Research map / Large language models / Retrieval and knowledge
AtlasKV: Augmenting LLMs with Billion-Scale Knowledge Graphs in 20GB VRAM
Haoyu Huang, Hong Ting Tsang, Jiaxin Bai, Xi Peng, Gong Zhang, Yangqiu Song
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train on KG2KV-generated Q-K-V data instead of KBLaM synthetic data | KBLaM-style knowledge injection · ATLAS-CC-QKV (10^3 triples) | ACC@1: 10.9 -> 61.8 | Improved |
| use KG2KV instead of synthetic Q-K-V generation | training data · data construction | diversity ratio: 0.003% -> 7.864% | Improved |
| use KG2KV instead of synthetic Q-K-V generation | training data · data construction | avg. token cost: 349.9 -> 165.7 | Improved |
| add HiKVP pruning at inference | AtlasKV · Enron (10^3 triples) | ACC@5: 80.0 -> 50.9 | Hurt |
| add HiKVP pruning at inference | AtlasKV · ATLAS-CC-QKV (10^4 triples) | ACC@1: 61.8 -> 40.0 | Hurt |
| add HiKVP pruning at inference | AtlasKV · GPU memory, KG size up to 1B | VRAM | Improved |
| add event entities to training keys/values (ablation) | AtlasKV w/o HiKVP · ATLAS-Pes2o-QKV (10^4 triples) | ACC@1: 9.1 -> 47.3 | Improved |
| add named entities to training keys/values (ablation) | AtlasKV w/o HiKVP · ATLAS-Pes2o-QKV (10^4 triples) | ACC@1: 3.6 -> 47.3 | Improved |
| add named entities to training keys/values (ablation) | AtlasKV w/o HiKVP · Enron (10^4 triples) | ACC@1: 1.8 -> 27.3 | Improved |
| add relation rewriting to training data (ablation) | AtlasKV w/o HiKVP · ATLAS-Pes2o-QKV (10^3 triples) | ACC@1: 55.8 -> 72.7 | Improved |
| add relation rewriting to test data (ablation) | AtlasKV w/o HiKVP · ATLAS-Pes2o-QKV (10^4 triples) | ACC@1: 26.9 -> 47.3 | Improved |
| injection frequency K=3 vs K=1 or 10 (ablation) | AtlasKV · ATLAS-Pes2o-QKV | ACC | Improved |
Closest papers by meaning
- ReMindRAG: Low-Cost LLM-Guided Knowledge Graph Traversal for Efficient RAG (NeurIPS 2025 · similarity 0.90)
- A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation (AAAI 2025 · similarity 0.90)
- RetrievalAttention: Accelerating Long-Context LLM Inference via Vector Retrieval (NeurIPS 2025 · similarity 0.90)
- Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation (ICLR 2025 · similarity 0.90)
- Inference Scaling for Long-Context Retrieval Augmented Generation (ICLR 2025 · similarity 0.89)
- Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
