AtlasKV: billion-scale knowledge graphs in 20GB VRAM, no retriever or vector DB needed

udmrzn · x · 2026-11-20

Researchers from HKUST and Huawei's Theory Lab propose AtlasKV, a parametric knowledge augmentation method that injects billion-scale knowledge graph triples (e.g. 1B) directly into an LLM's inherent attention mechanism using under 20GB VRAM. Via two new methods (KG2KV and HiKVP) it achieves sub-linear time and memory complexity, with no external retrievers, no long-context priors, and no retraining when adding new knowledge — a potential replacement for the whole RAG pipeline. Code and data are open-sourced.

Original post →

More from Infra

Infra channel →