Large Reasoning Models Struggle to Transfer Parametric Knowledge Across Scripts

Lucas Bandarkar Alan Ansell Trevor CohnGoogle Research · 2026

Read Large Reasoning Models Struggle to Transfer Parametric Knowledge Across Scripts by Lucas Bandarkar Alan Ansell Trevor CohnGoogle Research (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-2603-17070.txt.

Description

Bandarkar, Ansell, and Cohn show that large reasoning models often fail to move knowledge across writing systems. They study model answers to local knowledge questions on two datasets. Regression analysis shows that script match, not language or family, best predicts transfer failure. Giving the model key entities in the source language helps cross script questions the most. The team then builds training data that teaches models to reason about transliteration gaps. They report that this training shrinks the cross script gap in two models.

Catalog record

CREATORLucas Bandarkar Alan Ansell Trevor CohnGoogle Research
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE30980
LICENSEhttp://creativecommons.org/licenses/by-sa/4.0/
RIGHTSCC BY SA
SOURCEarXiv

Source: https://arxiv.org/abs/2603.17070

Browse the full library · About the Tower of Babel