Memory-Based Model Editing at Scale

TL;DR

The paper introduces SERAC, a retrieval-augmented model editing method that explicitly stores and reasons over edits, significantly improving the accuracy and robustness of language model updates across multiple tasks.

Contribution

SERAC is a novel semi-parametric approach that enhances model editing by explicit memory and reasoning, outperforming existing methods on diverse language tasks.

Findings

SERAC achieves high performance across question answering, fact-checking, and dialogue generation.

It outperforms existing model editing approaches by a significant margin.

The approach maintains effectiveness after multiple edits.

Abstract

Even the largest neural networks make errors, and once-correct predictions can become invalid as the world changes. Model editors make local updates to the behavior of base (pre-trained) models to inject updated knowledge or correct undesirable behaviors. Existing model editors have shown promise, but also suffer from insufficient expressiveness: they struggle to accurately model an edit's intended scope (examples affected by the edit), leading to inaccurate predictions for test inputs loosely related to the edit, and they often fail altogether after many edits. As a higher-capacity alternative, we propose Semi-Parametric Editing with a Retrieval-Augmented Counterfactual Model (SERAC), which stores edits in an explicit memory and learns to reason over them to modulate the base model's predictions as needed. To enable more rigorous evaluation of model editors, we introduce three challenging language model editing problems based on question answering, fact-checking, and dialogue generation. We find that only SERAC achieves high performance on all three problems, consistently outperforming existing approaches to model editing by a significant margin. Code, data, and additional project information will be made available at https://sites.google.com/view/serac-editing.