Learnijoy
ResearchAI ResearchAI Engineering & DevTools

Parametric Knowledge Graphs Show Storage-Retrieval Gap

Martino M. L. Pulici, Cuong Xuan Chu, Evgeny Kharlamov, Volker Tresp· August 27, 2026 View original

Key takeaways

  • LoRA adapters can effectively store factual knowledge from knowledge graphs parametrically.
  • This method offers zero query-time context cost compared to traditional RAG.
  • A significant "storage-retrieval gap" exists, as knowledge is not recoverable by semantic similarity.
  • New query-conditioned composition mechanisms are needed for effective retrieval from parametric memory.

Who benefits

AI/ML PlatformsEnterprise SearchKnowledge ManagementSoftware Development

Summary

This paper explores compiling knowledge graphs into LoRA adapters for parametric memory, finding that while adapters effectively store factual knowledge, retrieving it via semantic similarity or weight-space geometry is ineffective. This highlights a "storage-retrieval gap" and the need for new query-conditioned composition mechanisms.

Traditional graph retrieval-augmented generation (RAG) incurs recurring token costs by placing subgraphs into a model's context window. This research investigates an alternative: compiling a knowledge graph offline into a bank of LoRA adapters, with one adapter per entity, to serve as a parametric knowledge layer. This approach aims to eliminate query-time context costs by injecting weights instead of text. Using the MetaQA dataset, the study found that these subgraph-trained adapters successfully encode context-free factual knowledge, generalizing to unseen questions and significantly outperforming a base model. However, a critical "storage-retrieval gap" was identified: while knowledge is stored, it cannot be effectively retrieved by semantic similarity or weight-space geometry. This means that even semantically similar entities do not contain the answer in their adapters, indicating that knowledge is highly localized. The findings underscore the need for a learned, query-conditioned mechanism to select and compose the correct adapters.

Why it matters

For professionals designing knowledge-intensive AI systems, this research reveals both the potential and a significant challenge of parametric knowledge storage. It suggests that while knowledge can be embedded, effective retrieval requires innovative approaches beyond current similarity-based methods.

How to implement this in your domain

  1. 1Explore the concept of parametric knowledge storage using LoRA adapters for specific knowledge domains.
  2. 2Design experiments to quantify the storage capacity of adapters for factual knowledge.
  3. 3Investigate alternative retrieval mechanisms beyond semantic similarity for parametric knowledge graphs.
  4. 4Consider developing query-conditioned composition mechanisms for combining multiple adapters to answer complex queries.

Original post by Martino M. L. Pulici, Cuong Xuan Chu, Evgeny Kharlamov, Volker Tresp

"arXiv:2608.25489v1 Announce Type: new Abstract: Graph retrieval-augmented generation places retrieved subgraphs into the model's context window at query time, paying a recurring token cost and exposing source data on every call. We study an alternative: compiling a knowledge grap…"

View on X

Originally posted by Martino M. L. Pulici, Cuong Xuan Chu, Evgeny Kharlamov, Volker Tresp on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI Engineering & DevToolsAI Research

Resilient Decentralized Federated Learning for Wireless IoT Networks

This paper introduces QEF-GT-AdamW, a communication-efficient and outage-resilient algorithm for decentralized federated learning over wireless IoT networks. It combines gradient tracking, AdamW optimization, and dual-stream biased quantization with error feedback to improve robustness and convergence under heterogeneous data and unreliable communication.

Nguyen Van Thieu, Ti Ti Nguyen, Ons Aouedi, Vu Nguyen Ha, Symeon ChatzinotasAug 27, 2026
AI Engineering & DevToolsAI Research

FedQoS Predicts QoS Risk for Wireless Access Selection

This paper proposes FedQoS, a federated QoS-risk learning framework that predicts future QoS degradation for reliable access selection in heterogeneous indoor-outdoor wireless environments. It enables access nodes to locally learn from network logs and collaboratively train a global predictor without centralizing user data, significantly reducing QoS failure rates.

Nguyen Van Thieu, Ti Ti Nguyen, Ons Aouedi, Zerihun Huruy, Vu Nguyen Ha, Symeon ChatzinotasAug 27, 2026
AI ResearchAI Engineering & DevTools

DQC Improves Multi-Modal Regression by Resolving Mean-Collapse

This paper introduces Difference-Quotient Clustering (DQC) to address the mean-collapse problem in multi-modal regression, where regressors converge to a mean that misses all true modes. DQC partitions data to minimize intra-cluster output-vs-input discrepancy, assigning conditional labels that significantly improve performance on synthetic benchmarks.

Huang WeiquanAug 27, 2026