Transform Learning: Procedural Reasoning Without Programs in Semantic Topology Architectures

Description Procedural Intelligence Through Structural Operators This paper introduces Transform Learning, a core primitive of the Semantic Topology Reasoning Architecture (STRA), and proposes a new model of procedural intelligence grounded in structural operators, not stored programs or token‑level reasoning traces. Instead of memorizing solution paths or synthesizing explicit programs, STRA learns reusable reasoning procedures—called transforms—that operate directly on activation states within the semantic topology. Transforms such as SPREAD, FILTER, FORK, MERGE, ORDER, and ANCHOR function as domain‑agnostic operators that manipulate semantic activation patterns. Through repeated use, meta‑observer feedback, and stability‑based evaluation, STRA learns which transform sequences reliably solve classes of problems. This enables the emergence of procedural expertise without storing symbolic rules, solution traces, or chain‑of‑thought sequences. Transform Learning bridges the gap between symbolic reasoning (explicit procedures) and neural adaptability (learning from experience). It provides a mechanism for cross‑domain procedural generalization, computational efficiency, and transparent reasoning. This paper positions Transform Learning as an essential architectural layer in STRA, enabling the system to learn how to think, not just what to know or how to describe reasoning. Abstract This paper presents Transform Learning, a mechanism for acquiring reusable reasoning procedures within the Semantic Topology Reasoning Architecture (STRA). Unlike systems that store explicit programs or regenerate chain‑of‑thought traces for each problem, STRA learns topology‑level operators—transforms—that manipulate activation arrays to produce stable, convergent reasoning states. Transform selection is guided by a meta‑observer that evaluates state stability, resonance, contradiction reduction, and activation efficiency. We formalize six core transforms (SPREAD, FILTER, FORK, MERGE, ORDER, ANCHOR) and show how composite transforms emerge through repeated use. Transform Learning enables procedural memory without storing traces, supports cross‑domain generalization through structural similarity, and provides a scalable model of procedural intelligence. This work positions Transform Learning as a foundational component of STRA’s structure‑centric approach to reasoning. Background This paper is part of a six‑paper series introducing and expanding the Semantic Topology Reasoning Architecture: Paper 0 – Manifest: Semantic Topology Reasoning Architecture (STRA) (10.5281/zenodo.18207533) Paper 1: Activation Arrays as Working Memory in Semantic Reasoning Systems (10.5281/zenodo.18207539) Paper 2: Causal Signatures for Cross‑Domain Reasoning (10.5281/zenodo.18207546) Paper 3: Selection Pressure & Natural Selection of Thought (10.5281/zenodo.18207552) Paper 4: Transform Learning: Procedural Reasoning Without Programs (this paper) Paper 5: The Semantic Abacus: Quantitative, Symbolic, and Skill‑Based Reasoning (10.5281/zenodo.18207560) Transform Learning provides the procedural layer of STRA, enabling the emergence of expertise and efficient reasoning. Key Contributions Transform Framework: Defines transforms as reusable, domain‑agnostic reasoning operators. Meta‑Observer Mechanism: Introduces stability‑based feedback for transform selection. Procedural Memory Without Traces: Demonstrates how STRA learns procedures without storing paths or programs. Cross‑Domain Generalization: Shows how structural operators transfer across domains. Expertise Emergence: Explains how repeated success compresses reasoning into efficient transform sequences. Architectural Integration: Establishes Transform Learning as a core layer in STRA’s reasoning stack. Research Impact This work contributes to cognitive architectures, meta‑learning, and explainable AI by: Providing a transparent alternative to chain‑of‑thought and program synthesis Introducing a scalable model of procedural intelligence Demonstrating how reasoning can be learned through structural operators Enabling expertise without parameter inflation or trace storage Strengthening STRA’s position as a structure‑centric reasoning paradigm Access and Documentation ORCID: https://orcid.org/0009-0003-4876-9273 GitHub: https://github.com/Neuron-Soul-AI/Neuron-Soul-AI LinkedIn: https://www.linkedin.com/in/marcelo-emanuel-paradela-teixeira-702082382/ Email: marcelo.soul.ai@gmail.com License: CC BY‑NC 4.0 © Marcelo Emanuel Paradela Teixeira 2026