Generalization Boundaries of Fine-Tuned Small Language Models for Graph Structural Inference

Michal Podstawski · 2026

Read Generalization Boundaries of Fine-Tuned Small Language Models for Graph Structural Inference by Michal Podstawski (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-2604-18092.txt.

Description

Podstawski investigates generalization of tuned small language models in graph inference. The study tests two axes: graph size and graph family distribution. With three instruction tuned models of 3 to 4 billion parameters and two serialization formats, it tests graphs much larger than the training range. It also tests held out random graph families. Tuned models keep strong ordinal consistency across distinct families. They keep ranking graphs by structure on larger inputs, with varying degradation by architecture. The findings map reliable generalization zones for graph reasoning tasks.

Catalog record

CREATORMichal Podstawski
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43254
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.18092

Browse the full library · About the Tower of Babel