Results 61 - 70 of 24600
Language model hallucinations arise not only from gaps in knowledge, but also from failures in how models retrieve, select, and integrate information. In this talk, I will examine these failures from two complementary perspectives. First, I will present a mechanistic account of non-factual hallucinations, identifying distinct failures in knowledge enrichment and answer extraction within the model’s factual recall process. I will then extend this analysis to long-context language models, examining how parametric knowledge interacts, and sometimes conflicts, with information provided in context. The results show that stronger contextual retrieval does not necessarily lead to more reliable knowledge use and may even interfere with parametric recall. Together, these studies characterize hallucination as a failure of knowledge coordination and highlight implications for model interpretability, evaluation, mitigation, and design.
Interaction with AI systems is now happening at massive scale. People engage in multi-round exchanges with AI to guide real-world decisions, and increasingly, AI agents act on our behalf, drawing on humans and tools for support. Yet despite the prevalence of these interactions, we still lack fundamental principles that should govern them: what makes an interaction trustworthy, and how can we ensure it reliably improves the resulting decisions?
The first part of the talk takes a human-centric view of collaboration. The human chooses whether to engage with the AI in the first place and remains accountable for the final decision, so trustworthy collaboration must be built around them. From this view, I identify two fundamental levers governing the interaction. AI should preserve the judgments the human would have made correctly on their own, while providing meaningful help where the human is likely to err. This perspective changes the role of uncertainty quantification. Rather than describing the AI’s uncertainty in isolation, uncertainty becomes a mechanism for governing the effect of AI on the human’s decision. I will show how these principles lead to algorithms for both one-shot predictions and multi-round interactions, with distribution-free guarantees that hold for any human paired with any AI model, without requiring a model of human behavior.
The second part turns to agentic AI, where the traditional decision-support paradigm is reversed: the agent acts on our behalf, while humans and tools provide support. This shift requires rethinking uncertainty as a joint property of the agent and its available support, capturing whether that support can materially improve the agent’s decision. I will show how this perspective unifies diverse support settings within a common framework and provides a principled foundation for designing reliable agent-support interactions.
Across both settings, the talk develops a common language for trustworthy interaction and algorithms that translate these principles into guarantees across different humans, AI models, and forms of collaboration.
Abstract not available.
Register for the lecture here.
Expander graphs are sparse graphs for which, whenever you split the graph into two parts, the number of edges going between the parts is proportional to the size of the smaller part. There are extremely simple, strongly explicit expander families whose analysis requires only a tiny bit of group theory. This fact is perhaps not so widely known. In this lecture, Ryan O’Donnell will show some of these constructions/analyses, and also indicate how they can be used to get even stronger kinds of expanders, such as “super-expanders,” and “non-sofic property (T) groups.”
Ryan O’Donnell is a Professor of Computer Science at Carnegie Mellon University. He received his BSc from the University of Toronto in 1999, and his PhD in applied mathematics from MIT in 2003, where he was advised by Madhu Sudan. His research interests include complexity theory, quantum computation, and spectral graph theory. He is the author of the book Analysis of Boolean Functions.
Refreshments will be served at 3 p.m., before the event.
The Richard M. Karp Distinguished Lectures were created in Fall 2019 to celebrate the role of Simons Institute Founding Director Dick Karp in establishing the field of theoretical computer science, formulating its central problems, and contributing stunning results in the areas of computational complexity and algorithms. Formerly known as the Simons Institute Open Lectures, the series features visionary leaders in the field of theoretical computer science and is geared toward a broad scientific audience.
The lecture recording URL will be emailed to registered participants. This URL can be used for access to the livestream and recorded lecture. Lecture recordings will be publicly available on SimonsTV about five days following each presentation unless otherwise noted.
The Simons Institute regularly captures photos and video of activity around the Institute for use in publications and promotional materials.
If you require special accommodation, please contact our access coordinator at simonsevents@berkeley.edu with as much advance notice as possible.
Generative models are increasingly used in high-stakes applications such as clinical decision support, where a confident wrong answer is worse than an abstention. For a widespread adoption in such settings, models also need to know when they don't know and communicate this reliably to the user. In this talk I will present a principled and practical framework for uncertainty quantification. I will first introduce a general bias-variance decomposition that holds for any proper score that allows sample-specific uncertainty to be estimated directly from the loss function. I will then show how kernel scores extend this decomposition to generative models, yielding an uncertainty measure that can be computed from a finite number of outputs alone, both for discrete and for continuous domains. Building on this, I will introduce a spectral decomposition based on the von Neumann entropy that separates uncertainty caused by an ambiguous prompt from uncertainty caused by a lack of knowledge and show how these sampling-based estimates can be distilled into calibrated uncertainty estimates by post-training approaches.
No abstract available.
Detecting hallucinations in large language models is a critical open problem with significant implications for safety and reliability. While existing hallucination detection methods achieve strong performance in question-answering tasks, they remain less effective on tasks requiring reasoning. In this work, we revisit hallucination detection through the lens of out-of-distribution (OOD) detection, a well-studied problem in areas like computer vision. Treating next-token prediction in language models as a classification task allows us to apply OOD techniques, provided appropriate modifications are made to account for the structural differences in large language models. We show that OOD-based approaches yield training-free, single-sample-based detectors, achieving strong accuracy in hallucination detection for reasoning tasks. Overall, our work suggests that reframing hallucination detection as OOD detection provides a promising and scalable pathway toward language model safety.
No abstract available.