KG-GROUNDED CONTINUAL ADAPTATION FOR LLMS

 Introduction

Large language models can benefit from structured knowledge graphs (KG) as an external grounding mechanism. Current KG–LLM research shows that integrating KG information can improve interpretability, trustworthiness, and decision support, while also helping mitigate hallucinations in open-ended tasks. A promising research direction is to use KGs as a structured memory or adaptation signal when the underlying knowledge evolves over time.

Objectives

·  Explore how a knowledge graph can support continual adaptation of an LLM.

·  Compare static prompting, retrieval-based grounding, and incremental KG-aware adaptation.

·  Evaluate trustworthiness, consistency, and temporal relevance of generated answers.

Tasks

1.     Literature Review: Continual KG-augmented LLMs, temporal KGs, trustworthiness, grounding.

2.     Task Definition: Choose a QA, explanation, or assistant task where knowledge evolves over time.

3.     System Design: Build an LLM pipeline grounded on KG/TKG/STKG information.

4.     Adaptation Mechanism: Study how KG updates propagate into model behavior via prompting, retrieval, or parameter-efficient tuning.

5.     Evaluation: Measure factuality, temporal consistency, usefulness, and robustness to outdated or conflicting knowledge when using KG-aware continual adaptation compared to baselines.

6.     Optional Extension: Include a user interface for non-expert interaction.

Pre-requisites

Python, NLP/LLMs, basic knowledge of graphs or semantic technologies.

Work

25% Theory, 55% Programming/Experiments, 20% Writing

Contact

Ali Sabzi Khoshraftar (a.sabzikhoshraftar@utwente.nl)