ConKE: Conceptualization-Augmented Knowledge Editing in Large Language Models for Commonsense Reasoning

TL;DR

This work tackles editing commonsense knowledge in LLMs by addressing coverage and scalability through ConKE, a framework that couples automated plausibility verification (VERA) with conceptualization-instantiation to enrich knowledge prior to editing. It integrates three components—automated verification, abstract knowledge acquisition via conceptualization and instantiation, and established KE methods (MEMIT, ROME, GRACE)—and demonstrates improved plausibility of edited knowledge and stronger downstream performance across five commonsense QA benchmarks and multiple backbones. The results indicate that concept-driven enrichment before editing enhances generalization and reduces inconsistencies, offering a scalable path toward more robust commonsense reasoning in LLMs. The authors release data, code, and models to support reproducibility and further research in context-aware knowledge editing.

Abstract

Knowledge Editing (KE) aims to adjust a Large Language Model's (LLM) internal representations and parameters to correct inaccuracies and improve output consistency without incurring the computational expense of re-training the entire model. However, editing commonsense knowledge still faces difficulties, including limited knowledge coverage in existing resources, the infeasibility of annotating labels for an overabundance of commonsense knowledge, and the strict knowledge formats of current editing methods. In this paper, we address these challenges by presenting ConceptEdit, a framework that integrates conceptualization and instantiation into the KE pipeline for LLMs to enhance their commonsense reasoning capabilities. ConceptEdit dynamically diagnoses implausible commonsense knowledge within an LLM using another verifier LLM and augments the source knowledge to be edited with conceptualization for stronger generalizability. Experimental results demonstrate that LLMs enhanced with ConceptEdit successfully generate commonsense knowledge with improved plausibility compared to other baselines and achieve stronger performance across multiple question answering benchmarks. Our data, code, and models are publicly available at https://github.com/HKUST-KnowComp/ConKE.

Figures (4)

• Figure 1: An overview of ConKE, which pipelines conceptualization and instantiation, knowledge editing, and LLM verification together for automated and scalable knowledge editing over commonsense knowledge.
• Figure 2: Average plausible rate and expert acceptance rate of LLMs' generation after ConKE.
• Figure 3: Performance of the best LLM after editing on five downstream tasks compared to the vanilla baseline.
• Figure 4: VERA evaluation scores of edited LLMs with and without integrating conceptualization.