ThinkEval: Practical Evaluation of Knowledge Leakage in LLM Editing using Thought-based Knowledge Graphs

TL;DR

This paper introduces ThinkEval, a framework for systematically measuring indirect knowledge leakage in large language model editing, using knowledge graphs and a new benchmark dataset to evaluate multiple editing techniques.

Contribution

We propose ThinkEval, a novel framework that quantifies indirect knowledge leakage in LLM editing, supported by the KnowGIC benchmark dataset for complex knowledge transformation analysis.

Findings

Existing editing techniques struggle to suppress indirect knowledge leakage.

All evaluated techniques compromise contextual integrity to some extent.

Knowledge graphs effectively analyze causal structures before and after editing.

Abstract

Robust model-editing techniques are essential for deploying large language models (LLMs) in practical applications, as they enable cost-effective ways to deal with challenges such as privacy breaches, bias mitigation and misinformation spread. For example, an LLM-based healthcare assistance may need to update out-dated or incorrect knowledge to prevent harmful recommendations. However, many editing techniques focus on isolated facts, which critically fail to prevent indirect knowledge leakage -- the unintended reconstruction of edited-out information through persistent causal links and contextual relationships. To assist users in selecting the right editing technique, we develop and present ThinkEval, a framework to systematically quantify indirect knowledge leakage and ripple effects in model-editing. ThinkEval builds and employs specialized knowledge graphs to analyze the causal structure of facts before and after editing. To support this approach, we present KnowGIC, a benchmark dataset comprising multi-step reasoning paths that precisely measure these complex knowledge transformation effects. We evaluate five editing techniques: AlphaEdit, RECT, ROME, MEMIT, and PRUNE across multiple LLMs. Our results show that these techniques struggle to balance indirect fact suppression with the preservation of related knowledge, compromising the contextual integrity of a model's knowledge. Our dataset is available at: https://github.com/manitbaser/KnowGIC.