Evaluating Small Decoder-Only Language Models for Grammar Correction and Text Simplification

Anthony Lamelas ∙ \bullet New York University · 2026

Read Evaluating Small Decoder-Only Language Models for Grammar Correction and Text Simplification by Anthony Lamelas ∙ \bullet New York University (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-2601-03874.txt.

Description

Anthony Lamelas, working alone, tests whether small decoder-only language models can handle grammar correction and text simplification well enough to replace far larger models. Big models are costly to run, hard to deploy, and awkward to secure, so a capable small model would be valuable in schools, offices, and on private devices. The experiments try the small models as shipped, after fine-tuning, and in sequential pipelines.

Testing used the JFLEG set for grammar and the ASSET set for simplification, scored with established metrics. The small models learned some behaviors well but stayed below strong baselines and current large models overall. They also struggled to preserve meaning and sometimes hallucinated content. Lamelas concludes that, despite their efficiency, today's small models are not yet competitive for rewriting tasks, and further advances are needed.

Catalog record

CREATORAnthony Lamelas ∙ \bullet New York University
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-03874

Browse the full library · About the Tower of Babel