In-Context Bias Propagation in LLM-Based Tabular Data Generation
The study shows that statistical biases present in examples given to a Large Language Model (LLM) systematically distort synthetic tabular data, causing global statistical shifts and compromising downstream fairness. Researchers modeled this propagation, found that malicious injections can target specific subgroups, and tested mitigation strategies. The core finding is that in-context bias is an inherent vulnerability in LLM data generation pipelines.