Scalable Evaluation and Neural Models for Compositional Generalization

Giacomo Camposampiero, Pietro Barbiero, Michael Hersche, Roger Wattenhofer, Abbas Rahimi · 2025

Read Scalable Evaluation and Neural Models for Compositional Generalization by Giacomo Camposampiero, Pietro Barbiero, Michael Hersche, Roger Wattenhofer, Abbas Rahimi (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-2511-02667.txt.

Introduction

Giacomo Camposampiero and four colleagues study compositional generalization, the skill of handling new mixes of known ideas. A model might know red and squares yet fail on a red square it never saw paired. Their 2025 paper notes that tests for this skill lack shared standards, and that big vision models lack the built-in biases the skill seems to need.

They offer a stricter test protocol that cuts the computing cost from combinatorial to constant. They also trained more than five thousand models to take stock of current vision backbones. Their Attribute Invariant Networks set a new balance point, lifting accuracy by over 23 percent over baselines while adding only 16 percent more parameters, far less than fully split designs.

Catalog record

CREATORGiacomo Camposampiero, Pietro Barbiero, Michael Hersche, Roger Wattenhofer, Abbas Rahimi
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE161 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-02667

Browse the full library · About the Tower of Babel