Paolo Morosi (UAB): ‘The semantic abilities of Large Language Models are bound by logical reasoning but not by pragmatic enrichment: Evidence from conditionals’

Seminari del CLT

Divendres, 10 d’abril de 2026
Hora 15:30
Aula 202, Facultat de Filosofia i Lletres

Abstract:
Humans read between the lines, effortlessly going beyond the literal content of words and structures. For instance, when someone says, “If you mow the lawn, I’ll give you 50$”, it is implied that the speaker will pay only if the lawn is mowed (a so-called perfected reading). Conversely, “If you’re hungry, there’s pizza in the oven” implies that pizza is in the oven regardless of the hearer’s hunger (a biscuit reading). Large Language Models (LLMs) show human-like performance on many cognitive tasks, yet it is unclear whether they can reason like humans, and if not, why exactly this ability is still under development.

In this talk I will present a population-matching experiment that evaluates how 25 LLMs compute perfected and biscuit inferences with conditionals, across Catalan, English, Italian, and Spanish, and tested against an equal number of humans per language.

Our results indicate that humans tend to enrich logical reasoning through pragmatic inference cross-linguistically. Model behavior reveals a more variable pattern. Some LLMs perfectly follow the truth-table semantics of conditionals but they ignore pragmatic inferences, while others adhere to a perfected interpretation across the board, reflecting accurate rule-based processing but not human-like pragmatic reasoning. Overall, LLMs are accurate semantic operators, but they may fail to derive the pragmatic inferences that humans typically compute. Crucially, this variation in LLM performance is not predicted by factors such as open vs. closed status, training orientation, or type of architecture, suggesting that pragmatic reasoning is still an emerging ability in the cognitive toolkit of artificial systems.