Factual Accuracy in LLM-Powered Conversational Assistants: Challenges and Innovations

Day 2 | 09:45 – 10:10 | Main Hall

Photo. Portrait of Jens Lehmann.

Prof. Dr. Jens Lehmann

Amazon AGI

Abstract

This keynote talk highlights the critical challenge of factual accuracy in Large Language Model (LLM)-powered conversational assistants. While LLMs excel at generating fluent and natural conversations, they often struggle with maintaining factual correctness. This talk will explore the limitations of current systems, particularly focusing on how retrieval-augmented generation (RAG) using external APIs can both mitigate and introduce new challenges in achieving accurate outputs. We will introduce HumanIQ, a novel approach that combines structured and unstructured knowledge sources to mimic human-like reasoning, and highlight additional emerging strategies aimed at enhancing factual accuracy. Attendees will gain a deeper understanding of why factuality remains an open problem in LLM-based systems and what promising solutions are on the horizon. 

Prof. Dr. Jens Lehmann

Prof. Dr. Jens Lehmann is a Principal Scientist at Amazon where he works at the AGI (Artificial General Intelligence) organisation on advancing large language models, conversational AI and knowledge graphs. He also holds an honorary professorship at TU Dresden, was selected as a fellow of ELLIS and is a member of  InfAI. His academic activities at TU Dresden and InfAI support the Smart Data Analytics research group. Previously, he was jointly appointed full professor at the University of Bonn and Fraunhofer IAIS leading approximately 40 researchers. In this role, he was a lead scientist at Fraunhofer IAIS and coordinated the Dresden branch of the institute. His research interests involve knowledge graphs, machine learning as well as question answering & dialogue systems. His main research goal is to investigate and build generally intelligent systems by combining knowledge- and data-driven approaches. Prof. Lehmann authored more than 200 articles in international journals and conferences winning 15 best paper awards and obtaining more than 25000 citations and h-index 60+. He is a supporter and contributor to community research projects, including DBpedia, DL-Learner and LinkedGeoData. Previously, he led the AKSW research group and completed his PhD with „summa cum laude“ at the University of Leipzig with visits to the University of Oxford.