Paper Presentation: LLMS Know More Than They Show: On The In-Trinsic Representation Of LLM Hallucinations

​Paper Presentation

​LLMS Know More Than They Show: On The In-Trinsic Representation Of LLM Hallucinations

​- By Hadas Orgad, Ph.D. Candidate at the Technion

​Large Language Models (LLMs) have transformed many fields, from natural language processing to conversational AI, yet they face a critical challenge—generating “hallucinations” or errors, which include factual inaccuracies, biases, and reasoning failures. This session features Hadas Orgad, a leading researcher from Technion, who will present insights from her latest paper, "LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations."

​Orgad’s research reveals that LLMs possess intrinsic mechanisms for encoding truthfulness within their internal states. Don’t miss this session, ideal for AI professionals, data scientists, and researchers exploring advanced privacy techniques in AI.

​Join us to learn about innovative approaches to understanding and mitigating errors in LLMs. We will conclude with a dedicated Q&A session for participants to ask questions, engage with the speaker, and discuss the research findings in depth.

​Meet our Speaker:

​Hadas Orgad

​Hadas Orgad is a PhD candidate at the Technion, advised by Yonatan Belinkov. She specializes in interpretability research in language models and text-to-image models. Her work focuses on making interpretability insights practical and actionable for improving AI systems. Through her research, she tackled challenges in biases, fairness, and model hallucinations, and developed methods for updating and erasing information in models. Hadas’ contributions were recognized by the Apple Scholars in AIML PhD fellowship.

​Download the research paper: Here

​​This session is part of AI Paper-fest 2024 by The SSI Club. For more information and to register for other presentations, visit papers.ssiclub.ai