Golden Gate Bridge, as Always? Eliciting Prototypical Places From Autoregressive Large Language Models via Category Production

TL;DR

A study investigating whether autoregressive LLMs generate prototypical places for geographic categories.

Problem

Humans perceive categories with graded membership based on prototypicality, but it is unknown whether LLMs internally represent categories in a similar way. Moreover, the lack of diversity in LLM outputs may stem from such graded internal representations, a point that has been overlooked.

Approach

Adapting the cognitive psychology experiment of category production, we prompted 11 autoregressive LLMs to generate prototypical places for 297 geographic categories (e.g., 'city', 'river'). We analyzed the generated category norms to identify graded structures within and across models, and explained the lack of diversity through the lens of prototype theory.

Results & Contribution

We found a strong presence of prototypical places shared across different sampling temperatures, geographic categories, and models. We showed that the lack of geographic diversity in LLM outputs can be explained by prototype theory, providing a new theoretical lens for studying bias in LLMs.