Developmental Trajectories of Situation Modeling and Mentalizing in Transformer Language Models

Pamela D. Rivière, Cameron Jones, Sean Trott · 2026

Read Developmental Trajectories of Situation Modeling and Mentalizing in Transformer Language Models by Pamela D. Rivière, Cameron Jones, Sean Trott (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-2606-28524.txt.

Description

Pamela D. Rivière, Cameron Jones, and Sean Trott watch language models grow up. Models can pass false belief tasks, the classic test of understanding that someone can believe something untrue, but doubters question whether the success means real mentalizing. The team traces reasoning across training stages in two open model families, much as developmental psychologists track children.

Above-chance false belief performance needed both size and enough training, arrived late in pretraining, and improved most from post-training in the most diagnostic implicit condition. Yet the skill stayed fragile: verbs like thinks inflated false belief attributions even when beliefs were true. A plainer ability, situation modeling, reporting the facts of a scene, emerged earlier and stronger, but models still confused one character's knowledge with another's. Capability arrives in a sensible order, with coherence lagging behind.

Catalog record

CREATORPamela D. Rivière, Cameron Jones, Sean Trott
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.28524

Browse the full library · About the Tower of Babel