Learning to Extrapolate to New Tasks: A Relational Approach to Task Extrapolation

Adam Ousherovitch, Yixin Wang · 2026

Read Learning to Extrapolate to New Tasks: A Relational Approach to Task Extrapolation by Adam Ousherovitch, Yixin Wang (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-2605-30132.txt.

Description

Adam Ousherovitch and Yixin Wang tackle task extrapolation to unseen tasks outside the training distribution. Their Relational Task Extrapolator rests on one insight: extrapolation means learning how tasks transform into one another. RTE splits each target into a known anchor plus a transformation, then learns an operator mapping pairs to target predictions. It is tested on parameter, length and compositional extrapolation, and extended to sequence prediction in fine-tuning. Across studies it substantially beats existing approaches on novel tasks, even out-of-range ones.

Catalog record

CREATORAdam Ousherovitch, Yixin Wang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE103150
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.30132

Browse the full library · About the Tower of Babel