Task Generalization With AutoRegressive Compositional Structure: Can Learning From D D Tasks Generalize to D T D^{T} Tasks?

Amirhesam Abedsoltan 1 , Huaqing Zhang 3 , Kaiyue Wen 4 , Hongzhou Lin 5 , Jingzhao Zhang 3 , Mikhail Belkin 1,2 · 2025

Read Task Generalization With AutoRegressive Compositional Structure: Can Learning From D D Tasks Generalize to D T D^{T} Tasks? by Amirhesam Abedsoltan 1 , Huaqing Zhang 3 , Kaiyue Wen 4 , Hongzhou Lin 5 , Jingzhao Zhang 3 , Mikhail Belkin 1,2 (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-2502-08991.txt.

Description

Amirhesam Abedsoltan and co-authors ask when learning a few tasks can spread to a huge family of tasks in 2025. Each task here is a chain of steps. Each step comes from a fixed set of basic moves. The total family grows fast with chain length. Large language models show such spread in daily use.

Theory says training on a small set can cover the whole family. The needed set grows only with the count of basic moves, not the chain length. Tests with transformers bear this out on a math puzzle task. The models used a few demos and step by step work. Spread also showed in sums and in translation tasks. The work gives a frame for this striking skill.

Catalog record

CREATORAmirhesam Abedsoltan 1 , Huaqing Zhang 3 , Kaiyue Wen 4 , Hongzhou Lin 5 , Jingzhao Zhang 3 , Mikhail Belkin 1,2
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE64 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-08991

Browse the full library · About the Tower of Babel