Structure Development in List-Sorting Transformers

Einar Urdshals, Jasmina Urdshals * · 2025

Read Structure Development in List-Sorting Transformers by Einar Urdshals, Jasmina Urdshals * (2025) 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-2501-18666.txt.

Description

Einar Urdshals and Jasmina Urdshals watch a small Transformer learn to sort number lists in a 2025 arXiv paper. The model has one layer and uses attention only. At the end of training, its heads fall into two modes. One splits the number range among heads. The other blocks a head from copying a token it just used.

Range splitting showed up with or without a common shrink trick in training. That backs the view that nets lean toward plain plans on their own. The copy block links to a known trick in GPT-2. Data traits that push the model to its final shape are also named.

Catalog record

CREATOREinar Urdshals, Jasmina Urdshals *
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE75 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-18666

Browse the full library · About the Tower of Babel