data selection, sample selection bias, model collapse, data silos, and Wasserstein geometry.

Introductionedit

The topic treats synthetic data as both a resource and a risk. Generated samples can reduce data-access costs and support privacy-preserving workflows, but recursive use of selected synthetic data can also narrow the training distribution. Evidence for model collapse concerns indiscriminate recursive reuse rather than every use of synthetic data; retaining original data reduced degradation in the cited study.¹ Low-resource verification, biased local selection, and collaborative evaluation shape this tradeoff.

Research contextedit

Recursive selection can amplify bias and erase modes from the training distribution. Low-resource communities are particularly exposed to tail loss when local verifiers mistake rare but valid samples for low-quality generations.

Publicationsedit

Paper	Venue/status
When Sample Selection Bias Precipitates Model Collapse	ICML 2026, 6-11 July 2026, Seoul.

Connection to Qiao's workedit

When Sample Selection Bias Precipitates Model Collapse studies how local selection bias can trigger collapse in low-resource, siloed recursive training, then uses collaborative Wasserstein-style signals to diagnose the problem. This connects synthetic-data reliability to AI and networks because the key difficulty is not only generation quality, but also distributed access to evidence about the data distribution.

See alsoedit

• When Sample Selection Bias Precipitates Model Collapse
• Synthetic Data
• Model Collapse
• Data Silos
• Collaborative Evaluation

Footnotesedit

1. Shumailov et al., "AI models collapse when trained on recursively generated data", Nature 631, 755-759 (2024), is a primary reference for the recursive model-collapse framing and reports reduced degradation when part of the original data is preserved. ↩