News

Large language models struggle to solve research-level math questions. It takes a human to assess just how poorly they perform.

Have reports of AI replacing mathematicians been greatly exaggerated? Artificial intelligence has attained an impressive series of feats — solving problems from the International Math Olympiad, conducting encyclopedic surveys of academic literature, and even finding solutions to some longstanding research questions. Yet these systems largely remain unable to match top experts in the conceptual frontiers of research math.

When machine learning systems bridge from prediction to intervention — such as in statistical profiling of job seekers — seemingly minor modeling decisions can have profound consequences for who ultimately receives support. In her talk during the workshop on Bridging Prediction and Intervention Problems in Social Systems, Frauke Kreuter (LMU Munich and University of Maryland) examined how different choices in the data science pipeline affect not just predictive accuracy, but the actual composition of individuals flagged for intervention.

In many areas of machine learning, theory and practice have undergone a dramatic divergence; there is extremely little theory to guide our understanding of much of modern AI. Some of what’s likely called for is revolutionary new theory. In this Richard M. Karp Distinguished Lecture, however, Katrina Ligett (Hebrew University) explored a more conservative idea: that we sometimes approach the theory of learning in a way that leaves money on the table.

Greetings from Berkeley, where our program on Federated and Collaborative Learning is in full swing. In January, we hosted our winter Scientific Advisory Board meeting, preceded by a Theory Day with talks by some of our board members. We also had two groups of Circles participants here for a week of collaboration. And this was in addition to a workshop, a boot camp, a program reunion, and a Richard M. Karp Distinguished Lecture.

The successes of generative AI and large language models involve both powerful observable behavior and deep internal representations of the world that they construct for their own uses. How do these internal representations work, and to what extent are they similar to or different from the representations of the world that we build as humans? In this talk, Jon Kleinberg explores these questions through the lens of generative AI, drawing on examples from game-playing, geographic navigation, and other complex tasks.

In this episode of our Polylogues web series, Simons Institute Founding Associate Director Alistair Sinclair interviews newly appointed Institute Director Venkatesan Guruswami. Their wide-ranging conversation touches on the Institute’s mission and strategy, prospects for the field in the years to come, and engagement with the global research community as well as the broader public.

Happy New Year from Berkeley, where the magnolias are already in bud and we have just welcomed the participants in our Spring 2026 research program on Federated and Collaborative Learning. In addition to the periodic workshops associated with the program, we also have upcoming workshops on various aspects at the nexus of theoretical computer science and machine learning, ranging from the deployment of ML models in social systems, healthcare, and deep learning theory to the impact of techniques developed in learning theory on the theory of computing.

Large language models (LLMs) gain their encyclopedic knowledge and conversational tact by learning from an entire internet’s worth of human-generated text. But learning from language alone has shown diminishing returns. While LLMs have proved themselves to be masters of producing fluent language, their capabilities in other cognitive skills, like logic and reasoning, have lagged behind. At the Simons Institute workshop on LLMs, Cognitive Science, Linguistics, and Neuroscience last year, neuroscientists, linguists, and computer scientists came together to explore why this is the case — and how a different model, the human brain, could point the way forward.

Fast matrix multiplication is a central goal in algorithms research. The goal is to find the smallest real value omega such that n by n matrices can be multiplied in n{omega + o(1)} time in the worst case. The current best bound is omega < 2.37134. In this Richard M. Karp Distinguished Lecture in the Complexity and Linear Algebra program, Virginia Vassilevska Williams examines progress on matrix multiplication algorithms over the decades and offers some intuition about where the research area may be headed.