Grounding Functional Similarity by Invariance-Aware Model Stitching

Ioannis Athanasiadis, Anmar Karmush, Michael Felsberg · 2025

Read Grounding Functional Similarity by Invariance-Aware Model Stitching by Ioannis Athanasiadis, Anmar Karmush, Michael Felsberg (2025) 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-2505-20142.txt.

Description

Ioannis Athanasiadis, Anmar Karmush, and Michael Felsberg tackle a flaw in how researchers judge whether two neural models work in the same way. A common test joins parts of two models and checks if the joined system still solves a task. Models that rely on different cues can pass that test and look alike when they are not.

The authors trace the flaw to blindness about what each model ignores or keeps fixed. They add a backward test to the forward one, forming a stricter joining method. Studying key setups, they show the new test gives a sounder view of likeness and reveals gaps the old test hid.

Catalog record

CREATORIoannis Athanasiadis, Anmar Karmush, Michael Felsberg
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-20142

Browse the full library · About the Tower of Babel