Bridging LLMs and KGs without Fine-Tuning: Intermediate Probing Meets Subgraph-Aware Entity Descriptions

Traditional knowledge graph completion (KGC) methods rely solely on structural information, struggling with the inherent sparsity of knowledge graphs (KGs). By contrast, Large Language Models (LLMs) encapsulate extensive world knowledge and exhibit powerful context modeling capabilities, making them promising for mitigating the limitations of traditional methods. However, direct fine-tuning of LLMs for KGC, though effective, imposes substantial computational and memory overheads, while utilizing non-fine-tuned LLMs is efficient but yields suboptimal performance. In this work, we propose a novel framework that synergizes the strengths of LLMs with robust knowledge representation to enable effective and efficient KGC. We extract the context-aware hidden states of knowledge triples from the intermediate layers of LLMs, thereby capturing rich semantic and relational nuances. These representations are then utilized to train a data-efficient classifier tailored specifically for KGC tasks. To bridge the semantic gaps between LLMs and KGs, we employ subgraph sampling on KGs to generate model-friendly entity descriptions. We further adopt sliced mutual information (SMI) as a principled metric to quantify the task-specific information encoded in these representations. Extensive experiments on standard benchmarks validate the efficiency and effectiveness of our approach. We achieve a 47\% relative improvement over previous methods based on non-fine-tuned LLMs and, to our knowledge, are the first to achieve classification performance comparable to fine-tuned LLMs while enhancing GPU memory efficiency by $188\times$ and accelerating training and inference by $26.11\times$.