Synthetic data

Self-improving data pipelines where models generate candidate outputs, then validate them against criteria (correctness, safety, diversity) before accepting them as training data. This closed-loop approach produces higher-quality data than one-shot generation and enables models to bootstrap their own improvement without human annotation.

Creating high-quality synthetic training data using generative models. Developing pipelines that produce diverse, balanced datasets without the privacy concerns or biases of web-scraped data, while ensuring the generated data is genuinely useful for model training.

Understanding and preventing the phenomenon where models trained on synthetic data from previous model generations progressively degrade. Researching the feedback loops that cause model collapse and developing mitigation strategies.

Developing automated methods to assess, filter, and curate training data. Research on quality metrics, deduplication, bias detection, and data mixing strategies that optimize model performance per training dollar.

Using synthetic data generation as a lever for AI safety. Creating targeted safety training examples, red-teaming datasets, and alignment data that help models learn to refuse harmful requests while remaining helpful.