Mechanistic Circuit-Based Knowledge Editing in Large Language Models

TL;DR

This paper introduces MCircKE, a circuit-based framework for precise knowledge editing in large language models, effectively improving multi-hop reasoning by targeting causal circuits responsible for specific facts.

Contribution

The paper presents a novel circuit-based approach that identifies and surgically updates causal reasoning circuits to enhance knowledge editing in LLMs.

Findings

MCircKE effectively bridges the Reasoning Gap in knowledge editing.

The method improves multi-hop reasoning accuracy on the MQuAKE-3K benchmark.

Targeted circuit updates lead to more precise and reliable knowledge modifications.

Abstract

Deploying Large Language Models (LLMs) in real-world dynamic environments raises the challenge of updating their pre-trained knowledge. While existing knowledge editing methods can reliably patch isolated facts, they frequently suffer from a "Reasoning Gap", where the model recalls the edited fact but fails to utilize it in multi-step reasoning chains. To bridge this gap, we introduce MCircKE (\underline{M}echanistic \underline{Circ}uit-based \underline{K}nowledge \underline{E}diting), a novel framework that enables a precise "map-and-adapt" editing procedure. MCircKE first identifies the causal circuits responsible for a specific reasoning task, capturing both the storage of the fact and the routing of its logical consequences. It then surgically update parameters exclusively within this mapped circuit. Extensive experiments on the MQuAKE-3K benchmark demonstrate the effectiveness of the proposed method for multi-hop reasoning in knowledge editing.