Séminaire ayant lieu dans le cadre du doctorat en informatique cognitive, en collaboration avec le centre de recherche CRIA   

TITRE :  The unreasonable effectiveness of pattern matching

Thomas SERRE

Jeudi le 19 février 2026 à 10h30

Local PK-5115 (Il est possible d'y assister en virtuel en vous inscrivant ici)               

RÉSUMÉ

Large language models (LLMs) can often recover the meaning of texts in which most or all content words have been replaced by nonsense strings—so-called “Jabberwockified” language. In this talk, I argue that this ability reveals something important, and underappreciated, about both artificial and human cognition: the extraordinary power of large-scale pattern matching and constraint satisfaction. Drawing on new demonstrations showing that LLMs can reconstruct the gist—and sometimes surprisingly specific content—of highly degraded texts, I suggest that these systems are not best understood as parrots, databases, or blurry copies of the web. Instead, they exploit deeply learned structural regularities spanning syntax, discourse, and genre. I connect these findings to construction-grammar approaches to language, classic work on relational cognition, and evidence that human reasoning itself is graded, probabilistic, and pattern-based rather than strictly rule-governed. The broader implication is that pattern matching is not an alternative to “real” intelligence, but a central ingredient of it—one that helps clarify both the power and the limits of current language models.

BIOGRAPHIE

Gary LUPYAN is Professor of Psychology at the University of Wisconsin–Madison. His research examines how language shapes perception, thought, and learning, and how linguistic systems adapt to the needs of their users and learners. He has worked extensively on the role of language in human cognition, category learning, conceptual structure, and the evolution of communicative systems, with recent work exploring what large language models reveal about the nature of meaning and generalization.

RÉFÉRENCES 

Lupyan, G., & Agüera y Arcas, B. (2026). The unreasonable effectiveness of pattern matching. arXiv preprint.

Lupyan, G. (2025). Large language models have learned to use language. arXiv preprint arXiv:2512.12447.

Wigner, E. (1960). The Unreasonable Effectiveness of Mathematics in the Natural Sciences. Communications on pure and applied mathematics, 12, 1–14.