Grid-World Representations in Transformers Reflect Predictive Geometry

Sasha Brenner, Thomas R. Knösche, Nico Scherf · 2026

Read Grid-World Representations in Transformers Reflect Predictive Geometry by Sasha Brenner, Thomas R. Knösche, Nico Scherf (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-2603-16689.txt.

Description

Brenner, Knösche, and Scherf probe how transformers picture a grid world. Their walkers cross a lattice and must hit a fixed endpoint on time. Best prediction needs only position versus target plus time left. The team trains decoder-only transformers on samples from the exact process. Hidden activations align with those sufficient statistics. The first attention block pulls the statistic out. Later layers reshape it into next-step geometry. Recurrent networks match the loss but skip that clean stage.

Catalog record

CREATORSasha Brenner, Thomas R. Knösche, Nico Scherf
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59739
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.16689

Browse the full library · About the Tower of Babel