KGARevion: An AI Agent for Knowledge-Intensive Biomedical QA

TL;DR

KGARevion presents a knowledge-graph grounded AI agent for biomedical QA that jointly leverages LLMs and structured KGs through a four-action loop: Generate, Review, Revise, and Answer. It learns KG-aware triplet validity via TransE embeddings aligned to relation descriptions and LoRA fine-tuning, enabling rigorous verification of generated knowledge. Across multi-choice and open-ended tasks, including newly curated MedDDx and AfriMed-QA datasets, KGARevion outperforms baselines and shows robust performance under variations in question structure and KG source. The approach demonstrates strong zero-shot generalization in underrepresented contexts and offers a versatile, modular framework for knowledge-intensive medical reasoning with potential clinical applicability.

Abstract

Biomedical reasoning integrates structured, codified knowledge with tacit, experience-driven insights. Depending on the context, quantity, and nature of available evidence, researchers and clinicians use diverse strategies, including rule-based, prototype-based, and case-based reasoning. Effective medical AI models must handle this complexity while ensuring reliability and adaptability. We introduce KGARevion, a knowledge graph-based agent that answers knowledge-intensive questions. Upon receiving a query, KGARevion generates relevant triplets by leveraging the latent knowledge embedded in a large language model. It then verifies these triplets against a grounded knowledge graph, filtering out errors and retaining only accurate, contextually relevant information for the final answer. This multi-step process strengthens reasoning, adapts to different models of medical inference, and outperforms retrieval-augmented generation-based approaches that lack effective verification mechanisms. Evaluations on medical QA benchmarks show that KGARevion improves accuracy by over 5.2% over 15 models in handling complex medical queries. To further assess its effectiveness, we curated three new medical QA datasets with varying levels of semantic complexity, where KGARevion improved accuracy by 10.4%. The agent integrates with different LLMs and biomedical knowledge graphs for broad applicability across knowledge-intensive tasks. We evaluated KGARevion on AfriMed-QA, a newly introduced dataset focused on African healthcare, demonstrating its strong zero-shot generalization to underrepresented medical contexts.

Figures (8)

• Figure 1: a) Performance of existing LLMs on three new datasets (MedDDx-Basic, MedDDx-Intermediate, MedDDx-Expert) introduced in this paper with questions of varying difficulty. b) Sample questions from the new datasets.
• Figure 2: a) Overview of KGARevion agent. b) Overview of fine-tuning in the Review action.
• Figure 3: The accuracy of KGARevion and pure LLMs with the medical concepts increase under a) multi-choice reasoning setting and b) open-ended reasoning setting.
• Figure 4: The results of ablation studies across all datasets under two settings.
• Figure 5: a) Performance of KGARevion with different backbone LLMs across all datasets. b) Performance of KGARevion with different KGs used in the fine-tuning stage of the Review action.