# Locating and Editing Factual Knowledge in GPT

**Authors:** Kevin Meng, David Bau, Alex Andonian, Yonatan Belinkov  
**arXiv:** [2202.05262](https://arxiv.org/abs/2202.05262)  
**Date:** 2022-02-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2202.05262v1/

## Key points

- Study titled "Locating and Editing Factual Knowledge in GPT" by Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov
- Introduction of causal intervention technique to identify neuron activations influencing factual predictions in autoregressive transformer language models
- Discovery of two sets of neurons representing understanding an abstract fact and articulating a specific word in large GPT-style models
- Development of ROME approach for modifying facts stored in model weights
- Creation of CounterFact dataset with over twenty thousand counterfactual examples for evaluating knowledge editing capabilities
- Validation of differentiation between "saying" and "knowing" neurons using CounterFact dataset
- Demonstration that ROME outperforms other methods in knowledge editing proficiency
- Access to interactive demo notebook, complete code implementation, and CounterFact dataset on https://rome.baulab.info/
- Research offers valuable insights into enhancing factual knowledge manipulation within language models

## Summary

In their study titled "Locating and Editing Factual Knowledge in GPT," authors Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov delve into the mechanisms behind factual knowledge recall in autoregressive transformer language models. They introduce a causal intervention technique to identify neuron activations that can influence a model's factual predictions. Through this method, they discover two distinct sets of neurons within large GPT-style models, which they theorize represent understanding an abstract fact and articulating a specific word. This revelation leads to the development of ROME, a groundbreaking approach for modifying facts stored in model weights. To assess the effectiveness of ROME, the researchers compile CounterFact, a comprehensive dataset containing over twenty thousand counterfactual examples. This dataset enables them to conduct precise evaluations of knowledge editing capabilities. By utilizing CounterFact, they validate the differentiation between "saying" and "knowing" neurons and demonstrate that ROME outperforms other existing methods in terms of knowledge editing proficiency. The authors provide access to an interactive demo notebook, complete code implementation, and the CounterFact dataset on their website https://rome.baulab.info/. With 21 pages and 21 figures, this research offers valuable insights into enhancing factual knowledge manipulation within language models.