Focus shifts in contextual and lexical cue interactions in GPT models

Wonil Chung, Keonwoo Koo · 2026

Read Focus shifts in contextual and lexical cue interactions in GPT models by Wonil Chung, Keonwoo Koo (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/plos-pone-0351924.txt.

Description

Wonil Chung and Keonwoo Koo studied how GPT language models handle focus, the spotlight inside a sentence. Focus tells a listener which piece of news is at issue. The team used cut-off sentences with two possible completions. A question before the sentence set one spotlight. The small word only set another, placed before different items. Word-by-word surprise scores showed which ending each model expected.

The question's spotlight ruled in the first tests. It beat even a competing only. When only stood before the second object, the word-level spotlight grew stronger. It began to bend the expected endings its way. The models weigh the two cues in a graded tug-of-war, not by fixed rules. That matters wherever emphasis changes meaning in a text.

Catalog record

CREATORWonil Chung, Keonwoo Koo
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60665
LICENSECC_BY
RIGHTSCC BY
SOURCEPLOS

Source: https://doi.org/10.1371/journal.pone.0351924

Browse the full library · About the Tower of Babel