On Why Form Shapes Reason
Chao, Ching Hsueh
Chao, Ching Hsueh
Author
Supervisor
Department
Natural Language Processing
Embargo End Date
30/05/2026
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Human reasoning thrives on structure - our capacity to perceive, compose, and abstract patterns shapes our understanding of the world. Inspired by Kant’s insight that cognition actively imposes structure onto experience, this thesis explores how artificial intelligence can similarly acquire structured reasoning through latent representation. At the core of this inquiry lies the Latent Program Network (LPN), a neural architecture designed to implicitly learn transformations as structured, composable entities within a continuous latent space. Leveraging category theory as both a conceptual framework and an analytical tool, the work reinterprets neural representations through categorical constructs objects, morphisms, and functors-to probe their inherent compositionality. Empirical studies conducted on structured matrix transformation tasks demonstrate that LPNs indeed encode algebraic properties, spontaneously discovering structured, generalizable transformations without explicit symbolic instruction. These findings offer insights into the emergence of compositional understanding within neural models, contributing to our understanding of categorical abstraction in deep learning and the foundations of structured intelligence.
Citation
Ching Hsueh Chao, “On Why Form Shapes Reason,” Master of Science thesis, Natural Language Processing, MBZUAI, 2025.
Source
Conference
Keywords
Category Theory, Reasoning, Representation Learning, Latent Program Networks, Variational Autoencoder, Test-Time Adaptation