Interpretabilityedit
Understanding model behavior and the influence of training data.
Interpretability refers to methods that help people understand why a model behaves the way it does. Qiao's related research focuses on data influence, error diagnosis, and explanations that support trustworthiness decisions.
Research contextedit
Interpretability is a supporting topic for Trustworthy AI and Data Centric ML. A model can be accurate but still difficult to audit. If a researcher can explain which examples, groups, or synthetic-data processes caused a behavior, then the next action can be data selection, unlearning, correction, or collaborative evaluation. Interpretability therefore links explanation to intervention.
Connection to Qiao's workedit
Qiao uses influence functions and unlearning to study how data changes affect model behavior. Hessian-Free Online Certified Unlearning and Beyond Binary Erasure both rely on understanding how data changes affect model parameters or predictions. The synthetic-data line also needs interpretability in a broader sense: when model collapse occurs, the research asks what process caused the degeneration and how distributed parties can detect it. The NeurIPS 2026 paper Illusory Pattern Perception Drives Spurious Inference in Large Language Models uses sparse autoencoders to analyze internal features associated with illusory pattern perception and unsupported relational inference.