Abstract: The rapid advancement of Large Language Models (LLMs) has revolutionized natural language generation, yet their propensity to "hallucinate"—generating non-factual or unfaithful content—remains a critical challenge to their reliable deployment. This talk presents a research journey aimed at building more truthful AI, charting a course from granular detection to proactive mitigation. We begin with a detailed exploration of hallucination detection in the news domain. We present a framework for identifying hallucinations in generated news headlines. This work is extended through a fine-grained, multilingual typology of hallucinations, providing a more nuanced understanding of how models fail across different languages. Building on these diagnostic insights, we then shift to a new training-time hallucination mitigation framework. Specifically, we look at how search-augmented multi-step reinforcement learning can post-train LLMs for improved internal factual consistency. Finally, the talk concludes by outlining some interesting future research directions for building fundamentally more reliable generative systems.
Bio: Jiaming Shen is a senior research scientist at Google DeepMind, specializing in natural language processing, data mining, and machine learning. His research focuses on assisting humans with LLM agents for knowledge acquisition, decision making, and creative thinking. His research has been acknowledged through several awards, including the ACL Outstanding Paper Honorable Mention in 2023, the Yunni & Maxine Pao Memorial Fellowship in 2019, and the Brian Totty Graduate Fellowship in 2016. He earned his Ph.D. in Computer Science from the University of Illinois at Urbana-Champaign in 2021.