Parameter-Efficient Quality Estimation via Frozen Recursive Models

Umar Abubacar, Roman Bauer ,, Diptesh Kanojia , · 2026

Read Parameter-Efficient Quality Estimation via Frozen Recursive Models by Umar Abubacar, Roman Bauer ,, Diptesh Kanojia , (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-2603-14593.txt.

Description

Abubacar, Bauer and Kanojia test tiny recursive models for translation quality estimation. Recursive models refine answers by looping one shared network. The team checks if that loop helps score translations in low-resource languages. Tests span eight language pairs. Outside iteration hurts scores, they find, and inside recursion helps only in narrow ways. Frozen pretrained embeddings match fine-tuned ones while cutting trainable weights 37-fold. Their frozen model ties fine-tuned scores and beats a plain transformer of equal depth.

Catalog record

CREATORUmar Abubacar, Roman Bauer ,, Diptesh Kanojia ,
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE18617
LICENSEhttp://creativecommons.org/licenses/by-sa/4.0/
RIGHTSCC BY SA
SOURCEarXiv

Source: https://arxiv.org/abs/2603.14593

Browse the full library · About the Tower of Babel