# KARMA: Leveraging Multi-Agent LLMs for Automated Knowledge Graph Enrichment

**Authors:** Yuxing Lu, Jinzhuo Wang  
**arXiv:** [2502.06472](https://arxiv.org/abs/2502.06472)  
**Date:** 2025-02-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2502.06472v1/

## Key points

- Maintaining comprehensive and up-to-date knowledge graphs is crucial for optimal performance
- Manual curation struggles to keep pace with the rapid expansion of scientific literature
- Introduction of a groundbreaking framework called [Framework Name]
- Innovative approach leverages AI to automate KG enrichment by analyzing unstructured text in a structured manner
- Framework operates through nine collaborative agents specializing in tasks like entity discovery, relation extraction, schema alignment, and conflict resolution
- Agents work iteratively to parse documents, verify extracted knowledge, and integrate it into existing graph structures while adhering to domain-specific schema requirements
- Experiments conducted on 1,200 PubMed articles across three domains showcased impressive results:
- Identified up to 38,230 new entities with an LLM-verified correctness rate of 83.1%
- Reduced conflict edges by 18.6% through multi-layer assessments

## Summary

In the realm of , maintaining comprehensive and up-to-date  is crucial for optimal performance. The manual curation process often struggles to keep pace with the rapid expansion of scientific literature. To address this challenge, a groundbreaking framework called  has been introduced. This innovative approach leverages  to automate KG enrichment by analyzing unstructured text in a structured manner. The  operates through nine collaborative agents, each specializing in different tasks such as entity discovery, relation extraction, schema alignment, and conflict resolution. These agents work together iteratively to parse documents, verify extracted knowledge, and seamlessly integrate it into existing graph structures while adhering to domain-specific schema requirements. To validate the effectiveness of , experiments were conducted on 1,200 PubMed articles spanning three distinct domains. The results were impressive, showcasing the framework's ability to identify up to 38,230 new entities with an impressive LLM-verified correctness rate of 83.1%. Moreover,  demonstrated its capability to reduce conflict edges by 18.6% through multi-layer assessments.