AI Translates Preference Models into Natural Language for Editing.

Statistical learning algorithms are increasingly used to infer human preferences from complex choice data. However, the opacity of these models makes it difficult to understand which factors truly drive decisions and to correct errors. A new method, termed "weights to words," aims to bridge this gap by making AI preference models more interpretable and editable. This method takes a dataset of choice problems and automatically identifies relevant preference dimensions within the domain. Each dimension is then described in natural language and linked to a corresponding vector in the model's internal representation space. This dual representation addresses both the under-determination of factors influencing choices and the opacity of the models. Users can inspect the model's inferences in natural language and make real-time edits to these preference profiles. Experiments across various domains, including moral dilemmas and movie selection, show significant benefits. Regularizing a preference model towards the learned basis increases prediction accuracy on unseen choices, and incorporating user-structured edits further enhances this accuracy. Participants in human-subjects experiments preferred the method's inferred preference profiles and found its predictions more accurate, highlighting its potential for improving trust and utility in AI-driven preference systems.