HKUST and Huawei publish AtlasKV paper for LLaMA 3.1
Researchers at HKUST and Huawei's Theory Lab introduced AtlasKV in an ICLR 2026 paper, demonstrating that injecting knowledge graph triples into the attention layers of a frozen LLaMA 3.1 8B model can store one billion triples in under 20GB of VRAM. The method uses KG2KV for data pipelines and HiKVP hierarchical pruning to achieve sub-linear scaling. However, critics note that document retrieval remains necessary, accuracy has not been tested at the one-billion scale, and the VRAM figure excludes massive offline embedding costs. Open article →