Unifying Large Language Models and Knowledge Graphs for Faithful and Interpretable Reasoning by Gholamreza (Reza) Haffari

Venue: SIT001

Abstract: Large Language Models (LLMs) demonstrate strong general reasoning ability, yet still suffer from hallucination, limited faithfulness, and a lack of interpretability—especially in knowledge-intensive or domain-specific settings. Knowledge Graphs (KGs), on the other hand, provide structured, explicit, and verifiable representations of facts, but are incomplete and lack linguistic flexibility. This talk presents recent advances in unifying these two paradigms to achieve trustworthy and interpretable reasoning. In the first part, I will introduce Reasoning on Graphs (RoG) and Graph-Constrained Reasoning (GCR), two frameworks that guide or constrain LLM reasoning using KG structure. RoG enables planning–retrieval–reasoning with faithful relation paths, while GCR enforces KG-valid reasoning during decoding, eliminating hallucinated reasoning paths and improving accuracy and interpretability. The second part of the talk presents GFM-RAG, a graph foundation model trained on 60 diverse KGs with over 14 million triples for efficient, multi-hop retrieval-augmented generation. GFM-RAG achieves state-of-the-art performance across multiple QA benchmarks and generalizes zero-shot to new datasets. Together, these methods highlight a path toward unified, scalable, and reliable KG-LLM reasoning.

Bio:  Gholamreza (Reza) Haffari is a Professor in the Department of Data Science and Artificial Intelligence at Monash University, Australia. He is a former ARC Future Fellow and previously served as Director of the Vision and Language Group. His research sits at the intersection of Natural Language Processing, Deep Learning, and Machine Learning, with funding from ARC, Google Research, Amazon, eBay, Adobe, and other industry partners. Reza also serves as the Chief AI Scientist at Openstream AI.