Title: Abstract World Models for Learning and Reasoning
Abstract: Consider experimenting to learn how to use a new appliance, webpage, or toy: Within tens of minutes, we can learn how something new works, and use that knowledge to achieve novel goals, make sense of similar devices, and communicate our new knowledge in natural language. How could an AI system similarly acquire, transfer, and communicate its knowledge of how things work? To make progress on this question, we study AI systems which learn abstract world knowledge represented as symbolic programs written in Python, natural language, and/or logic. To test these systems, we assemble new interactive benchmarks focusing on interactive puzzle-like grid worlds, and simulated robotic environments, where the agent must interact with a new environment, and then learn enough about how it works to answer questions about the dynamics or achieve novel goals.
Bio: Kevin Ellis is an Assistant Professor at Cornell University in Computer Science, having previously completed his PhD at MIT in Cognitive Science. His research studies the intersection of program synthesis, AI, and human cognition, and was previously recognized with an NSF CAREER Award, coverage by the New York Times, a selection as one of the best human behavior articles in Nature Communications its year, and a paper award at the ARCPrize contest.
Title: Levin Tree Search: Search-and-Learn with Formal Guarantees via Time Sharing
Abstract: Solving complex combinatorial problems from scratch requires a tight integration between learning and search. This keynote presents the Levin Tree Search (LTS) family of algorithms and analyzes their performance guarantees using the simple yet fundamental concept of time-sharing. We show how allocating computational budgets based on policy and heuristic quality can be used to provide strict bounds on the number of node visits before finding a solution. We then describe how these algorithms operate within a search-and-learn bootstrap paradigm to solve combinatorial problems without prior knowledge, using only a training set of unsolved instances of varied difficulty. We introduce in particular the √LTS (root-LTS) algorithm, which extends the time-sharing perspective to all nodes in a search tree. By distributing the computational budget among multiple concurrent searches rooted at different nodes, √LTS implicitly decomposes the original problem into smaller subtasks. We present formal guarantees showing that this rerooting mechanism provides exponential speedups over standard LTS.
Bio: Laurent Orseau is a Research Scientist at Google DeepMind, a position he has held since 2014. From 2007 to 2014, he was a maître de conférences at AgroParisTech. He earned his PhD in Artificial Intelligence from INSA Rennes / Université de Rennes II in 2007. His research focuses on online learning, algorithmic safety, and policy-guided search, earning recognition including an IJCAI 2023 Distinguished Paper Award.
Title: The Complexity Gap: Understanding AI Behavior in Realistic Settings
Abstract: Modern AI systems achieve impressive performance on benchmarks composed of isolated, fully specified tasks. Real-world knowledge work, however, is rarely so clean. Users communicate goals through evolving conversations, leave important details implicit, revisit prior decisions, and combine tasks of varying complexity into long-horizon workflows. As complexity accumulates through under-specification, changing intent, long-range dependencies, and iterative transformations of content, AI systems can exhibit subtle failures that are difficult for both users and developers to anticipate, diagnose, and correct. In this talk, I will discuss recent efforts to move beyond traditional benchmarks and study AI behavior in realistic knowledge-work settings. Drawing on work spanning complex reasoning tasks, multi-turn conversations, and long-horizon workflows, I will present empirical and theoretical results that reveal recurring failure modes as complexity increases. Together, these findings provide a foundation for understanding how and why AI behavior changes in realistic settings, and for developing evaluation paradigms that better reflect the challenges of real-world knowledge work.
Bio: Jennifer Neville is a Partner Research Manager at Microsoft Research Redmond, where she leads the AI Interaction and Learning research group, and the Samuel Conte Chair Professor of Computer Science and Statistics at Purdue University. Her research focuses on understanding how machine learning and AI systems behave in realistic settings, from relational and networked data to modern AI assistants and agents. She is particularly interested in how mismatches between modeling assumptions and deployment environments shape system behavior. Jennifer has authored more than 150 publications with over 12,000 citations across machine learning and artificial intelligence. Her honors include an ICLR 2026 Best Paper Award, NSF CAREER Award, and IEEE's “10 to Watch in AI.” She served as Program Chair of AAAI 2023 and has held leadership roles at NeurIPS, ICML, and IJCAI.
Title: Are AI's key challenges solved?
Abstract: Generative Pre-trained Transformer models (GPT) are a form of Large-Language-Model (LLM) which has attracted wide-scale interest based on general open-ended query-answering. In this talk I will look at remaining challenges for the field of Artificial Intelligence. In particular, I will argue that although widely available LLM systems partially address Alan Turing's "Imitation Game" [1950 paper in the Journal Mind], now called the "Turing Test" they fall short of the "Super Criticality Challenge" which Turing provided within the same paper. While Turing's paper initiated modern discussions on these topics, it was John McCarthy [1956] who named the field "Artificial Intelligence". In a 2006 keynote speech at the Inductive Logic Programming conference (Santiago de Compostella), McCarthy introduced a "Discovery Challenge" which is closely related to Turing's Supercriticality. I will exemplify a range of human discoveries in Science, Engineering and Mathematics, and argue that neither Turing's nor McCarthy's challenges are addressed by existing LLM techniques. By contrast, to enable progress on Discovery Systems, further work is required on development of methods for identifying and explaining rare phenomena. Some initial work and directions for further studies will be described.
Bio: Professor Stephen Muggleton FREng (SM) is founder of the field of Inductive Logic Programming to which he has made contributions in theory, implementations and applications. SM has over 230 publications, and has a Google Scholar h-index of 77. He has been Executive Editor of the Machine Intelligence workshop series since 1992 and Editor-in-Chief of the series since 2000. In particular he acted as Programme Chair of the Machine Intelligence 20 and 21 workshop on Human-Like Computing (Cumberland Lodge, 23-25 October 2016; 30th June-3rd July 2019), co-chair of the 2022 Hooke meeting on Cognitive Artificial Intelligence at the Royal Society, Programme Chair of the International Joint Conference on Learning and Reasoning (IJCLR 2022 and IJCLR 2025) and co-editor of the book "Human-Like Machine Learning" (OUP, 2021). SM founded the Machine Learning group at the University of Oxford in 1993 and went on to become Oxford University Reader in Machine Learning and later Professor of Machine Learning at the University of York in 1997 and then EPSRC/BBSRC Professor of Computational Bioinformatics at Imperial College London in 2001. His work has received widespread international recognition as evidenced by his election to the fellowship of a number of learned bodies: FAAAI (2002), FBCS (2008), FIET (2008), FREng (2010), FRSB (2011) and FECCAI (2014). In 2016 SM acted as a witness to the UK Government’s Select Committee on Artificial Intelligence and Robotics.