Deficient executive control in transformer attention

Suketu Chandrakant Patel, Hongbin Wang, Jin Fan, Derek Abbott · 2026

Read Deficient executive control in transformer attention by Suketu Chandrakant Patel, Hongbin Wang, Jin Fan, Derek Abbott (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/pmc-13227603.txt.

Description

Suketu Chandrakant Patel and colleagues probe a 2026 weakness in large language models using the Stroop task, psychology's classic test of mental control. People asked to name the ink colour of a mismatched colour word slow down and err, yet they manage. Transformer models, the authors argue, have attention but no executive system for resolving such conflict, and they tested whether that shows.

On short lists the models displayed a human-like conflict effect, erring more on mismatched items. As lists lengthened, mismatched performance collapsed toward total failure while matched items stayed excellent and simple word reading stayed near perfect. Control did not scale up with rising interference. The authors suggest that adding executive-style control may be necessary for more general intelligence.

Catalog record

CREATORSuketu Chandrakant Patel, Hongbin Wang, Jin Fan, Derek Abbott
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE56 KB
VERSION1.0
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC13227603

Browse the full library · About the Tower of Babel