Language Models Can Resolve Reference Compositionally, But It’s Not Their Native Strength: The Case of the Personal Relation Task

Bart Evelo Meaghan Fowlie Denis Paperno · 2026

Read Language Models Can Resolve Reference Compositionally, But It’s Not Their Native Strength: The Case of the Personal Relation Task by Bart Evelo Meaghan Fowlie Denis Paperno (2026) 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-2605-31480.txt.

Description

Evelo, Fowlie, and Paperno test whether language models truly grasp reference in a compositional way. They use a task about people and their relations. One part asks what a phrase like Amber's parent's friend points to in the world. The other asks for its sense as a nested formula. People do better on the first part, and models on the second. The authors read this split as a sign that missing real-world grounding in model training blocks human-like understanding.

Catalog record

CREATORBart Evelo Meaghan Fowlie Denis Paperno
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40449
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.31480

Browse the full library · About the Tower of Babel