Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond Monolithic Models: Symbolic Seams for Composable Neuro-Symbolic Architectures

Nicolas Schuler, Vincenzo Scotti, Raffaela Mirandola

Abstract

Current Artificial Intelligence (AI) systems are frequently built around monolithic models that entangle perception, reasoning, and decision-making, a design that often conflicts with established software architecture principles. Large Language Models (LLMs) amplify this tendency, offering scale but limited transparency and adaptability. To address this, we argue for composability as a guiding principle that treats AI as a living architecture rather than a fixed artifact. We introduce symbolic seams: explicit architectural breakpoints where a system commits to inspectable, typed boundary objects, versioned constraint bundles, and decision traces. We describe how seams enable a composable neuro-symbolic design that combines the data-driven adaptability of learned components with the verifiability of explicit symbolic constraints -- combining strengths neither paradigm achieves alone. By treating AI systems as assemblies of interchangeable parts rather than indivisible wholes, we outline a direction for intelligent systems that are extensible, transparent, and amenable to principled evolution.

Beyond Monolithic Models: Symbolic Seams for Composable Neuro-Symbolic Architectures

Abstract

Current Artificial Intelligence (AI) systems are frequently built around monolithic models that entangle perception, reasoning, and decision-making, a design that often conflicts with established software architecture principles. Large Language Models (LLMs) amplify this tendency, offering scale but limited transparency and adaptability. To address this, we argue for composability as a guiding principle that treats AI as a living architecture rather than a fixed artifact. We introduce symbolic seams: explicit architectural breakpoints where a system commits to inspectable, typed boundary objects, versioned constraint bundles, and decision traces. We describe how seams enable a composable neuro-symbolic design that combines the data-driven adaptability of learned components with the verifiability of explicit symbolic constraints -- combining strengths neither paradigm achieves alone. By treating AI systems as assemblies of interchangeable parts rather than indivisible wholes, we outline a direction for intelligent systems that are extensible, transparent, and amenable to principled evolution.
Paper Structure (8 sections, 1 equation, 3 figures, 1 table)

This paper contains 8 sections, 1 equation, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Why Monolithic AI Resists Architectural Evolution
  4. Symbolic Seams for Composable Architecture
  5. Overview
  6. Design Commitments
  7. Illustrative Scenario: Enterprise Policy Assistant
  8. Discussion and Research Agenda

Figures (3)

  • Figure 1: Current entanglement postures: (a) system is substantially overlapped with its neural logic; (b) entanglement is hidden through informal mechanisms. $\Delta$s represent the changes that affect the system.
  • Figure 2: Proposed posture: (a) learned modules ($\theta$) are separated by symbolic seam connectors ($\phi$) that checkpoint constraints; the dashed circle shows change bounded at the affected seam. (b) Seams persist across evolution while modules are replaced ($\theta'$ and $\phi'$ are the updated versions). Effect of change $\Delta$ is contained.
  • Figure 3: A seam ($\phi$) commits to typed boundary objects ($\mathcal{T}_{\mathrm{in}}, \mathcal{T}_{\mathrm{out}}$), a versioned constraint bundle $\mathcal{C}^{(v)}$, and a trace schema $\mathcal{S}_{\mathrm{trace}}$. For stochastic components, the contract may specify distributional properties.