Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Breaking Up with Normatively Monolithic Agency with GRACE: A Reason-Based Neuro-Symbolic Architecture for Safe and Ethical AI Alignment

Felix Jahn, Yannic Muskalla, Lisa Dargasz, Patrick Schramowski, Kevin Baum

TL;DR

The paper tackles the problem of normatively misaligned AI arising from monolithic decision architectures by introducing GRACE, a neuro-symbolic governor that explicitly separates normative reasoning (MM) from instrumental optimization (DMM) and enforcement (Guard), guided by a case-based Moral Advisor. This modular design provides transparent interfaces, enabling interpretable justifications, contestability, and verifiable alignment in morally consequential settings. By grounding normative guidance in explicit reasons and macro-action types, GRACE supports principled reasoning, robust handling of normative uncertainty, and the possibility of formal verification. The TherapAI demonstrations illustrate how stakeholders can understand, contest, and refine agent behavior, highlighting practical implications for safer and more ethically aligned AI systems with broad applicability.

Abstract

As AI agents become increasingly autonomous, widely deployed in consequential contexts, and efficacious in bringing about real-world impacts, ensuring that their decisions are not only instrumentally effective but also normatively aligned has become critical. We introduce a neuro-symbolic reason-based containment architecture, Governor for Reason-Aligned ContainmEnt (GRACE), that decouples normative reasoning from instrumental decision-making and can contain AI agents of virtually any design. GRACE restructures decision-making into three modules: a Moral Module (MM) that determines permissible macro actions via deontic logic-based reasoning; a Decision-Making Module (DMM) that encapsulates the target agent while selecting instrumentally optimal primitive actions in accordance with derived macro actions; and a Guard that monitors and enforces moral compliance. The MM uses a reason-based formalism providing a semantic foundation for deontic logic, enabling interpretability, contestability, and justifiability. Its symbolic representation enriches the DMM's informational context and supports formal verification and statistical guarantees of alignment enforced by the Guard. We demonstrate GRACE on an example of a LLM therapy assistant, showing how it enables stakeholders to understand, contest, and refine agent behavior.

Breaking Up with Normatively Monolithic Agency with GRACE: A Reason-Based Neuro-Symbolic Architecture for Safe and Ethical AI Alignment

TL;DR

The paper tackles the problem of normatively misaligned AI arising from monolithic decision architectures by introducing GRACE, a neuro-symbolic governor that explicitly separates normative reasoning (MM) from instrumental optimization (DMM) and enforcement (Guard), guided by a case-based Moral Advisor. This modular design provides transparent interfaces, enabling interpretable justifications, contestability, and verifiable alignment in morally consequential settings. By grounding normative guidance in explicit reasons and macro-action types, GRACE supports principled reasoning, robust handling of normative uncertainty, and the possibility of formal verification. The TherapAI demonstrations illustrate how stakeholders can understand, contest, and refine agent behavior, highlighting practical implications for safer and more ethically aligned AI systems with broad applicability.

Abstract

As AI agents become increasingly autonomous, widely deployed in consequential contexts, and efficacious in bringing about real-world impacts, ensuring that their decisions are not only instrumentally effective but also normatively aligned has become critical. We introduce a neuro-symbolic reason-based containment architecture, Governor for Reason-Aligned ContainmEnt (GRACE), that decouples normative reasoning from instrumental decision-making and can contain AI agents of virtually any design. GRACE restructures decision-making into three modules: a Moral Module (MM) that determines permissible macro actions via deontic logic-based reasoning; a Decision-Making Module (DMM) that encapsulates the target agent while selecting instrumentally optimal primitive actions in accordance with derived macro actions; and a Guard that monitors and enforces moral compliance. The MM uses a reason-based formalism providing a semantic foundation for deontic logic, enabling interpretability, contestability, and justifiability. Its symbolic representation enriches the DMM's informational context and supports formal verification and statistical guarantees of alignment enforced by the Guard. We demonstrate GRACE on an example of a LLM therapy assistant, showing how it enables stakeholders to understand, contest, and refine agent behavior.
Paper Structure (20 sections, 4 equations, 4 figures)

This paper contains 20 sections, 4 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Background
  3. A Family of Problems
  4. AI Agents and Decision-Making
  5. Problem Statement
  6. Morally Charged Environments
  7. Action Abstraction and the Interface Challenge
  8. The Moral Advisor
  9. The Flattening Problem
  10. GRACE: A Modular Governor Architecture
  11. Normative Reasons as Moral Foundation
  12. A Modular Multi-Agent Architecture
  13. The Moral Module (MM)
  14. The Decision-Making Module (DMM)
  15. The Guard Module
  16. ...and 5 more sections

Figures (4)

  • Figure 1: The GRACE containment architecture. The Moral Module determines permissible macro action types via reason-based inference; the Decision-Making Module selects instrumentally optimal actions within these constraints; the Guard enforces compliance; and the Moral Advisor provides external corrective feedback---all without modifying the encapsulated core agent.
  • Figure 2: Graph representation of the reason theory and resulting moral reasoning in Cases C\ref{['ex:reasons']}-C\ref{['ex:Guard']} inside the MM.
  • Figure 3: Our containment architecture as multi-agent system: each module (as well as the overall system) implements an agent instance connected with the other modules (the environment) via their individual observation and action space.
  • Figure 4: Different levels of (temporal) action abstraction. Alternating sequences of environment states and primitive actions form a macro action, a specific set of macro actions constitutes a macro action type described by a logical formula $\varphi$.