Abstract: Hallucinations in LLMs severely limit their applicability in critical domains, confining them to shallow use cases, where hallucinating a fact incorrectly is not critical. Hallucinations remain endemic to general-purpose LLMs, even with advanced techniques like RAG and instruction-finetuning. It turns out that hallucinations are a fundamental technical problem with a concrete solution that changes the LLM objective, called Memory Tuning.
In this talk, you'll learn the conceptual framework behind this breakthrough research that reduces hallucinations by up to 95% on enterprise use cases for the Fortune 500. Dive into the technical implementation details, including the innovative Mixture of Memory Experts (MoME) model that makes Memory Tuning computationally feasible at scale. Discover how this technology paves the way for more reliable, trustworthy LLM applications across diverse, high-stakes domains.
Bio: Dr. Sharon Zhou is the cofounder and CEO of Lamini. As a former Stanford faculty member, she led a research group in generative AI and published award-winning papers in generative AI. Sharon teaches some of the most popular courses on Coursera, including Finetuning LLMs, reaching nearly half a million professionals. She received her PhD in AI from Stanford, advised by Dr. Andrew Ng. Before her PhD, she was an AI product manager at Google. She received her bachelor's degree from Harvard in computer science and Classics. Finally, Sharon has served as an AI advisor in Washington D.C. and has been featured in MIT Technology Review’s 35 Under 35 list.



