Reproduce “A Nurse is Blue and Elephant is Rugby: Cross Domain Alignment in Large Language Models Reveal Human-like Patterns” on OpenResearch

About this paper

Asaf Yehudai, Taelin Karidi, Gabriel Stanovsky, Ariel Goldstein, Omri Abend

Large Language Models from the Hebrew University of Jerusalem team demonstrate human-like conceptualization and reasoning patterns when performing a cross-domain mapping task adapted from cognitive science. The models not only align with human choices but also provide explanations reflecting similar cognitive dimensions.

Key insights

• LLMs exhibit conceptual reasoning and abstract alignment that closely mirrors human cognitive patterns, extending beyond superficial text generation.
• Models rely on similar underlying cognitive dimensions, such as abstract alignment, perceptual similarity, and common mediators, when forming non-literal cross-domain conceptual links.
• The ability of LLMs to generate sensible and human-aligned explanations supports their potential as cognitive models and offers insights into their internal representations, advancing explainable AI.

Summary by alphaXiv.