Synthetic Data (concept)edit
Generated data used for training, evaluation, and collaboration.
Synthetic Data refers to generated examples that are used in place of, alongside, or as a proxy for real data. In machine learning, synthetic data can expand coverage, reduce annotation cost, protect privacy, or enable evaluation when real data are scarce. It can also introduce failure modes when generated samples are recursively reused without enough real-data anchoring.
Research contextedit
Synthetic data can be useful for augmentation and other applications, while recursive reuse introduces risks of its own. Qiao's synthetic-data research studies how those risks depend on selection and access to real data.
Connection to Qiao's workedit
Qiao's ICML 2026 work studies synthetic data under selection bias, low-resource verification, and siloed access. The core concern is not merely that data are generated, but that generation is embedded inside a repeated training loop. When each generation learns from biased selections of earlier outputs, the synthetic distribution can drift away from the original, with low-resource communities especially exposed to tail-mode loss. The project connects synthetic data to data selection, model collapse, and collaborative evaluation.