OpenNovelty.org

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

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
Large language model; Knowledge augmentation; Knowledge graph;
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 AtlasKV\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.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper proposes AtlasKV, a parametric method for integrating billion-scale knowledge graphs directly into LLM parameters via key-value mappings. It occupies the 'Direct KG-to-Parameter Encoding' leaf within the 'Parametric Knowledge Integration into LLMs' branch. Notably, this leaf contains only the original paper itself—no sibling papers were identified in the taxonomy. This suggests the specific approach of converting KG triples to attention-compatible key-value structures at billion-scale represents a relatively sparse research direction within the broader parametric integration landscape.

The taxonomy reveals neighboring approaches in adjacent branches. 'Pre-trained Knowledge Graph Embeddings for LLMs' explores unified representations learned during pretraining, while 'KG Verbalization for Corpus Augmentation' converts triples to natural language text. 'Multi-Encoder Fusion' architectures combine graph neural networks with language models using separate encoders. AtlasKV diverges by avoiding external retrievers, text conversion, or separate graph encoders, instead leveraging the LLM's native attention mechanism. The taxonomy's scope notes explicitly exclude retrieval-based and text-mediated methods from this parametric branch, positioning AtlasKV as pursuing tighter integration than hybrid alternatives.

Among twenty-four candidates examined across three contributions, none were identified as clearly refuting the work. The AtlasKV framework examined ten candidates with zero refutable matches; KG2KV pipeline examined five with none refutable; HiKVP algorithm examined nine with none refutable. This suggests that within the limited search scope, the specific combination of billion-scale direct encoding, sub-linear complexity guarantees, and hierarchical pruning mechanisms appears relatively unexplored. However, the search examined only top-K semantic matches plus citations, not an exhaustive survey of parametric knowledge integration literature.

Given the sparse taxonomy leaf and absence of refuting candidates among twenty-four examined papers, the work appears to occupy a distinct position within parametric integration approaches. The limited search scope means potentially relevant work in adjacent areas—such as efficient attention mechanisms or knowledge distillation—may not have been captured. The analysis reflects what was found in targeted semantic search, not a comprehensive field review.

Taxonomy

Core-task Taxonomy Papers
6
3
Claimed Contributions
24
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: Augmenting large language models with billion-scale knowledge graphs. The field organizes around several complementary strategies for integrating structured knowledge into LLMs. Parametric Knowledge Integration into LLMs explores methods that encode graph facts directly into model parameters, often through continued pretraining or specialized embedding techniques. Text-Mediated Knowledge Integration converts structured triples into natural language passages that LLMs can consume more naturally, exemplified by approaches like Synthetic Corpus Generation[2]. Hybrid Knowledge Integration Architectures combine parametric and retrieval-based mechanisms, as seen in systems like Hybrid Neural Ensemble[5], to balance memorization with dynamic access. Knowledge Graph Construction and Schema Induction focuses on building and refining the graphs themselves, with works such as AutoSchemaKG[1] addressing schema-level challenges at scale. Enterprise Knowledge Graph Applications targets domain-specific deployments, including e-commerce scenarios explored by PKGM Ecommerce[6], where billion-scale graphs must support real-world query loads. A central tension across these branches involves the trade-off between embedding knowledge within model weights versus retrieving it on demand. Parametric approaches promise faster inference and tighter integration but face scalability limits and update challenges, while text-mediated and hybrid methods offer flexibility at the cost of additional retrieval overhead. Within the parametric branch, AtlasKV[0] pursues direct KG-to-parameter encoding, aiming to compress vast relational structures into the model itself. This contrasts with hybrid strategies like Hybrid Neural Ensemble[5], which maintain separate retrieval modules, and with text-mediated techniques such as Synthetic Corpus Generation[2], which rely on verbalized triples. Ongoing questions include how to maintain factual consistency, as highlighted by LLMs Knowledge Graphs Factuality[4], and how to efficiently update billion-scale embeddings when graphs evolve. AtlasKV[0] sits squarely in the parametric camp, emphasizing direct encoding over retrieval augmentation.

Claimed Contributions

AtlasKV framework for billion-scale KG augmentation

AtlasKV is a parametric knowledge integration framework that enables LLMs to incorporate billion-scale knowledge graphs with minimal GPU memory requirements. It maintains strong knowledge grounding and generalization without requiring external retrievers, long context priors, or retraining when adapting to new knowledge.

10 retrieved papers
KG2KV pipeline for converting KG triples to Q-K-V data

KG2KV is a pipeline that naturally transforms knowledge graph triples into high-quality query-key-value data by leveraging the structural similarity between KG triples and self-attention Q-K-V vectors. This design enhances generalization performance by ensuring diverse enquiry attributes from massive KG relations.

5 retrieved papers
HiKVP algorithm for hierarchical key-value pruning

HiKVP is a hierarchical pruning algorithm that organizes knowledge keys into a three-layer structure and progressively selects relevant key-value pairs. It achieves sub-linear time and memory complexity, enabling scalable integration of billion-scale KGs while preserving high accuracy.

9 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Within the taxonomy built over the current TopK core-task papers, the original paper is assigned to a leaf with no direct siblings and no cousin branches under the same grandparent topic. In this retrieved landscape, it appears structurally isolated, which is one partial signal of novelty, but still constrained by search coverage and taxonomy granularity.

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

AtlasKV framework for billion-scale KG augmentation

AtlasKV is a parametric knowledge integration framework that enables LLMs to incorporate billion-scale knowledge graphs with minimal GPU memory requirements. It maintains strong knowledge grounding and generalization without requiring external retrievers, long context priors, or retraining when adapting to new knowledge.

Contribution

KG2KV pipeline for converting KG triples to Q-K-V data

KG2KV is a pipeline that naturally transforms knowledge graph triples into high-quality query-key-value data by leveraging the structural similarity between KG triples and self-attention Q-K-V vectors. This design enhances generalization performance by ensuring diverse enquiry attributes from massive KG relations.

Contribution

HiKVP algorithm for hierarchical key-value pruning

HiKVP is a hierarchical pruning algorithm that organizes knowledge keys into a three-layer structure and progressively selects relevant key-value pairs. It achieves sub-linear time and memory complexity, enabling scalable integration of billion-scale KGs while preserving high accuracy.

Table of Contents