Date: June 2, 2026 (Tuesday)
Time: 3:30pm – 5:00pm
Venue: Rm 10.13, Run Run Shaw Tower, Centennial Campus, HKU
Speaker:
Professor Eduardo Barrio, University of Buenos Aires
Registration: Here
Abstract:
This talk investigates the limits of transitive reasoning in Large Language Models (LLMs), specifically examining their handling of inferential patterns that violate classical structural properties. While classical logic assumes the transitivity of entailment, empirical observation suggests that LLMs often default to transitive closure even in contexts where it is logically inappropriate. Drawing upon theoretical motivations for treating metainferences as formal entailments, I introduce a novel benchmark designed to evaluate LLM performance in cases where Meta-Hypothetical Syllogism fails. The methodology contrasts zero-shot and few-shot judgments with chain-of-thought (CoT) prompting to isolate systematic biases toward transitive defaults. I conclude by discussing the implications of these findings for the underlying inferential architecture of LLMs and evaluate the extent to which these models can be aligned with non-classical logical frameworks.
