Neuro-Symbolic Moral Self-Reflection Engines for LLMs

Building on the neuro-symbolic approach of Lazzari et al. (2024) and the explainability focus in the recent literature, this idea proposes a closed-loop system: after an LLM makes a moral or value classification, it is required to generate a structured explanation (using an ontology or symbolic logic), then “self-reflect” by critiquing its own reasoning (perhaps using a different reasoning scaffold or a meta-prompt). If inconsistencies or biases are detected—such as defaulting to a single moral foundation or failing to justify a decision—the system either revises its output or flags it for human review. This goes beyond current explainable AI approaches by giving the model a “second-order” moral reasoning capacity: not just classifying or explaining, but actively interrogating and refining its own moral logic. Such self-reflection could reduce spurious bias and promote more human-aligned, context-sensitive moral outputs.

References:

1. Explainable Moral Values: a neuro-symbolic approach to value classification. Nicolas Lazzari, S. D. Giorgis, Aldo Gangemi, V. Presutti (2024). ESWC Satellite Events.
2. Explainable Moral Values: a neuro-symbolic approach to value classification. Nicolas Lazzari, S. D. Giorgis, Aldo Gangemi, V. Presutti (2024). ESWC Satellite Events.
3. Explainable Moral Values: a neuro-symbolic approach to value classification. Nicolas Lazzari, S. D. Giorgis, Aldo Gangemi, V. Presutti (2024). ESWC Satellite Events.