2025 Symposium

13th Annual Marshall M. Weinberg Symposium

The symposium will take place in the Michigan League Ballroom on Friday, March 21, 2025.

Three big claims have dominated the last half century of thinking in cognitive science about language. First, language relies on a system of abstract syntactic rules and principles. Second, these rules and principles are too complex to be learned by children without an innate linguistic faculty. Third, developing competence in language requires social interaction grounded in the real world. The last few years have seen the advent of large language models, neural networks that learn to produce fluent language “from scratch”, trained on data from the internet. Do large language models force us to reconsider any of these big claims, and if so, how should we revise our understanding of language? The 2025 Weinberg Symposium features three world-leading experts who will discuss and debate these questions.

Schedule

1:00 - 1:10 PM | Opening Remarks

1:10 - 2:00 PM | Jordan Kodner

2:05 - 2:55 PM | Richard Futrell

3:00 - 3:50 PM | Ellie Pavlick

4:00-5:00 | Panel Discussion

5:00-6:00 | Poster Session and Reception

Learn More About the Poster Session

2025 Featured Speakers

Our featured speakers are renowned experts in their respective disciplines, offering diverse perspectives in cognitive science.

Please click on their names below to read more about them.

Richard Futrell

Richard Futrell is an Associate Professor in Language Science at the University of California, Irvine.

Prediction and Locality in Language and Language Models

Abstract: I argue that human language is shaped by constraints on memory in online language comprehension and production. One way the bottleneck shows up as a preference for locality in the order of elements. Using cross-linguistic corpora of 55+ languages, I show evidence for dependency locality, a pressure for syntactically related words in sentences to be close to each other. I show how a more sophisticated model of language processing, based on incremental probabilistic prediction under resource constraints, yields a generalization of dependency locality called information locality, which I show correctly predicts adjective order across languages. Next, I formulate a general information-theoretic measure of the complexity of sequential prediction, and show cross-linguistic corpus evidence that phonological forms and morphological paradigms are structured in a way that minimizes this complexity. Finally, I present evidence that modern large language models also have a bias towards information locality, and that this may partially explain their successes in learning human language.

Jordan Kodner

Jordan Kodner Assistant Professor in the Stony Brook University Department of Linguistics and an affiliate of the Institute for Advanced Computational Science, Department of Computer Science, and AI Institute. His primary research revolves around computational approaches to child language acquisition and their broader implications. In particular, algorithmic models of grammar acquisition, how those processes drive language variation and change, what insights they provide for low-resource NLP, and how we can evaluate and draw conclusions from computational models at the intersection of (low-resource) NLP and cognitive science. In 2020, Jordan received his PhD from the University of Pennsylvania Department of Linguistics, where he worked with Charles Yang and Mitch Marcus. Jordan received a master's degree from the University of Pennsylvania Department of Computer and Information Science in 2018. Before graduate school, he was a member of the Speech, Language, and Multimedia group at Raytheon BBN Technologies where he worked on defense and medical-related projects.

LLMs and linguistics: Use them for what they are, and don't try to make them what they aren't

Large language models are the most powerful tools we have for uncovering statistical distributions in large quantities of linguistic data. Leveraging this information, they perform on par with or superior to humans on a range of tasks traditionally argued to require a human-like knowledge of language. Nevertheless, LLMs cannot serve as satisfactory models of human language cognition. Rather than revealing how humans learn, process, and represent language, LLMs have shown us that there are sometimes multiple ways to succeed at a task. In this talk, I will discuss weaknesses in popular evaluation methods which inflate the apparent performance of LLMs on tests of linguistic knowledge, both through unintended confounds and cognitively superficial NLP-like evaluation methodologies. I ask whether a test without these problems could even demonstrate that an LLM is an appropriate model of language cognition in the first place. Finally, I propose how LLMs may contribute to the science of language when we use them for what they are rather than trying to force them to be what they aren't.

Ellie Pavlick

Ellie Pavlick is an Associate Professor of Computer Science, Cognitive Science, and Linguistics at Brown University, and a Research Scientist at Google Deepmind. She leads the Language Understanding and Representation (LUNAR) Lab, which seeks to understand how language "works" and to build computational models which can understand language the way that humans do. Her lab's projects focus on language broadly construed, and often includes the study of capacities more general than language, including conceptual representations, reasoning, learning, and generalization. They are interested in understanding how humans achieve these things, how computational models (especially large language models and similar types of "black box" AI systems) achieve these things, and what insights can be gained from comparing the two. We often collaborate with researchers outside of computer science, including cognitive science, neuroscience, and philosophy.

Not-Your-Mother's-Connectionism: LLMs as Cognitive Models

Abstract: Recent advances in AI have led to large neural network models which exhibit human-like behavior across a range of language and reasoning tasks. This (re-)opens important theoretical questions about the nature of the structure that is required to support such behaviors, leading to debates reminiscent of long-running arguments that pit neural network models against explicitly structured symbolic models of the mind. In this talk, I will describe a series of experiments which highlight the ways in which LLMs today appear importantly different from the connectionist systems that inspired these debates originally. I will argue for a more nuanced stance which does not assume neural networks to be diametrically opposed to traditional models of the mind, but still acknowledges the potential of LLMs to teach us something fundamentally new about the structures that govern language and cognition in humans.

Join Us Virtually

Whether you're joining us from around the globe or tuning in from next door, join us Friday, March 21st on our YouTube livestream. Don't miss a moment of the 13th Annual Marshall M. Weinberg Symposium.

Search: {{$root.lsaSearchQuery.q}}, Page {{$root.page}}

for