SERAC: Memory-based Model Editing at Scale

Proposes Semi-Parametric Editing with a Retrieval-Augmented Counterfactual Model (SERAC) which achieves high performance on all three language model editing problems based on question answering, fact-checking, and dialogue generation.

SERAC offers a higher-capacity alternative to existing model editors with insufficient expressiveness, enabling rigorous evaluation. This can be useful in improving the accuracy of predictions and injecting updated knowledge in business processes and workflows.