Meta-Abstraction Layer for AI Self-Reflection and Theory Revision
Year: 2025
Type: article
Abstract: We propose a novel Meta-Abstraction Layer (MAL) for intelligent agents that enables self-reflection and autonomous theory revision. This layer operates above conventional perceptual and reasoning modules, monitoring the agent’s own abstractions and models to detect inconsistencies, derive higher-order insights, and refine internal theories. We review related work in meta-learning (learn-to-learn), theory-of-mind, and cognitive architectures with meta-cognitive components such as CLARION and GLAIR. Two implementation paths are explored: a symbolic/meta-reasoning approach with explicit knowledge representations and revision rules, and a neural/self-organizing approach where a network dynamically reconfigures its representations. We outline proposed experimental tasks (e.g. concept drift adaptation, incremental concept learning, dynamic planning tasks) to evaluate the MAL, and we project that agents with this layer will show improved adaptability and concept formation (with some computational overhead). We analyze advantages (e.g. robust generalization, explainability) and risks (e.g. complexity, misaligned self-modification), and we discuss whether to integrate the MAL with epistemological building blocks of Existence → Change → Abstraction → Meta-Abstraction. This work offers a comprehensive blueprint for adding a reflective, theory-building capability to AI systems, bridging gaps between meta-learning and meta-cognition.
Authors: Nguyen, Phi Son, Ng, Mers
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6911936857
FWCI:
Cites:
Cited by:
Related to: