InfuserKI: Enhancing Large Language Models with Knowledge Graphs via Infuser-Guided Knowledge Integration

TL;DR

InfuserKI tackles the challenge of integrating unknown, domain-specific knowledge from knowledge graphs into large language models without eroding existing knowledge. It introduces an adaptive infusion framework that uses parallel knowledge adapters and an Infuser module to selectively fuse new information, keeping the base transformer frozen. The method incorporates knowledge-detection via multiple-choice generation, a knowledge adapter to store new facts, and a relation-classification task to generalize knowledge to unseen scenarios. Empirical results on UMLS and MetaQA show improved reliability, locality, and generality over PEFT and ME baselines, with reduced forgetting and strong performance at scale. The work suggests a practical path for domain-specific knowledge augmentation in LLMs without expensive retraining or catastrophic forgetting, with potential extensions to multi-hop reasoning.

Abstract

Large Language Models (LLMs) have achieved exceptional capabilities in open generation across various domains, yet they encounter difficulties with tasks that require intensive knowledge. To address these challenges, methods for integrating knowledge have been developed, which augment LLMs with domain-specific knowledge graphs through external modules. These approaches, however, face data inefficiency issues as they necessitate the processing of both known and unknown knowledge for fine-tuning. Thus, our research focuses on a novel problem: efficiently integrating unknown knowledge into LLMs without unnecessary overlap of known knowledge. A risk of introducing new knowledge is the potential forgetting of existing knowledge. To mitigate this risk, we propose the innovative {\method} framework. This framework employs transformer internal states to determine when to enrich LLM outputs with additional information, effectively preventing knowledge forgetting. Performance evaluations using the UMLS-2.5k and MetaQA domain knowledge graphs reveal that {\method} not only successfully integrates new knowledge but also outperforms state-of-the-art baselines, reducing knowledge forgetting by 9\% and 6\%, respectively.

Figures (7)

• Figure 1: An illustrative comparison among (a) Vanilla LLM, (b) Fine-Tuned LLM, and (c) our Knowledge-Infused LLM.
• Figure 2: Knowledge Areas in LLMs: Original ($\mathcal{N}_1$+$\mathcal{N}_2$), Post-Fine-Tuning ($\mathcal{N}_1$+$\mathcal{N}_3$), Forgotten ($\mathcal{N}_2$), and Failed Integration ($\mathcal{N}_4$).
• Figure 3: Infuser-Guided Knowledge Integration Framework.
• Figure 4: Infuser-Guided Knowledge Adapters.
• Figure 5: Impact of Adapter Positions on InfuserKI.