Minimization of Boolean Complexity in In-Context Concept Learning

Leroy Z. Wang, R. Thomas McCoy, Shane Steinert-Threlkeld · 2024

Read Minimization of Boolean Complexity in In-Context Concept Learning by Leroy Z. Wang, R. Thomas McCoy, Shane Steinert-Threlkeld (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-02823.txt.

Description

Wang, McCoy, and Steinert-Threlkeld tested a simplicity bias in language models in this paper from twenty twenty-four. People learn simple concepts more easily than tangled ones. The authors asked whether large language models show the same tilt when learning from examples in the prompt. They built concept tasks with known logical structure.

Each concept could be scored by its Boolean complexity, a count of the logical steps it needs. Model success tracked that score closely across tasks. Concepts with low complexity were learned well, while tangled ones tripped the models. The pattern mirrors findings in human concept learning. The authors suggest that in-context learning shares a bias for simple rules with human learners.

Catalog record

CREATORLeroy Z. Wang, R. Thomas McCoy, Shane Steinert-Threlkeld
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE18 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-02823

Browse the full library · About the Tower of Babel