AtlasKV: Augmenting LLMs with Billion-Scale Knowledge Graphs in 20GB VRAM

ICLR 2026 Conference Submission6710 Authors

Keywords: Large language model; Knowledge augmentation; Knowledge graph;

TL;DR: This paper proposes AtlasKV, a scalable, effective, and general way to augment LLMs with billion-scale KGs in less than 20GB GPU VRAM, where KG2KV and HiKVP are introduced to integrate KG triples at scale with sub-linear time and memory complexity.

Abstract: Retrieval-augmented generation (RAG) has shown some success in augmenting large language models (LLMs) with external knowledge. However, as a non-parametric knowledge integration paradigm for LLMs, RAG methods heavily rely on external retrieval modules and the retrieved textual context prior. Especially for very large scale knowledge augmentation, they would introduce substantial inference latency due to expensive searches and much longer relevant context. In this paper, we propose a parametric knowledge integration method, called $\textbf{AtlasKV}$, a scalable, effective, and general way to augment LLMs with billion-scale knowledge graphs (KGs) (e.g. 1B triples) using very little GPU memory cost (e.g. less than 20GB VRAM). In AtlasKV, we introduce KG2KV and HiKVP to integrate KG triples into LLMs at scale with sub-linear time and memory complexity. It maintains strong knowledge grounding and generalization performance using the LLMs' inherent attention mechanism, and requires no external retrievers, long context priors, or retraining when adapting to new knowledge.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 6710